Mercurial > pylearn
annotate dataset.py @ 216:4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
Previously, if a fieldname were associated with an integer column index (by
opposition to a column range or slice) then it would be returned as a Nx1
matrix.
Now if a fieldname is associated with an integer column index, then it will
make a field which is a vector of length N.
The old behaviour can still be achieved by associating a fieldname with
the slice(col, col+1).
author | James Bergstra <bergstrj@iro.umontreal.ca> |
---|---|
date | Thu, 22 May 2008 19:07:51 -0400 |
parents | bd728c83faff |
children | df3fae88ab46 |
rev | line source |
---|---|
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
1 |
12
ff4e551490f1
Added LookupList type in lookup_list.py and used it to keep order
bengioy@esprit.iro.umontreal.ca
parents:
11
diff
changeset
|
2 from lookup_list import LookupList |
ff4e551490f1
Added LookupList type in lookup_list.py and used it to keep order
bengioy@esprit.iro.umontreal.ca
parents:
11
diff
changeset
|
3 Example = LookupList |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
4 from misc import unique_elements_list_intersection |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
5 from string import join |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
6 from sys import maxint |
171
895b4b60f5e8
bugfix. Otherwise the example was writed over and not a new one was returned
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
167
diff
changeset
|
7 import numpy, copy |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
8 |
166 | 9 from exceptions import * |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
10 |
110
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
11 class AttributesHolder(object): |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
12 def __init__(self): pass |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
13 |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
14 def attributeNames(self): |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
15 raise AbstractFunction() |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
16 |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
17 def setAttributes(self,attribute_names,attribute_values,make_copies=False): |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
18 """ |
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
19 Allow the attribute_values to not be a list (but a single value) if the attribute_names is of length 1. |
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
20 """ |
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
21 if len(attribute_names)==1 and not (isinstance(attribute_values,list) or isinstance(attribute_values,tuple) ): |
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
22 attribute_values = [attribute_values] |
110
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
23 if make_copies: |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
24 for name,value in zip(attribute_names,attribute_values): |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
25 self.__setattr__(name,copy.deepcopy(value)) |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
26 else: |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
27 for name,value in zip(attribute_names,attribute_values): |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
28 self.__setattr__(name,value) |
193
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
29 |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
30 def getAttributes(self,attribute_names=None, return_copy=False): |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
31 """ |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
32 Return all (if attribute_names=None, in the order of attributeNames()) or a specified subset of attributes. |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
33 """ |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
34 if attribute_names is None: |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
35 attribute_names = self.attributeNames() |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
36 if return_copy: |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
37 return [copy.copy(self.__getattribute__(name)) for name in attribute_names] |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
38 else: |
cb6b945acf5a
Complete redesign of learner...
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
188
diff
changeset
|
39 return [self.__getattribute__(name) for name in attribute_names] |
110
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
40 |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
41 |
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
42 class DataSet(AttributesHolder): |
16 | 43 """A virtual base class for datasets. |
44 | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
45 A DataSet can be seen as a generalization of a matrix, meant to be used in conjunction |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
46 with learning algorithms (for training and testing them): rows/records are called examples, and |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
47 columns/attributes are called fields. The field value for a particular example can be an arbitrary |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
48 python object, which depends on the particular dataset. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
49 |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
50 We call a DataSet a 'stream' when its length is unbounded (otherwise its __len__ method |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
51 should return sys.maxint). |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
52 |
16 | 53 A DataSet is a generator of iterators; these iterators can run through the |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
54 examples or the fields in a variety of ways. A DataSet need not necessarily have a finite |
16 | 55 or known length, so this class can be used to interface to a 'stream' which |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
56 feeds on-line learning (however, as noted below, some operations are not |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
57 feasible or not recommanded on streams). |
16 | 58 |
59 To iterate over examples, there are several possibilities: | |
90
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
60 - for example in dataset: |
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
61 - for val1,val2,... in dataset: |
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
62 - for example in dataset(field1, field2,field3, ...): |
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
63 - for val1,val2,val3 in dataset(field1, field2,field3): |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
64 - for minibatch in dataset.minibatches([field1, field2, ...],minibatch_size=N): |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
65 - for mini1,mini2,mini3 in dataset.minibatches([field1, field2, field3], minibatch_size=N): |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
66 Each of these is documented below. All of these iterators are expected |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
67 to provide, in addition to the usual 'next()' method, a 'next_index()' method |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
68 which returns a non-negative integer pointing to the position of the next |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
69 example that will be returned by 'next()' (or of the first example in the |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
70 next minibatch returned). This is important because these iterators |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
71 can wrap around the dataset in order to do multiple passes through it, |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
72 in possibly unregular ways if the minibatch size is not a divisor of the |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
73 dataset length. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
74 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
75 To iterate over fields, one can do |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
76 - for field in dataset.fields(): |
46
c5b07e87b0cb
comments modif made by Yoshua
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
45
diff
changeset
|
77 for field_value in field: # iterate over the values associated to that field for all the dataset examples |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
78 - for field in dataset(field1,field2,...).fields() to select a subset of fields |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
79 - for field in dataset.fields(field1,field2,...) to select a subset of fields |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
80 and each of these fields is iterable over the examples: |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
81 - for field_examples in dataset.fields(): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
82 for example_value in field_examples: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
83 ... |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
84 but when the dataset is a stream (unbounded length), it is not recommanded to do |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
85 such things because the underlying dataset may refuse to access the different fields in |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
86 an unsynchronized ways. Hence the fields() method is illegal for streams, by default. |
132
f6505ec32dc3
Updated documentation slightly
Joseph Turian <turian@gmail.com>
parents:
128
diff
changeset
|
87 The result of fields() is a L{DataSetFields} object, which iterates over fields, |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
88 and whose elements are iterable over examples. A DataSetFields object can |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
89 be turned back into a DataSet with its examples() method:: |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
90 dataset2 = dataset1.fields().examples() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
91 and dataset2 should behave exactly like dataset1 (in fact by default dataset2==dataset1). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
92 |
16 | 93 Note: Fields are not mutually exclusive, i.e. two fields can overlap in their actual content. |
94 | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
95 Note: The content of a field can be of any type. Field values can also be 'missing' |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
96 (e.g. to handle semi-supervised learning), and in the case of numeric (numpy array) |
46
c5b07e87b0cb
comments modif made by Yoshua
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
45
diff
changeset
|
97 fields (i.e. an ArrayFieldsDataSet), NaN plays the role of a missing value. |
c5b07e87b0cb
comments modif made by Yoshua
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
45
diff
changeset
|
98 What about non-numeric values? None. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
99 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
100 Dataset elements can be indexed and sub-datasets (with a subset |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
101 of examples) can be extracted. These operations are not supported |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
102 by default in the case of streams. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
103 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
104 - dataset[:n] returns a dataset with the n first examples. |
16 | 105 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
106 - dataset[i1:i2:s] returns a dataset with the examples i1,i1+s,...i2-s. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
107 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
108 - dataset[i] returns an Example. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
109 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
110 - dataset[[i1,i2,...in]] returns a dataset with examples i1,i2,...in. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
111 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
112 - dataset[fieldname] an iterable over the values of the field fieldname across |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
113 the dataset (the iterable is obtained by default by calling valuesVStack |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
114 over the values for individual examples). |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
115 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
116 - dataset.<property> returns the value of a property associated with |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
117 the name <property>. The following properties should be supported: |
41 | 118 - 'description': a textual description or name for the dataset |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
119 - 'fieldtypes': a list of types (one per field) |
78 | 120 A DataSet may have other attributes that it makes visible to other objects. These are |
121 used to store information that is not example-wise but global to the dataset. | |
122 The list of names of these attributes is given by the attribute_names() method. | |
41 | 123 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
124 Datasets can be concatenated either vertically (increasing the length) or |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
125 horizontally (augmenting the set of fields), if they are compatible, using |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
126 the following operations (with the same basic semantics as numpy.hstack |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
127 and numpy.vstack): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
128 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
129 - dataset1 | dataset2 | dataset3 == dataset.hstack([dataset1,dataset2,dataset3]) |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
130 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
131 creates a new dataset whose list of fields is the concatenation of the list of |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
132 fields of the argument datasets. This only works if they all have the same length. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
133 |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
134 - dataset1 & dataset2 & dataset3 == dataset.vstack([dataset1,dataset2,dataset3]) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
135 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
136 creates a new dataset that concatenates the examples from the argument datasets |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
137 (and whose length is the sum of the length of the argument datasets). This only |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
138 works if they all have the same fields. |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
139 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
140 According to the same logic, and viewing a DataSetFields object associated to |
46
c5b07e87b0cb
comments modif made by Yoshua
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
45
diff
changeset
|
141 a DataSet as a kind of transpose of it, fields1 & fields2 concatenates fields of |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
142 a DataSetFields fields1 and fields2, and fields1 | fields2 concatenates their |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
143 examples. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
144 |
41 | 145 A dataset can hold arbitrary key-value pairs that may be used to access meta-data |
146 or other properties of the dataset or associated with the dataset or the result | |
147 of a computation stored in a dataset. These can be accessed through the [key] syntax | |
148 when key is a string (or more specifically, neither an integer, a slice, nor a list). | |
78 | 149 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
150 A DataSet sub-class should always redefine the following methods: |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
151 - __len__ if it is not a stream |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
152 - fieldNames |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
153 - minibatches_nowrap (called by DataSet.minibatches()) |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
154 - valuesHStack |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
155 - valuesVStack |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
156 For efficiency of implementation, a sub-class might also want to redefine |
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
157 - hasFields |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
158 - __getitem__ may not be feasible with some streams |
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
159 - __iter__ |
78 | 160 A sub-class should also append attributes to self._attribute_names |
161 (the default value returned by attributeNames()). | |
162 By convention, attributes not in attributeNames() should have a name | |
163 starting with an underscore. | |
164 @todo enforce/test that convention! | |
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
165 """ |
1
2cd82666b9a7
Added statscollector and started writing dataset and learner.
bengioy@esprit.iro.umontreal.ca
parents:
0
diff
changeset
|
166 |
83 | 167 numpy_vstack = lambda fieldname,values: numpy.vstack(values) |
168 numpy_hstack = lambda fieldnames,values: numpy.hstack(values) | |
77
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
169 |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
170 def __init__(self,description=None,fieldtypes=None): |
41 | 171 if description is None: |
172 # by default return "<DataSetType>(<SuperClass1>,<SuperClass2>,...)" | |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
173 description = type(self).__name__ + " ( " + join([x.__name__ for x in type(self).__bases__]) + " )" |
41 | 174 self.description=description |
60 | 175 self.fieldtypes=fieldtypes |
78 | 176 self._attribute_names = ["description"] |
177 if fieldtypes: | |
178 self._attribute_names.append("fieldtypes") | |
179 | |
180 def attributeNames(self): return self._attribute_names | |
181 | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
182 class MinibatchToSingleExampleIterator(object): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
183 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
184 Converts the result of minibatch iterator with minibatch_size==1 into |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
185 single-example values in the result. Therefore the result of |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
186 iterating on the dataset itself gives a sequence of single examples |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
187 (whereas the result of iterating over minibatches gives in each |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
188 Example field an iterable object over the individual examples in |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
189 the minibatch). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
190 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
191 def __init__(self, minibatch_iterator): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
192 self.minibatch_iterator = minibatch_iterator |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
193 self.minibatch = None |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
194 def __iter__(self): #makes for loop work |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
195 return self |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
196 def next(self): |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
197 size1_minibatch = self.minibatch_iterator.next() |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
198 if not self.minibatch: |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
199 self.minibatch = Example(size1_minibatch.keys(),[value[0] for value in size1_minibatch.values()]) |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
200 else: |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
201 self.minibatch._values = [value[0] for value in size1_minibatch.values()] |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
202 return self.minibatch |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
203 |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
204 def next_index(self): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
205 return self.minibatch_iterator.next_index() |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
206 |
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
207 def __iter__(self): |
16 | 208 """Supports the syntax "for i in dataset: ..." |
1
2cd82666b9a7
Added statscollector and started writing dataset and learner.
bengioy@esprit.iro.umontreal.ca
parents:
0
diff
changeset
|
209 |
16 | 210 Using this syntax, "i" will be an Example instance (or equivalent) with |
211 all the fields of DataSet self. Every field of "i" will give access to | |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
212 a field of a single example. Fields should be accessible via |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
213 i["fielname"] or i[3] (in the order defined by the elements of the |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
214 Example returned by this iterator), but the derived class is free |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
215 to accept any type of identifier, and add extra functionality to the iterator. |
16 | 216 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
217 The default implementation calls the minibatches iterator and extracts the first example of each field. |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
218 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
219 return DataSet.MinibatchToSingleExampleIterator(self.minibatches(None, minibatch_size = 1)) |
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
220 |
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
221 def __contains__(self, fieldname): |
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
222 return (fieldname in self.fieldNames()) \ |
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
223 or (fieldname in self.attributeNames()) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
224 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
225 class MinibatchWrapAroundIterator(object): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
226 """ |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
227 An iterator for minibatches that handles the case where we need to wrap around the |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
228 dataset because n_batches*minibatch_size > len(dataset). It is constructed from |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
229 a dataset that provides a minibatch iterator that does not need to handle that problem. |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
230 This class is a utility for dataset subclass writers, so that they do not have to handle |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
231 this issue multiple times, nor check that fieldnames are valid, nor handle the |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
232 empty fieldnames (meaning 'use all the fields'). |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
233 """ |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
234 def __init__(self,dataset,fieldnames,minibatch_size,n_batches,offset): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
235 self.dataset=dataset |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
236 self.fieldnames=fieldnames |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
237 self.minibatch_size=minibatch_size |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
238 self.n_batches=n_batches |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
239 self.n_batches_done=0 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
240 self.next_row=offset |
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
241 self.offset=offset |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
242 self.L=len(dataset) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
243 assert offset+minibatch_size<=self.L |
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
244 ds_nbatches = (self.L-self.next_row)/self.minibatch_size |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
245 if n_batches is not None: |
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
246 ds_nbatches = min(n_batches,ds_nbatches) |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
247 if fieldnames: |
216
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
248 if not dataset.hasFields(*fieldnames): |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
249 raise ValueError('field not present', fieldnames) |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
250 else: |
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
251 self.fieldnames=dataset.fieldNames() |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
252 self.iterator = self.dataset.minibatches_nowrap(self.fieldnames,self.minibatch_size, |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
253 ds_nbatches,self.next_row) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
254 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
255 def __iter__(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
256 return self |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
257 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
258 def next_index(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
259 return self.next_row |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
260 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
261 def next(self): |
43
e92244f30116
Corrected iterator logic errors
bengioy@grenat.iro.umontreal.ca
parents:
42
diff
changeset
|
262 if self.n_batches and self.n_batches_done==self.n_batches: |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
263 raise StopIteration |
101
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
264 elif not self.n_batches and self.next_row ==self.L: |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
265 raise StopIteration |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
266 upper = self.next_row+self.minibatch_size |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
267 if upper <=self.L: |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
268 minibatch = self.iterator.next() |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
269 else: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
270 if not self.n_batches: |
101
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
271 upper=min(upper, self.L) |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
272 # if their is not a fixed number of batch, we continue to the end of the dataset. |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
273 # this can create a minibatch that is smaller then the minibatch_size |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
274 assert (self.L-self.next_row)<=self.minibatch_size |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
275 minibatch = self.dataset.minibatches_nowrap(self.fieldnames,self.L-self.next_row,1,self.next_row).next() |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
276 else: |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
277 # we must concatenate (vstack) the bottom and top parts of our minibatch |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
278 # first get the beginning of our minibatch (top of dataset) |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
279 first_part = self.dataset.minibatches_nowrap(self.fieldnames,self.L-self.next_row,1,self.next_row).next() |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
280 second_part = self.dataset.minibatches_nowrap(self.fieldnames,upper-self.L,1,0).next() |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
281 minibatch = Example(self.fieldnames, |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
282 [self.dataset.valuesVStack(name,[first_part[name],second_part[name]]) |
a1740a99b81f
by default, in a minibatch without any fixed number of batchs, we need to finish at the end of the dataset. Now we return a minibatch at the end event if this minibacht size != the gived minibatch_size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
99
diff
changeset
|
283 for name in self.fieldnames]) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
284 self.next_row=upper |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
285 self.n_batches_done+=1 |
43
e92244f30116
Corrected iterator logic errors
bengioy@grenat.iro.umontreal.ca
parents:
42
diff
changeset
|
286 if upper >= self.L and self.n_batches: |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
287 self.next_row -= self.L |
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
288 ds_nbatches = (self.L-self.next_row)/self.minibatch_size |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
289 if self.n_batches is not None: |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
290 ds_nbatches = min(self.n_batches,ds_nbatches) |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
291 self.iterator = self.dataset.minibatches_nowrap(self.fieldnames,self.minibatch_size, |
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
292 ds_nbatches,self.next_row) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
293 return DataSetFields(MinibatchDataSet(minibatch,self.dataset.valuesVStack, |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
294 self.dataset.valuesHStack), |
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
295 minibatch.keys()) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
296 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
297 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
298 minibatches_fieldnames = None |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
299 minibatches_minibatch_size = 1 |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
300 minibatches_n_batches = None |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
301 def minibatches(self, |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
302 fieldnames = minibatches_fieldnames, |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
303 minibatch_size = minibatches_minibatch_size, |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
304 n_batches = minibatches_n_batches, |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
305 offset = 0): |
6
d5738b79089a
Removed MinibatchIterator and instead made minibatch_size a field of all DataSets,
bengioy@bengiomac.local
parents:
5
diff
changeset
|
306 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
307 Return an iterator that supports three forms of syntax: |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
308 |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
309 for i in dataset.minibatches(None,**kwargs): ... |
16 | 310 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
311 for i in dataset.minibatches([f1, f2, f3],**kwargs): ... |
16 | 312 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
313 for i1, i2, i3 in dataset.minibatches([f1, f2, f3],**kwargs): ... |
16 | 314 |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
315 Using the first two syntaxes, "i" will be an indexable object, such as a list, |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
316 tuple, or Example instance. In both cases, i[k] is a list-like container |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
317 of a batch of current examples. In the second case, i[0] is |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
318 list-like container of the f1 field of a batch current examples, i[1] is |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
319 a list-like container of the f2 field, etc. |
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
320 |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
321 Using the first syntax, all the fields will be returned in "i". |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
322 Using the third syntax, i1, i2, i3 will be list-like containers of the |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
323 f1, f2, and f3 fields of a batch of examples on each loop iteration. |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
324 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
325 The minibatches iterator is expected to return upon each call to next() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
326 a DataSetFields object, which is a LookupList (indexed by the field names) whose |
80 | 327 elements are iterable and indexable over the minibatch examples, and which keeps a pointer to |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
328 a sub-dataset that can be used to iterate over the individual examples |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
329 in the minibatch. Hence a minibatch can be converted back to a regular |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
330 dataset or its fields can be looked at individually (and possibly iterated over). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
331 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
332 PARAMETERS |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
333 - fieldnames (list of any type, default None): |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
334 The loop variables i1, i2, i3 (in the example above) should contain the |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
335 f1, f2, and f3 fields of the current batch of examples. If None, the |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
336 derived class can choose a default, e.g. all fields. |
16 | 337 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
338 - minibatch_size (integer, default 1) |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
339 On every iteration, the variables i1, i2, i3 will have |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
340 exactly minibatch_size elements. e.g. len(i1) == minibatch_size |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
341 |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
342 - n_batches (integer, default None) |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
343 The iterator will loop exactly this many times, and then stop. If None, |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
344 the derived class can choose a default. If (-1), then the returned |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
345 iterator should support looping indefinitely. |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
346 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
347 - offset (integer, default 0) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
348 The iterator will start at example 'offset' in the dataset, rather than the default. |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
349 |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
350 Note: A list-like container is something like a tuple, list, numpy.ndarray or |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
351 any other object that supports integer indexing and slicing. |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
352 |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
353 """ |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
354 return DataSet.MinibatchWrapAroundIterator(self,fieldnames,minibatch_size,n_batches,offset) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
355 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
356 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
357 """ |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
358 This is the minibatches iterator generator that sub-classes must define. |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
359 It does not need to worry about wrapping around multiple times across the dataset, |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
360 as this is handled by MinibatchWrapAroundIterator when DataSet.minibatches() is called. |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
361 The next() method of the returned iterator does not even need to worry about |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
362 the termination condition (as StopIteration will be raised by DataSet.minibatches |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
363 before an improper call to minibatches_nowrap's next() is made). |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
364 That next() method can assert that its next row will always be within [0,len(dataset)). |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
365 The iterator returned by minibatches_nowrap does not need to implement |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
366 a next_index() method either, as this will be provided by MinibatchWrapAroundIterator. |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
367 """ |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
368 raise AbstractFunction() |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
369 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
370 def __len__(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
371 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
372 len(dataset) returns the number of examples in the dataset. |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
373 By default, a DataSet is a 'stream', i.e. it has an unbounded length (sys.maxint). |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
374 Sub-classes which implement finite-length datasets should redefine this method. |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
375 Some methods only make sense for finite-length datasets. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
376 """ |
123 | 377 return maxint |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
378 |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
379 def is_unbounded(self): |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
380 """ |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
381 Tests whether a dataset is unbounded (e.g. a stream). |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
382 """ |
123 | 383 return len(self)==maxint |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
384 |
26
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
385 def hasFields(self,*fieldnames): |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
386 """ |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
387 Return true if the given field name (or field names, if multiple arguments are |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
388 given) is recognized by the DataSet (i.e. can be used as a field name in one |
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
389 of the iterators). |
29
46c5c90019c2
Changed apply_function so that it propagates methods of the source.
bengioy@grenat.iro.umontreal.ca
parents:
28
diff
changeset
|
390 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
391 The default implementation may be inefficient (O(# fields in dataset)), as it calls the fieldNames() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
392 method. Many datasets may store their field names in a dictionary, which would allow more efficiency. |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
393 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
394 return len(unique_elements_list_intersection(fieldnames,self.fieldNames()))>0 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
395 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
396 def fieldNames(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
397 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
398 Return the list of field names that are supported by the iterators, |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
399 and for which hasFields(fieldname) would return True. |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
400 """ |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
401 raise AbstractFunction() |
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
402 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
403 def __call__(self,*fieldnames): |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
404 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
405 Return a dataset that sees only the fields whose name are specified. |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
406 """ |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
407 assert self.hasFields(*fieldnames) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
408 return self.fields(*fieldnames).examples() |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
409 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
410 def fields(self,*fieldnames): |
29
46c5c90019c2
Changed apply_function so that it propagates methods of the source.
bengioy@grenat.iro.umontreal.ca
parents:
28
diff
changeset
|
411 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
412 Return a DataSetFields object associated with this dataset. |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
413 """ |
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
414 return DataSetFields(self,fieldnames) |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
415 |
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
416 def __getitem__(self,i): |
28
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
417 """ |
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
418 dataset[i] returns the (i+1)-th example of the dataset. |
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
419 dataset[i:j] returns the subdataset with examples i,i+1,...,j-1. |
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
420 dataset[i:j:s] returns the subdataset with examples i,i+2,i+4...,j-2. |
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
421 dataset[[i1,i2,..,in]] returns the subdataset with examples i1,i2,...,in. |
41 | 422 dataset['key'] returns a property associated with the given 'key' string. |
423 If 'key' is a fieldname, then the VStacked field values (iterable over | |
424 field values) for that field is returned. Other keys may be supported | |
425 by different dataset subclasses. The following key names are encouraged: | |
426 - 'description': a textual description or name for the dataset | |
427 - '<fieldname>.type': a type name or value for a given <fieldname> | |
1
2cd82666b9a7
Added statscollector and started writing dataset and learner.
bengioy@esprit.iro.umontreal.ca
parents:
0
diff
changeset
|
428 |
39 | 429 Note that some stream datasets may be unable to implement random access, i.e. |
430 arbitrary slicing/indexing | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
431 because they can only iterate through examples one or a minibatch at a time |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
432 and do not actually store or keep past (or future) examples. |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
433 |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
434 The default implementation of getitem uses the minibatches iterator |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
435 to obtain one example, one slice, or a list of examples. It may not |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
436 always be the most efficient way to obtain the result, especially if |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
437 the data are actually stored in a memory array. |
28
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
438 """ |
41 | 439 # check for an index |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
440 if type(i) is int: |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
441 return DataSet.MinibatchToSingleExampleIterator( |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
442 self.minibatches(minibatch_size=1,n_batches=1,offset=i)).next() |
41 | 443 rows=None |
444 # or a slice | |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
445 if type(i) is slice: |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
446 #print 'i=',i |
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
447 if not i.start: i=slice(0,i.stop,i.step) |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
448 if not i.stop: i=slice(i.start,len(self),i.step) |
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
449 if not i.step: i=slice(i.start,i.stop,1) |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
450 if i.step is 1: |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
451 return self.minibatches(minibatch_size=i.stop-i.start,n_batches=1,offset=i.start).next().examples() |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
452 rows = range(i.start,i.stop,i.step) |
41 | 453 # or a list of indices |
454 elif type(i) is list: | |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
455 rows = i |
41 | 456 if rows is not None: |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
457 examples = [self[row] for row in rows] |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
458 fields_values = zip(*examples) |
45
a5c70dc42972
Test functions for dataset.py
bengioy@grenat.iro.umontreal.ca
parents:
44
diff
changeset
|
459 return MinibatchDataSet( |
41 | 460 Example(self.fieldNames(),[ self.valuesVStack(fieldname,field_values) |
461 for fieldname,field_values | |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
462 in zip(self.fieldNames(),fields_values)]), |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
463 self.valuesVStack,self.valuesHStack) |
41 | 464 # else check for a fieldname |
465 if self.hasFields(i): | |
466 return self.minibatches(fieldnames=[i],minibatch_size=len(self),n_batches=1,offset=0).next()[0] | |
467 # else we are trying to access a property of the dataset | |
468 assert i in self.__dict__ # else it means we are trying to access a non-existing property | |
469 return self.__dict__[i] | |
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
470 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
471 def valuesHStack(self,fieldnames,fieldvalues): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
472 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
473 Return a value that corresponds to concatenating (horizontally) several field values. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
474 This can be useful to merge some fields. The implementation of this operation is likely |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
475 to involve a copy of the original values. When the values are numpy arrays, the |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
476 result should be numpy.hstack(values). If it makes sense, this operation should |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
477 work as well when each value corresponds to multiple examples in a minibatch |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
478 e.g. if each value is a Ni-vector and a minibatch of length L is a LxNi matrix, |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
479 then the result should be a Lx(N1+N2+..) matrix equal to numpy.hstack(values). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
480 The default is to use numpy.hstack for numpy.ndarray values, and a list |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
481 pointing to the original values for other data types. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
482 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
483 all_numpy=True |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
484 for value in fieldvalues: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
485 if not type(value) is numpy.ndarray: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
486 all_numpy=False |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
487 if all_numpy: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
488 return numpy.hstack(fieldvalues) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
489 # the default implementation of horizontal stacking is to put values in a list |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
490 return fieldvalues |
26
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
491 |
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
492 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
493 def valuesVStack(self,fieldname,values): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
494 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
495 Return a value that corresponds to concatenating (vertically) several values of the |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
496 same field. This can be important to build a minibatch out of individual examples. This |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
497 is likely to involve a copy of the original values. When the values are numpy arrays, the |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
498 result should be numpy.vstack(values). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
499 The default is to use numpy.vstack for numpy.ndarray values, and a list |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
500 pointing to the original values for other data types. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
501 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
502 all_numpy=True |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
503 for value in values: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
504 if not type(value) is numpy.ndarray: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
505 all_numpy=False |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
506 if all_numpy: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
507 return numpy.vstack(values) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
508 # the default implementation of vertical stacking is to put values in a list |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
509 return values |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
510 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
511 def __or__(self,other): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
512 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
513 dataset1 | dataset2 returns a dataset whose list of fields is the concatenation of the list of |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
514 fields of the argument datasets. This only works if they all have the same length. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
515 """ |
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
516 return HStackedDataSet([self,other]) |
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
517 |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
518 def __and__(self,other): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
519 """ |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
520 dataset1 & dataset2 is a dataset that concatenates the examples from the argument datasets |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
521 (and whose length is the sum of the length of the argument datasets). This only |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
522 works if they all have the same fields. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
523 """ |
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
524 return VStackedDataSet([self,other]) |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
525 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
526 def hstack(datasets): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
527 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
528 hstack(dataset1,dataset2,...) returns dataset1 | datataset2 | ... |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
529 which is a dataset whose fields list is the concatenation of the fields |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
530 of the individual datasets. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
531 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
532 assert len(datasets)>0 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
533 if len(datasets)==1: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
534 return datasets[0] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
535 return HStackedDataSet(datasets) |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
536 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
537 def vstack(datasets): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
538 """ |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
539 vstack(dataset1,dataset2,...) returns dataset1 & datataset2 & ... |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
540 which is a dataset which iterates first over the examples of dataset1, then |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
541 over those of dataset2, etc. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
542 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
543 assert len(datasets)>0 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
544 if len(datasets)==1: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
545 return datasets[0] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
546 return VStackedDataSet(datasets) |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
547 |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
548 class FieldsSubsetDataSet(DataSet): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
549 """ |
167 | 550 A sub-class of L{DataSet} that selects a subset of the fields. |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
551 """ |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
552 def __init__(self,src,fieldnames): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
553 self.src=src |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
554 self.fieldnames=fieldnames |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
555 assert src.hasFields(*fieldnames) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
556 self.valuesHStack = src.valuesHStack |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
557 self.valuesVStack = src.valuesVStack |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
558 |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
559 def __len__(self): return len(self.src) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
560 |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
561 def fieldNames(self): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
562 return self.fieldnames |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
563 |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
564 def __iter__(self): |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
565 class FieldsSubsetIterator(object): |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
566 def __init__(self,ds): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
567 self.ds=ds |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
568 self.src_iter=ds.src.__iter__() |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
569 self.example=None |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
570 def __iter__(self): return self |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
571 def next(self): |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
572 complete_example = self.src_iter.next() |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
573 if self.example: |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
574 self.example._values=[complete_example[field] |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
575 for field in self.ds.fieldnames] |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
576 else: |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
577 self.example=Example(self.ds.fieldnames, |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
578 [complete_example[field] for field in self.ds.fieldnames]) |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
579 return self.example |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
580 return FieldsSubsetIterator(self) |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
581 |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
582 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
583 assert self.hasFields(*fieldnames) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
584 return self.src.minibatches_nowrap(fieldnames,minibatch_size,n_batches,offset) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
585 def __getitem__(self,i): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
586 return FieldsSubsetDataSet(self.src[i],self.fieldnames) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
587 |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
588 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
589 class DataSetFields(LookupList): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
590 """ |
167 | 591 Although a L{DataSet} iterates over examples (like rows of a matrix), an associated |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
592 DataSetFields iterates over fields (like columns of a matrix), and can be understood |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
593 as a transpose of the associated dataset. |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
594 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
595 To iterate over fields, one can do |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
596 * for fields in dataset.fields() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
597 * for fields in dataset(field1,field2,...).fields() to select a subset of fields |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
598 * for fields in dataset.fields(field1,field2,...) to select a subset of fields |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
599 and each of these fields is iterable over the examples: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
600 * for field_examples in dataset.fields(): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
601 for example_value in field_examples: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
602 ... |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
603 but when the dataset is a stream (unbounded length), it is not recommanded to do |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
604 such things because the underlying dataset may refuse to access the different fields in |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
605 an unsynchronized ways. Hence the fields() method is illegal for streams, by default. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
606 The result of fields() is a DataSetFields object, which iterates over fields, |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
607 and whose elements are iterable over examples. A DataSetFields object can |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
608 be turned back into a DataSet with its examples() method: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
609 dataset2 = dataset1.fields().examples() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
610 and dataset2 should behave exactly like dataset1 (in fact by default dataset2==dataset1). |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
611 |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
612 DataSetFields can be concatenated vertically or horizontally. To be consistent with |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
613 the syntax used for DataSets, the | concatenates the fields and the & concatenates |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
614 the examples. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
615 """ |
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
616 def __init__(self,dataset,fieldnames): |
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
617 original_dataset=dataset |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
618 if not fieldnames: |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
619 fieldnames=dataset.fieldNames() |
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
620 elif not fieldnames==dataset.fieldNames(): |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
621 dataset = FieldsSubsetDataSet(dataset,fieldnames) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
622 assert dataset.hasFields(*fieldnames) |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
623 self.dataset=dataset |
66
dde1fb1b63ba
fixed test and removed print
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
65
diff
changeset
|
624 |
64
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
625 if isinstance(dataset,MinibatchDataSet): |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
626 LookupList.__init__(self,fieldnames,list(dataset._fields)) |
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
627 elif isinstance(original_dataset,MinibatchDataSet): |
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
628 LookupList.__init__(self,fieldnames, |
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
629 [original_dataset._fields[field] |
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
630 for field in fieldnames]) |
64
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
631 else: |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
632 minibatch_iterator = dataset.minibatches(fieldnames, |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
633 minibatch_size=len(dataset), |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
634 n_batches=1) |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
635 minibatch=minibatch_iterator.next() |
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
636 LookupList.__init__(self,fieldnames,minibatch) |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
637 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
638 def examples(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
639 return self.dataset |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
640 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
641 def __or__(self,other): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
642 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
643 fields1 | fields2 is a DataSetFields that whose list of examples is the concatenation |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
644 of the list of examples of DataSetFields fields1 and fields2. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
645 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
646 return (self.examples() + other.examples()).fields() |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
647 |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
648 def __and__(self,other): |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
649 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
650 fields1 + fields2 is a DataSetFields that whose list of fields is the concatenation |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
651 of the fields of DataSetFields fields1 and fields2. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
652 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
653 return (self.examples() | other.examples()).fields() |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
654 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
655 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
656 class MinibatchDataSet(DataSet): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
657 """ |
167 | 658 Turn a L{LookupList} of same-length (iterable) fields into an example-iterable dataset. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
659 Each element of the lookup-list should be an iterable and sliceable, all of the same length. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
660 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
661 def __init__(self,fields_lookuplist,values_vstack=DataSet().valuesVStack, |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
662 values_hstack=DataSet().valuesHStack): |
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
663 """ |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
664 The user can (and generally should) also provide values_vstack(fieldname,fieldvalues) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
665 and a values_hstack(fieldnames,fieldvalues) functions behaving with the same |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
666 semantics as the DataSet methods of the same name (but without the self argument). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
667 """ |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
668 |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
669 self._fields=fields_lookuplist |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
670 assert len(fields_lookuplist)>0 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
671 self.length=len(fields_lookuplist[0]) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
672 for field in fields_lookuplist[1:]: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
673 assert self.length==len(field) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
674 self.values_vstack=values_vstack |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
675 self.values_hstack=values_hstack |
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
676 |
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
677 def __len__(self): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
678 return self.length |
28
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
679 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
680 def __getitem__(self,i): |
80 | 681 if type(i) in (slice,list): |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
682 return DataSetFields(MinibatchDataSet( |
80 | 683 Example(self._fields.keys(),[field[i] for field in self._fields])),self.fieldNames()) |
684 if type(i) is int: | |
85 | 685 return Example(self._fields.keys(),[field[i] for field in self._fields]) |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
686 if self.hasFields(i): |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
687 return self._fields[i] |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
688 assert i in self.__dict__ # else it means we are trying to access a non-existing property |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
689 return self.__dict__[i] |
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
690 |
29
46c5c90019c2
Changed apply_function so that it propagates methods of the source.
bengioy@grenat.iro.umontreal.ca
parents:
28
diff
changeset
|
691 def fieldNames(self): |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
692 return self._fields.keys() |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
693 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
694 def hasFields(self,*fieldnames): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
695 for fieldname in fieldnames: |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
696 if fieldname not in self._fields.keys(): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
697 return False |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
698 return True |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
699 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
700 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
701 class Iterator(object): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
702 def __init__(self,ds): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
703 self.ds=ds |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
704 self.next_example=offset |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
705 assert minibatch_size > 0 |
41 | 706 if offset+minibatch_size > ds.length: |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
707 raise NotImplementedError() |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
708 def __iter__(self): |
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
709 return self |
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
710 def next(self): |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
711 upper = self.next_example+minibatch_size |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
712 assert upper<=self.ds.length |
61
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
713 minibatch = Example(self.ds._fields.keys(), |
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
714 [field[self.next_example:upper] |
a8b70a9117ad
bugfix: in MinibatchDataSet renamed the class variable fields to _fields as parent class have a function called field.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
60
diff
changeset
|
715 for field in self.ds._fields]) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
716 self.next_example+=minibatch_size |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
717 return minibatch |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
718 |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
719 return Iterator(self) |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
720 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
721 def valuesVStack(self,fieldname,fieldvalues): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
722 return self.values_vstack(fieldname,fieldvalues) |
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
723 |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
724 def valuesHStack(self,fieldnames,fieldvalues): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
725 return self.values_hstack(fieldnames,fieldvalues) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
726 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
727 class HStackedDataSet(DataSet): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
728 """ |
167 | 729 A L{DataSet} that wraps several datasets and shows a view that includes all their fields, |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
730 i.e. whose list of fields is the concatenation of their lists of fields. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
731 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
732 If a field name is found in more than one of the datasets, then either an error is |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
733 raised or the fields are renamed (either by prefixing the __name__ attribute |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
734 of the dataset + ".", if it exists, or by suffixing the dataset index in the argument list). |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
735 |
167 | 736 @todo: automatically detect a chain of stacked datasets due to A | B | C | D ... |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
737 """ |
41 | 738 def __init__(self,datasets,accept_nonunique_names=False,description=None,field_types=None): |
739 DataSet.__init__(self,description,field_types) | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
740 self.datasets=datasets |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
741 self.accept_nonunique_names=accept_nonunique_names |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
742 self.fieldname2dataset={} |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
743 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
744 def rename_field(fieldname,dataset,i): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
745 if hasattr(dataset,"__name__"): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
746 return dataset.__name__ + "." + fieldname |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
747 return fieldname+"."+str(i) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
748 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
749 # make sure all datasets have the same length and unique field names |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
750 self.length=None |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
751 names_to_change=[] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
752 for i in xrange(len(datasets)): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
753 dataset = datasets[i] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
754 length=len(dataset) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
755 if self.length: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
756 assert self.length==length |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
757 else: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
758 self.length=length |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
759 for fieldname in dataset.fieldNames(): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
760 if fieldname in self.fieldname2dataset: # name conflict! |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
761 if accept_nonunique_names: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
762 fieldname=rename_field(fieldname,dataset,i) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
763 names2change.append((fieldname,i)) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
764 else: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
765 raise ValueError("Incompatible datasets: non-unique field name = "+fieldname) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
766 self.fieldname2dataset[fieldname]=i |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
767 for fieldname,i in names_to_change: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
768 del self.fieldname2dataset[fieldname] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
769 self.fieldname2dataset[rename_field(fieldname,self.datasets[i],i)]=i |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
770 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
771 def hasFields(self,*fieldnames): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
772 for fieldname in fieldnames: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
773 if not fieldname in self.fieldname2dataset: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
774 return False |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
775 return True |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
776 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
777 def fieldNames(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
778 return self.fieldname2dataset.keys() |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
779 |
41 | 780 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
781 |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
782 class HStackedIterator(object): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
783 def __init__(self,hsds,iterators): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
784 self.hsds=hsds |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
785 self.iterators=iterators |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
786 def __iter__(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
787 return self |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
788 def next(self): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
789 # concatenate all the fields of the minibatches |
140 | 790 l=LookupList() |
791 for iter in self.iterators: | |
792 l.append_lookuplist(iter.next()) | |
793 return l | |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
794 |
125 | 795 assert self.hasFields(*fieldnames) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
796 # find out which underlying datasets are necessary to service the required fields |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
797 # and construct corresponding minibatch iterators |
140 | 798 if fieldnames and fieldnames!=self.fieldNames(): |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
799 datasets=set([]) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
800 fields_in_dataset=dict([(dataset,[]) for dataset in datasets]) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
801 for fieldname in fieldnames: |
136 | 802 dataset=self.datasets[self.fieldname2dataset[fieldname]] |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
803 datasets.add(dataset) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
804 fields_in_dataset[dataset].append(fieldname) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
805 datasets=list(datasets) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
806 iterators=[dataset.minibatches(fields_in_dataset[dataset],minibatch_size,n_batches,offset) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
807 for dataset in datasets] |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
808 else: |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
809 datasets=self.datasets |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
810 iterators=[dataset.minibatches(None,minibatch_size,n_batches,offset) for dataset in datasets] |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
811 return HStackedIterator(self,iterators) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
812 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
813 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
814 def valuesVStack(self,fieldname,fieldvalues): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
815 return self.datasets[self.fieldname2dataset[fieldname]].valuesVStack(fieldname,fieldvalues) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
816 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
817 def valuesHStack(self,fieldnames,fieldvalues): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
818 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
819 We will use the sub-dataset associated with the first fieldname in the fieldnames list |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
820 to do the work, hoping that it can cope with the other values (i.e. won't care |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
821 about the incompatible fieldnames). Hence this heuristic will always work if |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
822 all the fieldnames are of the same sub-dataset. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
823 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
824 return self.datasets[self.fieldname2dataset[fieldnames[0]]].valuesHStack(fieldnames,fieldvalues) |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
825 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
826 class VStackedDataSet(DataSet): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
827 """ |
167 | 828 A L{DataSet} that wraps several datasets and shows a view that includes all their examples, |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
829 in the order provided. This clearly assumes that they all have the same field names |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
830 and all (except possibly the last one) are of finite length. |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
831 |
167 | 832 @todo: automatically detect a chain of stacked datasets due to A + B + C + D ... |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
833 """ |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
834 def __init__(self,datasets): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
835 self.datasets=datasets |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
836 self.length=0 |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
837 self.index2dataset={} |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
838 assert len(datasets)>0 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
839 fieldnames = datasets[-1].fieldNames() |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
840 self.datasets_start_row=[] |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
841 # We use this map from row index to dataset index for constant-time random access of examples, |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
842 # to avoid having to search for the appropriate dataset each time and slice is asked for. |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
843 for dataset,k in enumerate(datasets[0:-1]): |
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
844 assert dataset.is_unbounded() # All VStacked datasets (except possibly the last) must be bounded (have a length). |
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
845 L=len(dataset) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
846 for i in xrange(L): |
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
847 self.index2dataset[self.length+i]=k |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
848 self.datasets_start_row.append(self.length) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
849 self.length+=L |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
850 assert dataset.fieldNames()==fieldnames |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
851 self.datasets_start_row.append(self.length) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
852 self.length+=len(datasets[-1]) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
853 # If length is very large, we should use a more memory-efficient mechanism |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
854 # that does not store all indices |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
855 if self.length>1000000: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
856 # 1 million entries would require about 60 meg for the index2dataset map |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
857 # TODO |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
858 print "A more efficient mechanism for index2dataset should be implemented" |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
859 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
860 def __len__(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
861 return self.length |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
862 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
863 def fieldNames(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
864 return self.datasets[0].fieldNames() |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
865 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
866 def hasFields(self,*fieldnames): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
867 return self.datasets[0].hasFields(*fieldnames) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
868 |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
869 def locate_row(self,row): |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
870 """Return (dataset_index, row_within_dataset) for global row number""" |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
871 dataset_index = self.index2dataset[row] |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
872 row_within_dataset = self.datasets_start_row[dataset_index] |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
873 return dataset_index, row_within_dataset |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
874 |
41 | 875 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
876 |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
877 class VStackedIterator(object): |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
878 def __init__(self,vsds): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
879 self.vsds=vsds |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
880 self.next_row=offset |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
881 self.next_dataset_index,self.next_dataset_row=self.vsds.locate_row(offset) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
882 self.current_iterator,self.n_left_at_the_end_of_ds,self.n_left_in_mb= \ |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
883 self.next_iterator(vsds.datasets[0],offset,n_batches) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
884 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
885 def next_iterator(self,dataset,starting_offset,batches_left): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
886 L=len(dataset) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
887 ds_nbatches = (L-starting_offset)/minibatch_size |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
888 if batches_left is not None: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
889 ds_nbatches = max(batches_left,ds_nbatches) |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
890 if minibatch_size>L: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
891 ds_minibatch_size=L |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
892 n_left_in_mb=minibatch_size-L |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
893 ds_nbatches=1 |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
894 else: |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
895 n_left_in_mb=0 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
896 return dataset.minibatches(fieldnames,minibatch_size,ds_nbatches,starting_offset), \ |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
897 L-(starting_offset+ds_nbatches*minibatch_size), n_left_in_mb |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
898 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
899 def move_to_next_dataset(self): |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
900 if self.n_left_at_the_end_of_ds>0: |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
901 self.current_iterator,self.n_left_at_the_end_of_ds,self.n_left_in_mb= \ |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
902 self.next_iterator(vsds.datasets[self.next_dataset_index], |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
903 self.n_left_at_the_end_of_ds,1) |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
904 else: |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
905 self.next_dataset_index +=1 |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
906 if self.next_dataset_index==len(self.vsds.datasets): |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
907 self.next_dataset_index = 0 |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
908 self.current_iterator,self.n_left_at_the_end_of_ds,self.n_left_in_mb= \ |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
909 self.next_iterator(vsds.datasets[self.next_dataset_index],starting_offset,n_batches) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
910 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
911 def __iter__(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
912 return self |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
913 |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
914 def next(self): |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
915 dataset=self.vsds.datasets[self.next_dataset_index] |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
916 mb = self.next_iterator.next() |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
917 if self.n_left_in_mb: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
918 extra_mb = [] |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
919 while self.n_left_in_mb>0: |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
920 self.move_to_next_dataset() |
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
921 extra_mb.append(self.next_iterator.next()) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
922 mb = Example(fieldnames, |
40
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
923 [dataset.valuesVStack(name, |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
924 [mb[name]]+[b[name] for b in extra_mb]) |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
925 for name in fieldnames]) |
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
926 |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
927 self.next_row+=minibatch_size |
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
928 self.next_dataset_row+=minibatch_size |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
929 if self.next_row+minibatch_size>len(dataset): |
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
930 self.move_to_next_dataset() |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
931 return examples |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
932 return VStackedIterator(self) |
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
933 |
41 | 934 class ArrayFieldsDataSet(DataSet): |
935 """ | |
936 Virtual super-class of datasets whose field values are numpy array, | |
937 thus defining valuesHStack and valuesVStack for sub-classes. | |
938 """ | |
939 def __init__(self,description=None,field_types=None): | |
940 DataSet.__init__(self,description,field_types) | |
941 def valuesHStack(self,fieldnames,fieldvalues): | |
942 """Concatenate field values horizontally, e.g. two vectors | |
943 become a longer vector, two matrices become a wider matrix, etc.""" | |
944 return numpy.hstack(fieldvalues) | |
945 def valuesVStack(self,fieldname,values): | |
946 """Concatenate field values vertically, e.g. two vectors | |
947 become a two-row matrix, two matrices become a longer matrix, etc.""" | |
948 return numpy.vstack(values) | |
949 | |
950 class ArrayDataSet(ArrayFieldsDataSet): | |
951 """ | |
952 An ArrayDataSet stores the fields as groups of columns in a numpy tensor, | |
953 whose first axis iterates over examples, second axis determines fields. | |
954 If the underlying array is N-dimensional (has N axes), then the field | |
955 values are (N-2)-dimensional objects (i.e. ordinary numbers if N=2). | |
956 """ | |
957 | |
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
958 def __init__(self, data_array, fields_columns, **kwargs): |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
959 """ |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
960 Construct an ArrayDataSet from the underlying numpy array (data) and |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
961 a map (fields_columns) from fieldnames to field columns. The columns of a field are specified |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
962 using the standard arguments for indexing/slicing: integer for a column index, |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
963 slice for an interval of columns (with possible stride), or iterable of column indices. |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
964 """ |
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
965 ArrayFieldsDataSet.__init__(self, **kwargs) |
41 | 966 self.data=data_array |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
967 self.fields_columns=fields_columns |
41 | 968 |
969 # check consistency and complete slices definitions | |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
970 for fieldname, fieldcolumns in self.fields_columns.items(): |
41 | 971 if type(fieldcolumns) is int: |
972 assert fieldcolumns>=0 and fieldcolumns<data_array.shape[1] | |
216
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
973 |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
974 if 0: |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
975 #I changed this because it didn't make sense to me, |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
976 # and it made it more difficult to write my learner. |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
977 # If it breaks stuff, let's talk about it. |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
978 # - James 22/05/2008 |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
979 self.fields_columns[fieldname]=[fieldcolumns] |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
980 else: |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
981 self.fields_columns[fieldname]=fieldcolumns |
4b7e89b75e2b
Modified ArrayDataSet's handling of column fields.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
211
diff
changeset
|
982 |
41 | 983 elif type(fieldcolumns) is slice: |
984 start,step=None,None | |
985 if not fieldcolumns.start: | |
986 start=0 | |
987 if not fieldcolumns.step: | |
988 step=1 | |
989 if start or step: | |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
990 self.fields_columns[fieldname]=slice(start,fieldcolumns.stop,step) |
41 | 991 elif hasattr(fieldcolumns,"__iter__"): # something like a list |
992 for i in fieldcolumns: | |
993 assert i>=0 and i<data_array.shape[1] | |
994 | |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
995 def fieldNames(self): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
996 return self.fields_columns.keys() |
41 | 997 |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
998 def __len__(self): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
999 return len(self.data) |
41 | 1000 |
80 | 1001 def __getitem__(self,key): |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1002 """More efficient implementation than the default __getitem__""" |
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1003 fieldnames=self.fields_columns.keys() |
80 | 1004 if type(key) is int: |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1005 return Example(fieldnames, |
80 | 1006 [self.data[key,self.fields_columns[f]] for f in fieldnames]) |
1007 if type(key) is slice: | |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1008 return MinibatchDataSet(Example(fieldnames, |
80 | 1009 [self.data[key,self.fields_columns[f]] for f in fieldnames])) |
1010 if type(key) is list: | |
1011 for i in range(len(key)): | |
1012 if self.hasFields(key[i]): | |
1013 key[i]=self.fields_columns[key[i]] | |
1014 return MinibatchDataSet(Example(fieldnames, | |
88
6749d18e11c8
bugfix as numpy numpy don't support self.data[[i1,...],[i2,...]] when their is more then two i1 and i2
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
85
diff
changeset
|
1015 #we must separate differently for list as numpy |
128 | 1016 # doesn't support self.data[[i1,...],[i2,...]] |
88
6749d18e11c8
bugfix as numpy numpy don't support self.data[[i1,...],[i2,...]] when their is more then two i1 and i2
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
85
diff
changeset
|
1017 # when their is more then two i1 and i2 |
6749d18e11c8
bugfix as numpy numpy don't support self.data[[i1,...],[i2,...]] when their is more then two i1 and i2
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
85
diff
changeset
|
1018 [self.data[key,:][:,self.fields_columns[f]] |
6749d18e11c8
bugfix as numpy numpy don't support self.data[[i1,...],[i2,...]] when their is more then two i1 and i2
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
85
diff
changeset
|
1019 if isinstance(self.fields_columns[f],list) else |
128 | 1020 self.data[key,self.fields_columns[f]] for f in fieldnames]), |
1021 | |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1022 self.valuesVStack,self.valuesHStack) |
80 | 1023 |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1024 # else check for a fieldname |
80 | 1025 if self.hasFields(key): |
105
8c0a1b11b007
bugfix, we keep all the line, but only a some columns
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
101
diff
changeset
|
1026 return self.data[:,self.fields_columns[key]] |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1027 # else we are trying to access a property of the dataset |
80 | 1028 assert key in self.__dict__ # else it means we are trying to access a non-existing property |
1029 return self.__dict__[key] | |
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1030 |
41 | 1031 |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1032 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1033 class ArrayDataSetIterator(object): |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1034 def __init__(self,dataset,fieldnames,minibatch_size,n_batches,offset): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1035 if fieldnames is None: fieldnames = dataset.fieldNames() |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1036 # store the resulting minibatch in a lookup-list of values |
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1037 self.minibatch = LookupList(fieldnames,[0]*len(fieldnames)) |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1038 self.dataset=dataset |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1039 self.minibatch_size=minibatch_size |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1040 assert offset>=0 and offset<len(dataset.data) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1041 assert offset+minibatch_size<=len(dataset.data) |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1042 self.current=offset |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1043 def __iter__(self): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1044 return self |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1045 def next(self): |
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1046 sub_data = self.dataset.data[self.current:self.current+self.minibatch_size] |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1047 self.minibatch._values = [sub_data[:,self.dataset.fields_columns[f]] for f in self.minibatch._names] |
43
e92244f30116
Corrected iterator logic errors
bengioy@grenat.iro.umontreal.ca
parents:
42
diff
changeset
|
1048 self.current+=self.minibatch_size |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1049 return self.minibatch |
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1050 |
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1051 return ArrayDataSetIterator(self,fieldnames,minibatch_size,n_batches,offset) |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1052 |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1053 |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1054 class CachedDataSet(DataSet): |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1055 """ |
167 | 1056 Wrap a L{DataSet} whose values are computationally expensive to obtain |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1057 (e.g. because they involve some computation, or disk access), |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1058 so that repeated accesses to the same example are done cheaply, |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1059 by caching every example value that has been accessed at least once. |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1060 |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1061 Optionally, for finite-length dataset, all the values can be computed |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1062 (and cached) upon construction of the CachedDataSet, rather at the |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1063 first access. |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1064 |
167 | 1065 @todo: when cache_all_upon_construction create mini-batches that are as |
77
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1066 large as possible but not so large as to fill up memory. |
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1067 |
167 | 1068 @todo: add disk-buffering capability, so that when the cache becomes too |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1069 big for memory, we cache things on disk, trying to keep in memory only |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1070 the record most likely to be accessed next. |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1071 """ |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1072 def __init__(self,source_dataset,cache_all_upon_construction=False): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1073 self.source_dataset=source_dataset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1074 self.cache_all_upon_construction=cache_all_upon_construction |
152
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1075 self.cached_examples = [] |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1076 if cache_all_upon_construction: |
77
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1077 # this potentially brings all the source examples |
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1078 # into memory at once, which may be too much |
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1079 # the work could possibly be done by minibatches |
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1080 # that are as large as possible but no more than what memory allows. |
146
8173e196e291
Trying to make CacheDataSet work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
144
diff
changeset
|
1081 fields_values = source_dataset.minibatches(minibatch_size=len(source_dataset)).__iter__().next() |
152
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1082 assert all([len(self)==len(field_values) for field_values in fields_values]) |
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1083 for example in fields_values.examples(): |
171
895b4b60f5e8
bugfix. Otherwise the example was writed over and not a new one was returned
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
167
diff
changeset
|
1084 self.cached_examples.append(copy.copy(example)) |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1085 |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1086 self.fieldNames = source_dataset.fieldNames |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1087 self.hasFields = source_dataset.hasFields |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1088 self.valuesHStack = source_dataset.valuesHStack |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1089 self.valuesVStack = source_dataset.valuesVStack |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1090 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1091 def __len__(self): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1092 return len(self.source_dataset) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1093 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1094 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1095 class CacheIterator(object): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1096 def __init__(self,dataset): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1097 self.dataset=dataset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1098 self.current=offset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1099 def __iter__(self): return self |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1100 def next(self): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1101 upper = self.current+minibatch_size |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1102 cache_len = len(self.dataset.cached_examples) |
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
1103 if upper>cache_len: # whole minibatch is not already in cache |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1104 # cache everything from current length to upper |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1105 for example in self.dataset.source_dataset[cache_len:upper]: |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1106 self.dataset.cached_examples.append(example) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1107 all_fields_minibatch = Example(self.dataset.fieldNames(), |
152
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1108 zip(*self.dataset.cached_examples[self.current:self.current+minibatch_size])) |
163
d7d67651d67c
bugfix, we should advence by the minibatch size.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
159
diff
changeset
|
1109 self.current+=minibatch_size |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1110 if self.dataset.fieldNames()==fieldnames: |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1111 return all_fields_minibatch |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1112 return Example(fieldnames,[all_fields_minibatch[name] for name in fieldnames]) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1113 return CacheIterator(self) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1114 |
153 | 1115 def __getitem__(self,i): |
1116 if type(i)==int and len(self.cached_examples)>i: | |
1117 return self.cached_examples[i] | |
1118 else: | |
1119 return DataSet.__getitem__(self,i) | |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1120 |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1121 class ApplyFunctionDataSet(DataSet): |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1122 """ |
167 | 1123 A L{DataSet} that contains as fields the results of applying a |
1124 given function example-wise or minibatch-wise to all the fields of | |
1125 an input dataset. The output of the function should be an iterable | |
1126 (e.g. a list or a LookupList) over the resulting values. | |
197
96d891448107
more explicit comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
196
diff
changeset
|
1127 |
96d891448107
more explicit comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
196
diff
changeset
|
1128 The function take as input the fields of the dataset, not the examples. |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1129 |
167 | 1130 In minibatch mode, the function is expected to work on minibatches |
1131 (takes a minibatch in input and returns a minibatch in output). More | |
1132 precisely, it means that each element of the input or output list | |
1133 should be iterable and indexable over the individual example values | |
1134 (typically these elements will be numpy arrays). All of the elements | |
1135 in the input and output lists should have the same length, which is | |
1136 the length of the minibatch. | |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1137 |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1138 The function is applied each time an example or a minibatch is accessed. |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1139 To avoid re-doing computation, wrap this dataset inside a CachedDataSet. |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1140 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1141 If the values_{h,v}stack functions are not provided, then |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1142 the input_dataset.values{H,V}Stack functions are used by default. |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1143 """ |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1144 def __init__(self,input_dataset,function,output_names,minibatch_mode=True, |
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
1145 values_hstack=None,values_vstack=None, |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1146 description=None,fieldtypes=None): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1147 """ |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1148 Constructor takes an input dataset that has as many fields as the function |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1149 expects as inputs. The resulting dataset has as many fields as the function |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1150 produces as outputs, and that should correspond to the number of output names |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1151 (provided in a list). |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1152 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1153 Note that the expected semantics of the function differs in minibatch mode |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1154 (it takes minibatches of inputs and produces minibatches of outputs, as |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1155 documented in the class comment). |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
1156 |
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
1157 TBM: are filedtypes the old field types (from input_dataset) or the new ones |
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
1158 (for the new dataset created)? |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1159 """ |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1160 self.input_dataset=input_dataset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1161 self.function=function |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1162 self.output_names=output_names |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1163 self.minibatch_mode=minibatch_mode |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1164 DataSet.__init__(self,description,fieldtypes) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1165 self.valuesHStack = values_hstack if values_hstack else input_dataset.valuesHStack |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1166 self.valuesVStack = values_vstack if values_vstack else input_dataset.valuesVStack |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1167 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1168 def __len__(self): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1169 return len(self.input_dataset) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1170 |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1171 def fieldNames(self): |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1172 return self.output_names |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1173 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1174 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1175 class ApplyFunctionIterator(object): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1176 def __init__(self,output_dataset): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1177 self.input_dataset=output_dataset.input_dataset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1178 self.output_dataset=output_dataset |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1179 self.input_iterator=self.input_dataset.minibatches(minibatch_size=minibatch_size, |
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1180 n_batches=n_batches,offset=offset).__iter__() |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1181 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1182 def __iter__(self): return self |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1183 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1184 def next(self): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1185 function_inputs = self.input_iterator.next() |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1186 all_output_names = self.output_dataset.output_names |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1187 if self.output_dataset.minibatch_mode: |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1188 function_outputs = self.output_dataset.function(*function_inputs) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1189 else: |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1190 input_examples = zip(*function_inputs) |
201 | 1191 output_examples = [self.output_dataset.function(*input_example) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1192 for input_example in input_examples] |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1193 function_outputs = [self.output_dataset.valuesVStack(name,values) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1194 for name,values in zip(all_output_names, |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1195 zip(*output_examples))] |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1196 all_outputs = Example(all_output_names,function_outputs) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1197 if fieldnames==all_output_names: |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1198 return all_outputs |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1199 return Example(fieldnames,[all_outputs[name] for name in fieldnames]) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1200 |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
203
diff
changeset
|
1201 |
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1202 return ApplyFunctionIterator(self) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1203 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1204 def __iter__(self): # only implemented for increased efficiency |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1205 class ApplyFunctionSingleExampleIterator(object): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1206 def __init__(self,output_dataset): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1207 self.current=0 |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1208 self.output_dataset=output_dataset |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1209 self.input_iterator=output_dataset.input_dataset.__iter__() |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1210 def __iter__(self): return self |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1211 def next(self): |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1212 if self.output_dataset.minibatch_mode: |
200 | 1213 function_inputs = [[input] for input in self.input_iterator.next()] |
199 | 1214 outputs = self.output_dataset.function(*function_inputs) |
1215 assert all([hasattr(output,'__iter__') for output in outputs]) | |
1216 function_outputs = [output[0] for output in outputs] | |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1217 else: |
200 | 1218 function_inputs = self.input_iterator.next() |
199 | 1219 function_outputs = self.output_dataset.function(*function_inputs) |
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1220 return Example(self.output_dataset.output_names,function_outputs) |
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1221 return ApplyFunctionSingleExampleIterator(self) |
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1222 |
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1223 |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1224 def supervised_learning_dataset(src_dataset,input_fields,target_fields,weight_field=None): |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1225 """ |
167 | 1226 Wraps an arbitrary L{DataSet} into one for supervised learning tasks |
1227 by forcing the user to define a set of fields as the 'input' field | |
1228 and a set of fields as the 'target' field. Optionally, a single | |
1229 weight_field can also be defined. | |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1230 """ |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1231 args = ((input_fields,'input'),(output_fields,'target')) |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1232 if weight_field: args+=(([weight_field],'weight')) |
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
1233 return src_dataset.merge_fields(*args) |
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1234 |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1235 |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1236 |
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1237 |