Mercurial > pylearn
annotate dataset.py @ 192:f62a03c9d485
Redesign of StatsCollector
| author | Yoshua Bengio <bengioy@iro.umontreal.ca> |
|---|---|
| date | Thu, 15 May 2008 12:46:21 -0400 |
| parents | f01ac276c6fb |
| children | cb6b945acf5a |
| 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) |
|
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
29 |
|
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
30 |
|
8fa1ef2411a0
Worked on OneShotTLearner and implementation of LinearRegression
bengioy@bengiomac.local
parents:
105
diff
changeset
|
31 class DataSet(AttributesHolder): |
| 16 | 32 """A virtual base class for datasets. |
| 33 | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
34 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
|
35 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
|
36 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
|
37 python object, which depends on the particular dataset. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
38 |
|
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
|
39 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
|
40 should return sys.maxint). |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
41 |
| 16 | 42 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
|
43 examples or the fields in a variety of ways. A DataSet need not necessarily have a finite |
| 16 | 44 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
|
45 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
|
46 feasible or not recommanded on streams). |
| 16 | 47 |
| 48 To iterate over examples, there are several possibilities: | |
|
90
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
49 - for example in dataset: |
|
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
50 - for val1,val2,... in dataset: |
|
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
51 - for example in dataset(field1, field2,field3, ...): |
|
a289b8bed64c
corrected comment
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
88
diff
changeset
|
52 - for val1,val2,val3 in dataset(field1, field2,field3): |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
53 - for minibatch in dataset.minibatches([field1, field2, ...],minibatch_size=N): |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
54 - 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
|
55 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
|
56 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
|
57 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
|
58 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
|
59 next minibatch returned). This is important because these iterators |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
60 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
|
61 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
|
62 dataset length. |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
63 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
64 To iterate over fields, one can do |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
65 - for field in dataset.fields(): |
|
46
c5b07e87b0cb
comments modif made by Yoshua
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
45
diff
changeset
|
66 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
|
67 - 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
|
68 - 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
|
69 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
|
70 - for field_examples in dataset.fields(): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
71 for example_value in field_examples: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
72 ... |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
73 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
|
74 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
|
75 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
|
76 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
|
77 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
|
78 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
|
79 dataset2 = dataset1.fields().examples() |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
80 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
|
81 |
| 16 | 82 Note: Fields are not mutually exclusive, i.e. two fields can overlap in their actual content. |
| 83 | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
84 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
|
85 (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
|
86 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
|
87 What about non-numeric values? None. |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
88 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
89 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
|
90 of examples) can be extracted. These operations are not supported |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
91 by default in the case of streams. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
92 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
93 - dataset[:n] returns a dataset with the n first examples. |
| 16 | 94 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
95 - 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
|
96 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
97 - dataset[i] returns an Example. |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
98 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
99 - 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
|
100 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
101 - 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
|
102 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
|
103 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
|
104 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
105 - dataset.<property> returns the value of a property associated with |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
106 the name <property>. The following properties should be supported: |
| 41 | 107 - '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
|
108 - 'fieldtypes': a list of types (one per field) |
| 78 | 109 A DataSet may have other attributes that it makes visible to other objects. These are |
| 110 used to store information that is not example-wise but global to the dataset. | |
| 111 The list of names of these attributes is given by the attribute_names() method. | |
| 41 | 112 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
113 Datasets can be concatenated either vertically (increasing the length) or |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
114 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
|
115 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
|
116 and numpy.vstack): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
117 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
118 - 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
|
119 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
120 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
|
121 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
|
122 |
|
72
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
123 - dataset1 & dataset2 & dataset3 == dataset.vstack([dataset1,dataset2,dataset3]) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
124 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
125 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
|
126 (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
|
127 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
|
128 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
129 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
|
130 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
|
131 a DataSetFields fields1 and fields2, and fields1 | fields2 concatenates their |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
132 examples. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
133 |
| 41 | 134 A dataset can hold arbitrary key-value pairs that may be used to access meta-data |
| 135 or other properties of the dataset or associated with the dataset or the result | |
| 136 of a computation stored in a dataset. These can be accessed through the [key] syntax | |
| 137 when key is a string (or more specifically, neither an integer, a slice, nor a list). | |
| 78 | 138 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
139 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
|
140 - __len__ if it is not a stream |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
141 - fieldNames |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
142 - minibatches_nowrap (called by DataSet.minibatches()) |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
143 - valuesHStack |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
144 - valuesVStack |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
145 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
|
146 - hasFields |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
147 - __getitem__ may not be feasible with some streams |
|
2b6656b2ef52
Changed docs slightly
Joseph Turian <turian@iro.umontreal.ca>
parents:
66
diff
changeset
|
148 - __iter__ |
| 78 | 149 A sub-class should also append attributes to self._attribute_names |
| 150 (the default value returned by attributeNames()). | |
| 151 By convention, attributes not in attributeNames() should have a name | |
| 152 starting with an underscore. | |
| 153 @todo enforce/test that convention! | |
|
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
154 """ |
|
1
2cd82666b9a7
Added statscollector and started writing dataset and learner.
bengioy@esprit.iro.umontreal.ca
parents:
0
diff
changeset
|
155 |
| 83 | 156 numpy_vstack = lambda fieldname,values: numpy.vstack(values) |
| 157 numpy_hstack = lambda fieldnames,values: numpy.hstack(values) | |
|
77
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
158 |
|
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
159 def __init__(self,description=None,fieldtypes=None): |
| 41 | 160 if description is None: |
| 161 # by default return "<DataSetType>(<SuperClass1>,<SuperClass2>,...)" | |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
162 description = type(self).__name__ + " ( " + join([x.__name__ for x in type(self).__bases__]) + " )" |
| 41 | 163 self.description=description |
| 60 | 164 self.fieldtypes=fieldtypes |
| 78 | 165 self._attribute_names = ["description"] |
| 166 if fieldtypes: | |
| 167 self._attribute_names.append("fieldtypes") | |
| 168 | |
| 169 def attributeNames(self): return self._attribute_names | |
| 170 | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
171 class MinibatchToSingleExampleIterator(object): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
172 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
173 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
|
174 single-example values in the result. Therefore the result of |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
175 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
|
176 (whereas the result of iterating over minibatches gives in each |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
177 Example field an iterable object over the individual examples in |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
178 the minibatch). |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
179 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
180 def __init__(self, minibatch_iterator): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
181 self.minibatch_iterator = minibatch_iterator |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
182 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
|
183 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
|
184 return self |
|
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
185 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
|
186 size1_minibatch = self.minibatch_iterator.next() |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
187 if not self.minibatch: |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
188 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
|
189 else: |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
190 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
|
191 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
|
192 |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
193 def next_index(self): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
194 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
|
195 |
|
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
196 def __iter__(self): |
| 16 | 197 """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
|
198 |
| 16 | 199 Using this syntax, "i" will be an Example instance (or equivalent) with |
| 200 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
|
201 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
|
202 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
|
203 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
|
204 to accept any type of identifier, and add extra functionality to the iterator. |
| 16 | 205 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
206 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
|
207 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
208 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
|
209 |
|
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
210 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
|
211 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
|
212 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
|
213 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
214 class MinibatchWrapAroundIterator(object): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
215 """ |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
216 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
|
217 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
|
218 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
|
219 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
|
220 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
|
221 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
|
222 """ |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
223 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
|
224 self.dataset=dataset |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
225 self.fieldnames=fieldnames |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
226 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
|
227 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
|
228 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
|
229 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
|
230 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
|
231 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
|
232 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
|
233 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
|
234 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
|
235 ds_nbatches = min(n_batches,ds_nbatches) |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
236 if fieldnames: |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
237 assert dataset.hasFields(*fieldnames) |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
238 else: |
|
98
7186e4f502d1
bugfix in DataSet.minibatch to correctly wrap in all corner case.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
95
diff
changeset
|
239 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
|
240 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
|
241 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
|
242 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
243 def __iter__(self): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
244 return self |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
245 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
246 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
|
247 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
|
248 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
249 def next(self): |
|
43
e92244f30116
Corrected iterator logic errors
bengioy@grenat.iro.umontreal.ca
parents:
42
diff
changeset
|
250 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
|
251 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
|
252 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
|
253 raise StopIteration |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
254 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
|
255 if upper <=self.L: |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
256 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
|
257 else: |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
258 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
|
259 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
|
260 # 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
|
261 # 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
|
262 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
|
263 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
|
264 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
|
265 # 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
|
266 # 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
|
267 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
|
268 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
|
269 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
|
270 [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
|
271 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
|
272 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
|
273 self.n_batches_done+=1 |
|
43
e92244f30116
Corrected iterator logic errors
bengioy@grenat.iro.umontreal.ca
parents:
42
diff
changeset
|
274 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
|
275 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
|
276 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
|
277 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
|
278 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
|
279 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
|
280 ds_nbatches,self.next_row) |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
281 return DataSetFields(MinibatchDataSet(minibatch,self.dataset.valuesVStack, |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
282 self.dataset.valuesHStack), |
|
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
283 minibatch.keys()) |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
284 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
285 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
286 minibatches_fieldnames = None |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
287 minibatches_minibatch_size = 1 |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
288 minibatches_n_batches = None |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
289 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
|
290 fieldnames = minibatches_fieldnames, |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
291 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
|
292 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
|
293 offset = 0): |
|
6
d5738b79089a
Removed MinibatchIterator and instead made minibatch_size a field of all DataSets,
bengioy@bengiomac.local
parents:
5
diff
changeset
|
294 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
295 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
|
296 |
|
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
297 for i in dataset.minibatches(None,**kwargs): ... |
| 16 | 298 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
299 for i in dataset.minibatches([f1, f2, f3],**kwargs): ... |
| 16 | 300 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
301 for i1, i2, i3 in dataset.minibatches([f1, f2, f3],**kwargs): ... |
| 16 | 302 |
|
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
303 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
|
304 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
|
305 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
|
306 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
|
307 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
|
308 |
|
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
309 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
|
310 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
|
311 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
|
312 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
313 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
|
314 a DataSetFields object, which is a LookupList (indexed by the field names) whose |
| 80 | 315 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
|
316 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
|
317 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
|
318 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
|
319 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
320 PARAMETERS |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
321 - 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
|
322 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
|
323 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
|
324 derived class can choose a default, e.g. all fields. |
| 16 | 325 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
326 - minibatch_size (integer, default 1) |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
327 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
|
328 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
|
329 |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
330 - n_batches (integer, default None) |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
331 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
|
332 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
|
333 iterator should support looping indefinitely. |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
334 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
335 - 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
|
336 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
|
337 |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
338 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
|
339 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
|
340 |
|
11
be128b9127c8
Debugged (to the extent of my tests) the new version of dataset
bengioy@esprit.iro.umontreal.ca
parents:
9
diff
changeset
|
341 """ |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
342 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
|
343 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
344 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
|
345 """ |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
346 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
|
347 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
|
348 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
|
349 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
|
350 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
|
351 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
|
352 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
|
353 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
|
354 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
|
355 """ |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
356 raise AbstractFunction() |
|
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
357 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
358 def __len__(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
359 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
360 len(dataset) returns the number of examples in the dataset. |
|
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
361 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
|
362 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
|
363 Some methods only make sense for finite-length datasets. |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
364 """ |
| 123 | 365 return maxint |
|
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
366 |
|
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
367 def is_unbounded(self): |
|
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
368 """ |
|
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
369 Tests whether a dataset is unbounded (e.g. a stream). |
|
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
370 """ |
| 123 | 371 return len(self)==maxint |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
372 |
|
26
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
373 def hasFields(self,*fieldnames): |
|
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
374 """ |
|
22
b6b36f65664f
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet,
bengioy@esprit.iro.umontreal.ca
parents:
20
diff
changeset
|
375 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
|
376 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
|
377 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
|
378 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
379 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
|
380 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
|
381 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
382 return len(unique_elements_list_intersection(fieldnames,self.fieldNames()))>0 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
383 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
384 def fieldNames(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
385 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
386 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
|
387 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
|
388 """ |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
389 raise AbstractFunction() |
|
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
390 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
391 def __call__(self,*fieldnames): |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
392 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
393 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
|
394 """ |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
395 assert self.hasFields(*fieldnames) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
396 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
|
397 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
398 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
|
399 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
400 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
|
401 """ |
|
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
402 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
|
403 |
|
2
3fddb1c8f955
Rewrote DataSet interface and created FiniteDataSet interface.
bengioy@bengiomac.local
parents:
1
diff
changeset
|
404 def __getitem__(self,i): |
|
28
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
405 """ |
|
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
406 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
|
407 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
|
408 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
|
409 dataset[[i1,i2,..,in]] returns the subdataset with examples i1,i2,...,in. |
| 41 | 410 dataset['key'] returns a property associated with the given 'key' string. |
| 411 If 'key' is a fieldname, then the VStacked field values (iterable over | |
| 412 field values) for that field is returned. Other keys may be supported | |
| 413 by different dataset subclasses. The following key names are encouraged: | |
| 414 - 'description': a textual description or name for the dataset | |
| 415 - '<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
|
416 |
| 39 | 417 Note that some stream datasets may be unable to implement random access, i.e. |
| 418 arbitrary slicing/indexing | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
419 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
|
420 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
|
421 |
|
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
422 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
|
423 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
|
424 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
|
425 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
|
426 """ |
| 41 | 427 # 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
|
428 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
|
429 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
|
430 self.minibatches(minibatch_size=1,n_batches=1,offset=i)).next() |
| 41 | 431 rows=None |
| 432 # 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
|
433 if type(i) is slice: |
|
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
434 if not i.start: i=slice(0,i.stop,i.step) |
|
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
435 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
|
436 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
|
437 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
|
438 rows = range(i.start,i.stop,i.step) |
| 41 | 439 # or a list of indices |
| 440 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
|
441 rows = i |
| 41 | 442 if rows is not None: |
|
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
443 examples = [self[row] for row in rows] |
|
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
444 fields_values = zip(*examples) |
|
45
a5c70dc42972
Test functions for dataset.py
bengioy@grenat.iro.umontreal.ca
parents:
44
diff
changeset
|
445 return MinibatchDataSet( |
| 41 | 446 Example(self.fieldNames(),[ self.valuesVStack(fieldname,field_values) |
| 447 for fieldname,field_values | |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
448 in zip(self.fieldNames(),fields_values)]), |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
449 self.valuesVStack,self.valuesHStack) |
| 41 | 450 # else check for a fieldname |
| 451 if self.hasFields(i): | |
| 452 return self.minibatches(fieldnames=[i],minibatch_size=len(self),n_batches=1,offset=0).next()[0] | |
| 453 # else we are trying to access a property of the dataset | |
| 454 assert i in self.__dict__ # else it means we are trying to access a non-existing property | |
| 455 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
|
456 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
457 def valuesHStack(self,fieldnames,fieldvalues): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
458 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
459 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
|
460 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
|
461 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
|
462 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
|
463 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
|
464 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
|
465 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
|
466 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
|
467 pointing to the original values for other data types. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
468 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
469 all_numpy=True |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
470 for value in fieldvalues: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
471 if not type(value) is numpy.ndarray: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
472 all_numpy=False |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
473 if all_numpy: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
474 return numpy.hstack(fieldvalues) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
475 # 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
|
476 return fieldvalues |
|
26
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
477 |
|
672fe4b23032
Fixed dataset errors so that _test_dataset.py works again.
bengioy@grenat.iro.umontreal.ca
parents:
23
diff
changeset
|
478 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
479 def valuesVStack(self,fieldname,values): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
480 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
481 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
|
482 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
|
483 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
|
484 result should be numpy.vstack(values). |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
485 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
|
486 pointing to the original values for other data types. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
487 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
488 all_numpy=True |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
489 for value in values: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
490 if not type(value) is numpy.ndarray: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
491 all_numpy=False |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
492 if all_numpy: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
493 return numpy.vstack(values) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
494 # 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
|
495 return values |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
496 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
497 def __or__(self,other): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
498 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
499 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
|
500 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
|
501 """ |
|
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
502 return HStackedDataSet([self,other]) |
|
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
503 |
|
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
|
504 def __and__(self,other): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
505 """ |
|
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
|
506 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
|
507 (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
|
508 works if they all have the same fields. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
509 """ |
|
135
0d8e721cc63c
Fixed bugs in dataset to make test_mlp.py work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
134
diff
changeset
|
510 return VStackedDataSet([self,other]) |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
511 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
512 def hstack(datasets): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
513 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
514 hstack(dataset1,dataset2,...) returns dataset1 | datataset2 | ... |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
515 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
|
516 of the individual datasets. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
517 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
518 assert len(datasets)>0 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
519 if len(datasets)==1: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
520 return datasets[0] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
521 return HStackedDataSet(datasets) |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
522 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
523 def vstack(datasets): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
524 """ |
|
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
|
525 vstack(dataset1,dataset2,...) returns dataset1 & datataset2 & ... |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
526 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
|
527 over those of dataset2, etc. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
528 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
529 assert len(datasets)>0 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
530 if len(datasets)==1: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
531 return datasets[0] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
532 return VStackedDataSet(datasets) |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
533 |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
534 class FieldsSubsetDataSet(DataSet): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
535 """ |
| 167 | 536 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
|
537 """ |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
538 def __init__(self,src,fieldnames): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
539 self.src=src |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
540 self.fieldnames=fieldnames |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
541 assert src.hasFields(*fieldnames) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
542 self.valuesHStack = src.valuesHStack |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
543 self.valuesVStack = src.valuesVStack |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
544 |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
545 def __len__(self): return len(self.src) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
546 |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
547 def fieldNames(self): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
548 return self.fieldnames |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
549 |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
550 def __iter__(self): |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
551 class FieldsSubsetIterator(object): |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
552 def __init__(self,ds): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
553 self.ds=ds |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
554 self.src_iter=ds.src.__iter__() |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
555 self.example=None |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
556 def __iter__(self): return self |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
557 def next(self): |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
558 complete_example = self.src_iter.next() |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
559 if self.example: |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
560 self.example._values=[complete_example[field] |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
561 for field in self.ds.fieldnames] |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
562 else: |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
563 self.example=Example(self.ds.fieldnames, |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
564 [complete_example[field] for field in self.ds.fieldnames]) |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
565 return self.example |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
566 return FieldsSubsetIterator(self) |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
567 |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
568 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
|
569 assert self.hasFields(*fieldnames) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
570 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
|
571 def __getitem__(self,i): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
572 return FieldsSubsetDataSet(self.src[i],self.fieldnames) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
573 |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
574 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
575 class DataSetFields(LookupList): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
576 """ |
| 167 | 577 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
|
578 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
|
579 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
|
580 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
581 To iterate over fields, one can do |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
582 * for fields in dataset.fields() |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
583 * 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
|
584 * 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
|
585 and each of these fields is iterable over the examples: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
586 * for field_examples in dataset.fields(): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
587 for example_value in field_examples: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
588 ... |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
589 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
|
590 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
|
591 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
|
592 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
|
593 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
|
594 be turned back into a DataSet with its examples() method: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
595 dataset2 = dataset1.fields().examples() |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
596 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
|
597 |
|
88fd1cce08b9
replaced infinity for length by raise UnboundedDataSet and use & instead of + to concatenate datasets
bengioy@esprit.iro.umontreal.ca
parents:
39
diff
changeset
|
598 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
|
599 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
|
600 the examples. |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
601 """ |
|
74
b4159cbdc06b
Fixed errors raised by test_dataset
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
73
diff
changeset
|
602 def __init__(self,dataset,fieldnames): |
|
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
603 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
|
604 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
|
605 fieldnames=dataset.fieldNames() |
|
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
606 elif not fieldnames==dataset.fieldNames(): |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
607 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
|
608 assert dataset.hasFields(*fieldnames) |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
609 self.dataset=dataset |
|
66
dde1fb1b63ba
fixed test and removed print
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
65
diff
changeset
|
610 |
|
64
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
611 if isinstance(dataset,MinibatchDataSet): |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
612 LookupList.__init__(self,fieldnames,list(dataset._fields)) |
|
65
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
613 elif isinstance(original_dataset,MinibatchDataSet): |
|
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
614 LookupList.__init__(self,fieldnames, |
|
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
615 [original_dataset._fields[field] |
|
d48eba49a2f4
fixed the infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
64
diff
changeset
|
616 for field in fieldnames]) |
|
64
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
617 else: |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
618 minibatch_iterator = dataset.minibatches(fieldnames, |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
619 minibatch_size=len(dataset), |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
620 n_batches=1) |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
621 minibatch=minibatch_iterator.next() |
|
863da25a60f1
trying to fix infinite loop
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
62
diff
changeset
|
622 LookupList.__init__(self,fieldnames,minibatch) |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
623 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
624 def examples(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
625 return self.dataset |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
626 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
627 def __or__(self,other): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
628 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
629 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
|
630 of the list of examples of DataSetFields fields1 and fields2. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
631 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
632 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
|
633 |
|
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
|
634 def __and__(self,other): |
|
17
759d17112b23
more comments, looping ArrayDataSet iterator, bugfixes to lookup_list, more tests
bergstrj@iro.umontreal.ca
diff
changeset
|
635 """ |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
636 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
|
637 of the fields of DataSetFields fields1 and fields2. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
638 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
639 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
|
640 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
641 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
642 class MinibatchDataSet(DataSet): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
643 """ |
| 167 | 644 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
|
645 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
|
646 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
647 def __init__(self,fields_lookuplist,values_vstack=DataSet().valuesVStack, |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
648 values_hstack=DataSet().valuesHStack): |
|
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 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
|
651 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
|
652 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
|
653 """ |
|
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
|
654 self._fields=fields_lookuplist |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
655 assert len(fields_lookuplist)>0 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
656 self.length=len(fields_lookuplist[0]) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
657 for field in fields_lookuplist[1:]: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
658 assert self.length==len(field) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
659 self.values_vstack=values_vstack |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
660 self.values_hstack=values_hstack |
|
3
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
661 |
|
378b68d5c4ad
Added first (untested) version of ArrayDataSet
bengioy@bengiomac.local
parents:
2
diff
changeset
|
662 def __len__(self): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
663 return self.length |
|
28
541a273bc89f
Removed __array__ method from dataset, whose
bengioy@grenat.iro.umontreal.ca
parents:
26
diff
changeset
|
664 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
665 def __getitem__(self,i): |
| 80 | 666 if type(i) in (slice,list): |
|
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
667 return DataSetFields(MinibatchDataSet( |
| 80 | 668 Example(self._fields.keys(),[field[i] for field in self._fields])),self.fieldNames()) |
| 669 if type(i) is int: | |
| 85 | 670 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
|
671 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
|
672 return self._fields[i] |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
673 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
|
674 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
|
675 |
|
29
46c5c90019c2
Changed apply_function so that it propagates methods of the source.
bengioy@grenat.iro.umontreal.ca
parents:
28
diff
changeset
|
676 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
|
677 return self._fields.keys() |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
678 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
679 def hasFields(self,*fieldnames): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
680 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
|
681 if fieldname not in self._fields.keys(): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
682 return False |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
683 return True |
|
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
684 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
685 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
|
686 class Iterator(object): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
687 def __init__(self,ds): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
688 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
|
689 self.next_example=offset |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
690 assert minibatch_size > 0 |
| 41 | 691 if offset+minibatch_size > ds.length: |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
692 raise NotImplementedError() |
|
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
693 def __iter__(self): |
|
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
694 return self |
|
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
695 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
|
696 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
|
697 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
|
698 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
|
699 [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
|
700 for field in self.ds._fields]) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
701 self.next_example+=minibatch_size |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
702 return minibatch |
|
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
703 |
|
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
|
704 return Iterator(self) |
|
20
266c68cb6136
Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents:
19
diff
changeset
|
705 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
706 def valuesVStack(self,fieldname,fieldvalues): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
707 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
|
708 |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
709 def valuesHStack(self,fieldnames,fieldvalues): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
710 return self.values_hstack(fieldnames,fieldvalues) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
711 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
712 class HStackedDataSet(DataSet): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
713 """ |
| 167 | 714 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
|
715 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
|
716 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
717 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
|
718 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
|
719 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
|
720 |
| 167 | 721 @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
|
722 """ |
| 41 | 723 def __init__(self,datasets,accept_nonunique_names=False,description=None,field_types=None): |
| 724 DataSet.__init__(self,description,field_types) | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
725 self.datasets=datasets |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
726 self.accept_nonunique_names=accept_nonunique_names |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
727 self.fieldname2dataset={} |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
728 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
729 def rename_field(fieldname,dataset,i): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
730 if hasattr(dataset,"__name__"): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
731 return dataset.__name__ + "." + fieldname |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
732 return fieldname+"."+str(i) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
733 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
734 # 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
|
735 self.length=None |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
736 names_to_change=[] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
737 for i in xrange(len(datasets)): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
738 dataset = datasets[i] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
739 length=len(dataset) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
740 if self.length: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
741 assert self.length==length |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
742 else: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
743 self.length=length |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
744 for fieldname in dataset.fieldNames(): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
745 if fieldname in self.fieldname2dataset: # name conflict! |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
746 if accept_nonunique_names: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
747 fieldname=rename_field(fieldname,dataset,i) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
748 names2change.append((fieldname,i)) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
749 else: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
750 raise ValueError("Incompatible datasets: non-unique field name = "+fieldname) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
751 self.fieldname2dataset[fieldname]=i |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
752 for fieldname,i in names_to_change: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
753 del self.fieldname2dataset[fieldname] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
754 self.fieldname2dataset[rename_field(fieldname,self.datasets[i],i)]=i |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
755 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
756 def hasFields(self,*fieldnames): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
757 for fieldname in fieldnames: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
758 if not fieldname in self.fieldname2dataset: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
759 return False |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
760 return True |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
761 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
762 def fieldNames(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
763 return self.fieldname2dataset.keys() |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
764 |
| 41 | 765 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
|
766 |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
767 class HStackedIterator(object): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
768 def __init__(self,hsds,iterators): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
769 self.hsds=hsds |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
770 self.iterators=iterators |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
771 def __iter__(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
772 return self |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
773 def next(self): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
774 # concatenate all the fields of the minibatches |
| 140 | 775 l=LookupList() |
| 776 for iter in self.iterators: | |
| 777 l.append_lookuplist(iter.next()) | |
| 778 return l | |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
779 |
| 125 | 780 assert self.hasFields(*fieldnames) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
781 # 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
|
782 # and construct corresponding minibatch iterators |
| 140 | 783 if fieldnames and fieldnames!=self.fieldNames(): |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
784 datasets=set([]) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
785 fields_in_dataset=dict([(dataset,[]) for dataset in datasets]) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
786 for fieldname in fieldnames: |
| 136 | 787 dataset=self.datasets[self.fieldname2dataset[fieldname]] |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
788 datasets.add(dataset) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
789 fields_in_dataset[dataset].append(fieldname) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
790 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
|
791 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
|
792 for dataset in datasets] |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
793 else: |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
794 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
|
795 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
|
796 return HStackedIterator(self,iterators) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
797 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
798 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
799 def valuesVStack(self,fieldname,fieldvalues): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
800 return self.datasets[self.fieldname2dataset[fieldname]].valuesVStack(fieldname,fieldvalues) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
801 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
802 def valuesHStack(self,fieldnames,fieldvalues): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
803 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
804 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
|
805 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
|
806 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
|
807 all the fieldnames are of the same sub-dataset. |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
808 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
809 return self.datasets[self.fieldname2dataset[fieldnames[0]]].valuesHStack(fieldnames,fieldvalues) |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
810 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
811 class VStackedDataSet(DataSet): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
812 """ |
| 167 | 813 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
|
814 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
|
815 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
|
816 |
| 167 | 817 @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
|
818 """ |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
819 def __init__(self,datasets): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
820 self.datasets=datasets |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
821 self.length=0 |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
822 self.index2dataset={} |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
823 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
|
824 fieldnames = datasets[-1].fieldNames() |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
825 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
|
826 # 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
|
827 # 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
|
828 for dataset,k in enumerate(datasets[0:-1]): |
|
48
b6730f9a336d
Fixing MinibatchDataSet getitem
bengioy@grenat.iro.umontreal.ca
parents:
46
diff
changeset
|
829 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
|
830 L=len(dataset) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
831 for i in xrange(L): |
|
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
832 self.index2dataset[self.length+i]=k |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
833 self.datasets_start_row.append(self.length) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
834 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
|
835 assert dataset.fieldNames()==fieldnames |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
836 self.datasets_start_row.append(self.length) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
837 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
|
838 # 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
|
839 # 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
|
840 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
|
841 # 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
|
842 # TODO |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
843 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
|
844 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
845 def __len__(self): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
846 return self.length |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
847 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
848 def fieldNames(self): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
849 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
|
850 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
851 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
|
852 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
|
853 |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
854 def locate_row(self,row): |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
855 """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
|
856 dataset_index = self.index2dataset[row] |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
857 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
|
858 return dataset_index, row_within_dataset |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
859 |
| 41 | 860 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
|
861 |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
862 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
|
863 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
|
864 self.vsds=vsds |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
865 self.next_row=offset |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
866 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
|
867 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
|
868 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
|
869 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
870 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
|
871 L=len(dataset) |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
872 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
|
873 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
|
874 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
|
875 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
|
876 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
|
877 n_left_in_mb=minibatch_size-L |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
878 ds_nbatches=1 |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
879 else: |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
880 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
|
881 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
|
882 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
|
883 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
884 def move_to_next_dataset(self): |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
885 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
|
886 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
|
887 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
|
888 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
|
889 else: |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
890 self.next_dataset_index +=1 |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
891 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
|
892 self.next_dataset_index = 0 |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
893 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
|
894 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
|
895 |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
896 def __iter__(self): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
897 return self |
|
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 next(self): |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
900 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
|
901 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
|
902 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
|
903 extra_mb = [] |
|
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
904 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
|
905 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
|
906 extra_mb.append(self.next_iterator.next()) |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
907 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
|
908 [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
|
909 [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
|
910 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
|
911 |
|
37
73c4212ba5b3
Factored the minibatch-writing code into an iterator class inside DataSet
bengioy@esprit.iro.umontreal.ca
parents:
36
diff
changeset
|
912 self.next_row+=minibatch_size |
|
38
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
913 self.next_dataset_row+=minibatch_size |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
914 if self.next_row+minibatch_size>len(dataset): |
|
d637ad8f7352
Finished first untested version of VStackedDataset
bengioy@esprit.iro.umontreal.ca
parents:
37
diff
changeset
|
915 self.move_to_next_dataset() |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
916 return examples |
|
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
917 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
|
918 |
| 41 | 919 class ArrayFieldsDataSet(DataSet): |
| 920 """ | |
| 921 Virtual super-class of datasets whose field values are numpy array, | |
| 922 thus defining valuesHStack and valuesVStack for sub-classes. | |
| 923 """ | |
| 924 def __init__(self,description=None,field_types=None): | |
| 925 DataSet.__init__(self,description,field_types) | |
| 926 def valuesHStack(self,fieldnames,fieldvalues): | |
| 927 """Concatenate field values horizontally, e.g. two vectors | |
| 928 become a longer vector, two matrices become a wider matrix, etc.""" | |
| 929 return numpy.hstack(fieldvalues) | |
| 930 def valuesVStack(self,fieldname,values): | |
| 931 """Concatenate field values vertically, e.g. two vectors | |
| 932 become a two-row matrix, two matrices become a longer matrix, etc.""" | |
| 933 return numpy.vstack(values) | |
| 934 | |
| 935 class ArrayDataSet(ArrayFieldsDataSet): | |
| 936 """ | |
| 937 An ArrayDataSet stores the fields as groups of columns in a numpy tensor, | |
| 938 whose first axis iterates over examples, second axis determines fields. | |
| 939 If the underlying array is N-dimensional (has N axes), then the field | |
| 940 values are (N-2)-dimensional objects (i.e. ordinary numbers if N=2). | |
| 941 """ | |
| 942 | |
|
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
943 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
|
944 """ |
|
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
945 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
|
946 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
|
947 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
|
948 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
|
949 """ |
|
188
f01ac276c6fb
added __contains__ to Dataset, added parent constructor call to ArrayDataSet
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
171
diff
changeset
|
950 ArrayFieldsDataSet.__init__(self, **kwargs) |
| 41 | 951 self.data=data_array |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
952 self.fields_columns=fields_columns |
| 41 | 953 |
| 954 # check consistency and complete slices definitions | |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
955 for fieldname, fieldcolumns in self.fields_columns.items(): |
| 41 | 956 if type(fieldcolumns) is int: |
| 957 assert fieldcolumns>=0 and fieldcolumns<data_array.shape[1] | |
|
99
a8da709eb6a9
in ArrayDataSet.__init__ if a columns is an index, we change it to be a list that containt only this index. This way, we remove the special case where the columns is an index for all subsequent call.
Frederic Bastien <bastienf@iro.umontreal.ca>
parents:
98
diff
changeset
|
958 self.fields_columns[fieldname]=[fieldcolumns] |
| 41 | 959 elif type(fieldcolumns) is slice: |
| 960 start,step=None,None | |
| 961 if not fieldcolumns.start: | |
| 962 start=0 | |
| 963 if not fieldcolumns.step: | |
| 964 step=1 | |
| 965 if start or step: | |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
966 self.fields_columns[fieldname]=slice(start,fieldcolumns.stop,step) |
| 41 | 967 elif hasattr(fieldcolumns,"__iter__"): # something like a list |
| 968 for i in fieldcolumns: | |
| 969 assert i>=0 and i<data_array.shape[1] | |
| 970 | |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
971 def fieldNames(self): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
972 return self.fields_columns.keys() |
| 41 | 973 |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
974 def __len__(self): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
975 return len(self.data) |
| 41 | 976 |
| 80 | 977 def __getitem__(self,key): |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
978 """More efficient implementation than the default __getitem__""" |
|
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
979 fieldnames=self.fields_columns.keys() |
| 80 | 980 if type(key) is int: |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
981 return Example(fieldnames, |
| 80 | 982 [self.data[key,self.fields_columns[f]] for f in fieldnames]) |
| 983 if type(key) is slice: | |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
984 return MinibatchDataSet(Example(fieldnames, |
| 80 | 985 [self.data[key,self.fields_columns[f]] for f in fieldnames])) |
| 986 if type(key) is list: | |
| 987 for i in range(len(key)): | |
| 988 if self.hasFields(key[i]): | |
| 989 key[i]=self.fields_columns[key[i]] | |
| 990 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
|
991 #we must separate differently for list as numpy |
| 128 | 992 # 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
|
993 # 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
|
994 [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
|
995 if isinstance(self.fields_columns[f],list) else |
| 128 | 996 self.data[key,self.fields_columns[f]] for f in fieldnames]), |
| 997 | |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
998 self.valuesVStack,self.valuesHStack) |
| 80 | 999 |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1000 # else check for a fieldname |
| 80 | 1001 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
|
1002 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
|
1003 # else we are trying to access a property of the dataset |
| 80 | 1004 assert key in self.__dict__ # else it means we are trying to access a non-existing property |
| 1005 return self.__dict__[key] | |
|
55
66619ce44497
Efficient implementation of getitem for ArrayDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
48
diff
changeset
|
1006 |
| 41 | 1007 |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1008 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
|
1009 class ArrayDataSetIterator(object): |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1010 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
|
1011 if fieldnames is None: fieldnames = dataset.fieldNames() |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1012 # 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
|
1013 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
|
1014 self.dataset=dataset |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1015 self.minibatch_size=minibatch_size |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1016 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
|
1017 assert offset+minibatch_size<=len(dataset.data) |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1018 self.current=offset |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1019 def __iter__(self): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1020 return self |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1021 def next(self): |
|
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1022 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
|
1023 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
|
1024 self.current+=self.minibatch_size |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1025 return self.minibatch |
|
42
9b68774fcc6b
Testing basic functionality and removing obvious bugs
bengioy@grenat.iro.umontreal.ca
parents:
41
diff
changeset
|
1026 |
|
44
5a85fda9b19b
Fixed some more iterator bugs
bengioy@grenat.iro.umontreal.ca
parents:
43
diff
changeset
|
1027 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
|
1028 |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1029 |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1030 class CachedDataSet(DataSet): |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1031 """ |
| 167 | 1032 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
|
1033 (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
|
1034 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
|
1035 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
|
1036 |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1037 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
|
1038 (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
|
1039 first access. |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1040 |
| 167 | 1041 @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
|
1042 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
|
1043 |
| 167 | 1044 @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
|
1045 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
|
1046 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
|
1047 """ |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1048 def __init__(self,source_dataset,cache_all_upon_construction=False): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1049 self.source_dataset=source_dataset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1050 self.cache_all_upon_construction=cache_all_upon_construction |
|
152
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1051 self.cached_examples = [] |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1052 if cache_all_upon_construction: |
|
77
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1053 # this potentially brings all the source examples |
|
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1054 # into memory at once, which may be too much |
|
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1055 # the work could possibly be done by minibatches |
|
1e2bb5bad636
toying with different ways to implement learners
bengioy@bengiomac.local
parents:
74
diff
changeset
|
1056 # 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
|
1057 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
|
1058 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
|
1059 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
|
1060 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
|
1061 |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1062 self.fieldNames = source_dataset.fieldNames |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1063 self.hasFields = source_dataset.hasFields |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1064 self.valuesHStack = source_dataset.valuesHStack |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1065 self.valuesVStack = source_dataset.valuesVStack |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1066 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1067 def __len__(self): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1068 return len(self.source_dataset) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1069 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1070 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1071 class CacheIterator(object): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1072 def __init__(self,dataset): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1073 self.dataset=dataset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1074 self.current=offset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1075 def __iter__(self): return self |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1076 def next(self): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1077 upper = self.current+minibatch_size |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1078 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
|
1079 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
|
1080 # cache everything from current length to upper |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1081 for example in self.dataset.source_dataset[cache_len:upper]: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1082 self.dataset.cached_examples.append(example) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1083 all_fields_minibatch = Example(self.dataset.fieldNames(), |
|
152
3f627e844cba
Fixes in CacheDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
151
diff
changeset
|
1084 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
|
1085 self.current+=minibatch_size |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1086 if self.dataset.fieldNames()==fieldnames: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1087 return all_fields_minibatch |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1088 return Example(fieldnames,[all_fields_minibatch[name] for name in fieldnames]) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1089 return CacheIterator(self) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1090 |
| 153 | 1091 def __getitem__(self,i): |
| 1092 if type(i)==int and len(self.cached_examples)>i: | |
| 1093 return self.cached_examples[i] | |
| 1094 else: | |
| 1095 return DataSet.__getitem__(self,i) | |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1096 |
|
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1097 class ApplyFunctionDataSet(DataSet): |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1098 """ |
| 167 | 1099 A L{DataSet} that contains as fields the results of applying a |
| 1100 given function example-wise or minibatch-wise to all the fields of | |
| 1101 an input dataset. The output of the function should be an iterable | |
| 1102 (e.g. a list or a LookupList) over the resulting values. | |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1103 |
| 167 | 1104 In minibatch mode, the function is expected to work on minibatches |
| 1105 (takes a minibatch in input and returns a minibatch in output). More | |
| 1106 precisely, it means that each element of the input or output list | |
| 1107 should be iterable and indexable over the individual example values | |
| 1108 (typically these elements will be numpy arrays). All of the elements | |
| 1109 in the input and output lists should have the same length, which is | |
| 1110 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
|
1111 |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1112 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
|
1113 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
|
1114 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1115 If the values_{h,v}stack functions are not provided, then |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1116 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
|
1117 """ |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1118 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
|
1119 values_hstack=None,values_vstack=None, |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1120 description=None,fieldtypes=None): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1121 """ |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1122 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
|
1123 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
|
1124 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
|
1125 (provided in a list). |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1126 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1127 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
|
1128 (it takes minibatches of inputs and produces minibatches of outputs, as |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1129 documented in the class comment). |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1130 """ |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1131 self.input_dataset=input_dataset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1132 self.function=function |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1133 self.output_names=output_names |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1134 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
|
1135 DataSet.__init__(self,description,fieldtypes) |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1136 self.valuesHStack = values_hstack if values_hstack else input_dataset.valuesHStack |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1137 self.valuesVStack = values_vstack if values_vstack else input_dataset.valuesVStack |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1138 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1139 def __len__(self): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1140 return len(self.input_dataset) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1141 |
|
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1142 def fieldNames(self): |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1143 return self.output_names |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1144 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1145 def minibatches_nowrap(self,fieldnames,minibatch_size,n_batches,offset): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1146 class ApplyFunctionIterator(object): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1147 def __init__(self,output_dataset): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1148 self.input_dataset=output_dataset.input_dataset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1149 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
|
1150 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
|
1151 n_batches=n_batches,offset=offset).__iter__() |
|
73
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 def __iter__(self): return self |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1154 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1155 def next(self): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1156 function_inputs = self.input_iterator.next() |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1157 all_output_names = self.output_dataset.output_names |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1158 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
|
1159 function_outputs = self.output_dataset.function(*function_inputs) |
|
73
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1160 else: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1161 input_examples = zip(*function_inputs) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1162 output_examples = [self.output_dataset.function(input_example) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1163 for input_example in input_examples] |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1164 function_outputs = [self.output_dataset.valuesVStack(name,values) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1165 for name,values in zip(all_output_names, |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1166 zip(*output_examples))] |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1167 all_outputs = Example(all_output_names,function_outputs) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1168 if fieldnames==all_output_names: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1169 return all_outputs |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1170 return Example(fieldnames,[all_outputs[name] for name in fieldnames]) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1171 |
|
134
3f4e5c9bdc5e
Fixes to ApplyFunctionDataSet and other things to make learner and mlp work
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
132
diff
changeset
|
1172 return ApplyFunctionIterator(self) |
|
73
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 __iter__(self): # only implemented for increased efficiency |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1175 class ApplyFunctionSingleExampleIterator(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.current=0 |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1178 self.output_dataset=output_dataset |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1179 self.input_iterator=output_dataset.input_dataset.__iter__() |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1180 def __iter__(self): return self |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1181 def next(self): |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1182 function_inputs = self.input_iterator.next() |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1183 if self.output_dataset.minibatch_mode: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1184 function_outputs = [output[0] for output in self.output_dataset.function(function_inputs)] |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1185 else: |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1186 function_outputs = self.output_dataset.function(function_inputs) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1187 return Example(self.output_dataset.output_names,function_outputs) |
|
69f97aad3faf
Coded untested ApplyFunctionDataSet and CacheDataSet
bengioy@bengiomac.local
parents:
72
diff
changeset
|
1188 return ApplyFunctionSingleExampleIterator(self) |
|
57
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1189 |
|
1aabd2e2bb5f
Added empty classes with doc: CachedDataSet and ApplyFunctionDataSet
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
56
diff
changeset
|
1190 |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1191 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
|
1192 """ |
| 167 | 1193 Wraps an arbitrary L{DataSet} into one for supervised learning tasks |
| 1194 by forcing the user to define a set of fields as the 'input' field | |
| 1195 and a set of fields as the 'target' field. Optionally, a single | |
| 1196 weight_field can also be defined. | |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1197 """ |
|
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1198 args = ((input_fields,'input'),(output_fields,'target')) |
|
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1199 if weight_field: args+=(([weight_field],'weight')) |
|
36
438440ba0627
Rewriting dataset.py completely
bengioy@zircon.iro.umontreal.ca
parents:
29
diff
changeset
|
1200 return src_dataset.merge_fields(*args) |
|
23
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1201 |
|
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1202 |
|
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1203 |
|
526e192b0699
Working on ApplyFunctionDataSet, added constraint that
bengioy@esprit.iro.umontreal.ca
parents:
22
diff
changeset
|
1204 |