Mercurial > pylearn

comparison datasets/smallNorb.py @ 505:74b3e65f5f24

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added smallNorb dataset, switched to PYLEARN_DATA_ROOT
author James Bergstra <bergstrj@iro.umontreal.ca>
date Wed, 29 Oct 2008 17:09:04 -0400
parents
children eda3d576ee97
comparison
equal deleted inserted replaced
504:19ab9ce916e3 505:74b3e65f5f24
1 import os
2 import numpy
3 from ..filetensor import read
4 from .config import data_root
5
6 #Path = '/u/bergstrj/pub/data/smallnorb'
7 #Path = '/home/fringant2/lisa/louradoj/data/smallnorb'
8 #Path = '/home/louradou/data/norb'
9
10 class Paths(object):
11 """File-related operations on smallNorb
12 """
13 def __init__(self):
14 smallnorb = [data_root(), 'smallnorb']
15 self.train_dat = os.path.join(*\
16 smallnorb + ['smallnorb-5x46789x9x18x6x2x96x96-training-dat.mat'])
17 self.test_dat = os.path.join(*\
18 smallnorb + ['smallnorb-5x01235x9x18x6x2x96x96-testing-dat.mat'])
19 self.train_cat = os.path.join(*\
20 smallnorb + ['smallnorb-5x46789x9x18x6x2x96x96-training-cat.mat'])
21 self.test_cat = os.path.join(*\
22 smallnorb + ['smallnorb-5x01235x9x18x6x2x96x96-testing-cat.mat'])
23 self.train_info = os.path.join(*\
24 smallnorb + ['smallnorb-5x46789x9x18x6x2x96x96-training-info.mat'])
25 self.test_info = os.path.join(*\
26 smallnorb + ['smallnorb-5x01235x9x18x6x2x96x96-testing-info.mat'])
27
28 def load_append_train_test(self, normalize_pixels=True, downsample_amt=1, dtype='float64'):
29
30 def downsample(dataset):
31 return dataset[:, 0, ::downsample_amt, ::downsample_amt]
32 samples = downsample(read(open(self.train_dat)))
33 samples = numpy.vstack((samples, downsample(read(open(self.test_dat)))))
34 samples = numpy.asarray(samples, dtype=dtype)
35 if normalize_pixels:
36 samples *= (1.0 / 255.0)
37
38 labels = read(open(self.train_cat))
39 labels = numpy.hstack((labels, read(open(self.test_cat))))
40
41 infos = read(open(self.train_info))
42 infos = numpy.vstack((infos, read(open(self.test_info))))
43
44 return samples, labels, infos
45
46 def smallnorb_iid(ntrain=29160, nvalid=9720, ntest=9720, dtype='float64', normalize_pixels=True):
47 """Variation of the smallNorb task in which we randomly shuffle all the object instances
48 together before dividing into train/valid/test.
49
50 The default train/valid/test sizes correspond to 60/20/20 split of the entire dataset.
51
52 :returns: 5, (train_x, train_labels), (valid_x, valid_labels), (test_x, test_labels)
53
54 """
55 # cut from /u/louradoj/theano/hpu/expcode1.py
56 rng = numpy.random.RandomState(1)
57 samples, labels, infos = Paths().load_append_train_test(downsample_amt=3, dtype=dtype, normalize_pixels=normalize_pixels)
58
59 nsamples = samples.shape[0]
60 if ntrain + nvalid + ntest > nsamples:
61 raise Exception("ntrain+nvalid+ntest exceeds number of samples (%i)" % nsamples,
62 (ntrain, nvalid, ntest))
63 i0 = 0
64 i1 = ntrain
65 i2 = ntrain + nvalid
66 i3 = ntrain + nvalid + ntest
67
68 indices = rng.permutation(nsamples)
69 train_rows = indices[i0:i1]
70 valid_rows = indices[i1:i2]
71 test_rows = indices[i2:i3]
72
73 n_labels = 5
74
75 def _pick_rows(rows):
76 a = numpy.array([ samples[i].flatten() for i in train_rows])
77 b = numpy.array([ [labels[i]] for i in train_rows])
78 return a, b
79
80 return n_labels, [_pick_rows(r) for r in (train_rows, valid_rows, test_rows)]
81
82 def smallnorb_azSplit():
83 # cut from /u/louradoj/theano/hpu/expcode1.py
84 # WARNING NOT NECESSARILY WORKING CODE
85
86 samples, labels, infos = _load_append_train_test()
87 train_rows, valid_rows, test_rows = [], [], []
88 train_rows_azimuth = []
89 for instance in range(10):
90 az_min = 4*instance
91 az_max = 4*instance + 18
92 train_rows_azimuth.append( [a % 36 for a in range(az_min,az_max,2)] )
93 #print "train_rows_azimuth", train_rows_azimuth
94 for i, info in enumerate(infos):
95 if info[2] in train_rows_azimuth[info[0]]:
96 train_rows.append(i)
97 elif info[2] / 2 % 2 == 0:
98 test_rows.append(i)
99 else:
100 valid_rows.append(i)
101
102 return [_pick_rows(samples, labels, r) for r in (train_rows, valid_rows, test_rows)]