Mercurial > pylearn

comparison pylearn/dataset_ops/image_patches.py @ 998:8ba8b08e0442

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added the image_patches dataset used in RanzatoHinton2010 modified mcRBM to use it.
author James Bergstra <bergstrj@iro.umontreal.ca>
date Tue, 24 Aug 2010 16:51:53 -0400
parents 507159eea97e
children 976539956475
comparison
equal deleted inserted replaced
997:71b0132b694a 998:8ba8b08e0442
1 import os, numpy 1 import os, numpy
2 import theano 2 import theano
3 3
4 from pylearn.datasets.image_patches import ( 4 from pylearn.datasets.image_patches import (
5 data_root,
5 olshausen_field_1996_whitened_images, 6 olshausen_field_1996_whitened_images,
6 extract_random_patches) 7 extract_random_patches)
7 8
8 from .protocol import TensorFnDataset # protocol.py __init__.py 9 from .protocol import TensorFnDataset # protocol.py __init__.py
9 from .memo import memo 10 from .memo import memo
11
12 import scipy.io
13 from pylearn.io import image_tiling
10 14
11 @memo 15 @memo
12 def get_dataset(N,R,C,dtype,center,unitvar): 16 def get_dataset(N,R,C,dtype,center,unitvar):
13 seed=98234 17 seed=98234
14 rng = numpy.random.RandomState(seed) 18 rng = numpy.random.RandomState(seed)
46 else: 50 else:
47 assert False, 'what happened?' 51 assert False, 'what happened?'
48 52
49 return x 53 return x
50 54
55
56
57 @memo
58 def ranzato_hinton_2010(path=None):
59 if path is None:
60 path = os.path.join(data_root(), 'image_patches', 'mcRBM',
61 'training_colorpatches_16x16_demo.mat')
62 dct = scipy.io.loadmat(path)
63 return dct
64 def ranzato_hinton_2010_whitened_patches(path=None):
65 """Return the pca of the data, which is 10240 x 105
66 """
67 dct = ranzato_hinton_2010(path)
68 return dct['whitendata'].astype('float32')
69
70 def undo_pca_filters_of_ranzato_hinton_2010(X, path=None):
71 """Return tuple (R,G,B,None) of matrices for matrix `X` of filters (one per row)
72
73 Return value can be passed to `image_tiling.tile_raster_images`.
74 """
75 dct = ranzato_hinton_2010(path)
76 X = numpy.dot(X, dct['invpcatransf'].T)
77 return (X[:,:256], X[:,256:512], X[:,512:], None)
78
79 def save_filters_of_ranzato_hinton_2010(X, fname, min_dynamic_range=1e-3, data_path=None):
80 _img = image_tiling.tile_raster_images(
81 undo_pca_filters_of_ranzato_hinton_2010(X, path=data_path),
82 img_shape=(16,16),
83 min_dynamic_range=min_dynamic_range)
84 image_tiling.save_tiled_raster_images(_img, fname)
85
86 def ranzato_hinton_2010_op(s_idx,
87 split='train',
88 dtype=theano.config.floatX, rasterized=True,
89 center=True,
90 unitvar=True):
91 N = 10240
92
93 if split != 'train':
94 raise NotImplementedError('train/test/valid splits for randomly sampled image patches?')
95
96 if not rasterized:
97 # the data is provided as PCA-sphered, so rasterizing does not make sense
98 # TODO: add a param to enable/disable 'PCA', and if disabled, then consider
99 # rasterizing or not
100 raise NotImplementedError('only pca data is provided')
101
102 if dtype != 'float32':
103 raise NotImplementedError('dtype not float32')
104
105 op = TensorFnDataset(dtype,
106 bcast=(False,),
107 fn=ranzato_hinton_2010_whitened_patches,
108 single_shape=(105,))
109 x = op(s_idx%N)
110 return x
111