Chunk / ImageR

Download as
Browse Code ยป

Commit 4e8ab1ef343355a45dfe59c92d605d3e2c849004

Authored by Chunk
1 parent 4eac6680
Exists in master and in 1 other branch refactor

staged.

Showing 1 changed file with 8 additions and 8 deletions   Show diff stats
mmodel/theano/THEANO.py
  Diff comments   View file @4e8ab1e
@@ -48,7 +48,7 @@ class ModelTHEANO(ModelBase): @@ -48,7 +48,7 @@ class ModelTHEANO(ModelBase):
48 X_train, X_test, Y_train, Y_test = cross_validation.train_test_split(X, Y, test_size=0.2, random_state=0) 48 X_train, X_test, Y_train, Y_test = cross_validation.train_test_split(X, Y, test_size=0.2, random_state=0)
49 49
50 train_set_x, train_set_y = shared_dataset((X_train, Y_train)) 50 train_set_x, train_set_y = shared_dataset((X_train, Y_train))
51 - valid_set_x, valid_set_y = train_set_x[:1000], train_set_y[:1000] 51 + valid_set_x, valid_set_y = shared_dataset((X_train[:1000], Y_train[:1000]))
52 test_set_x, test_set_y = shared_dataset((X_test, Y_test)) 52 test_set_x, test_set_y = shared_dataset((X_test, Y_test))
53 53
54 # compute number of minibatches for training, validation and testing 54 # compute number of minibatches for training, validation and testing
@@ -89,8 +89,8 @@ class ModelTHEANO(ModelBase): @@ -89,8 +89,8 @@ class ModelTHEANO(ModelBase):
89 89
90 # Construct the second convolutional pooling layer 90 # Construct the second convolutional pooling layer
91 # filtering reduces the image size to (148-5+1, 148-5+1) = (144, 144) 91 # filtering reduces the image size to (148-5+1, 148-5+1) = (144, 144)
92 - # maxpooling reduces this further to (144/4, 144/4) = (38, 38)  
93 - # 4D output tensor is thus of shape (batch_size, nkerns[1], 38, 38) 92 + # maxpooling reduces this further to (144/4, 144/4) = (36, 36)
  93 + # 4D output tensor is thus of shape (batch_size, nkerns[1], 36, 36)
94 layer1 = ConvPoolLayer( 94 layer1 = ConvPoolLayer(
95 rng, 95 rng,
96 input=layer0.output, 96 input=layer0.output,
@@ -101,15 +101,15 @@ class ModelTHEANO(ModelBase): @@ -101,15 +101,15 @@ class ModelTHEANO(ModelBase):
101 101
102 # the HiddenLayer being fully-connected, it operates on 2D matrices of 102 # the HiddenLayer being fully-connected, it operates on 2D matrices of
103 # shape (batch_size, num_pixels) (i.e matrix of rasterized images). 103 # shape (batch_size, num_pixels) (i.e matrix of rasterized images).
104 - # This will generate a matrix of shape (batch_size, nkerns[1] * 4 * 4),  
105 - # or (500, 50 * 4 * 4) = (500, 800) with the default values. 104 + # This will generate a matrix of shape (batch_size, nkerns[1] * 36 * 36),
  105 + # or (500, 50 * 36 * 36) = (500, 800) with the default values.
106 layer2_input = layer1.output.flatten(2) 106 layer2_input = layer1.output.flatten(2)
107 107
108 # construct a fully-connected sigmoidal layer 108 # construct a fully-connected sigmoidal layer
109 layer2 = HiddenLayer( 109 layer2 = HiddenLayer(
110 rng, 110 rng,
111 input=layer2_input, 111 input=layer2_input,
112 - n_in=nkerns[1] * 38 * 38, 112 + n_in=nkerns[1] * 36 * 36,
113 n_out=500, 113 n_out=500,
114 activation=T.tanh 114 activation=T.tanh
115 ) 115 )
@@ -156,8 +156,8 @@ class ModelTHEANO(ModelBase): @@ -156,8 +156,8 @@ class ModelTHEANO(ModelBase):
156 ] 156 ]
157 """ 157 """
158 Total Parameters: 158 Total Parameters:
159 - >>> 20 * 64 + 1000 * 25 + 50 * 38 * 38 * 500 + 500 * 2  
160 - 36127280 159 + >>> 20 * 64 + 1000 * 25 + 50 * 36 * 36 * 500 + 500 * 2
  160 + 32427280
161 """ 161 """
162 train_model = theano.function( 162 train_model = theano.function(
163 [index], 163 [index],