Add support for dim_ordering in initializations

Author: fchollet

keras/initializations.py

@@ -3,9 +3,26 @@
 from . import backend as K
 
 
-def get_fans(shape):
-    fan_in = shape[0] if len(shape) == 2 else np.prod(shape[1:])
-    fan_out = shape[1] if len(shape) == 2 else shape[0]
+def get_fans(shape, dim_ordering='th'):
+    if len(shape) == 2:
+        fan_in = shape[0]
+        fan_out = shape[1]
+    elif len(shape) == 4 or len(shape) == 5:
+        # assuming convolution kernels (2D or 3D).
+        # TH kernel shape: (depth, input_depth, ...)
+        # TF kernel shape: (..., input_depth, depth)
+        if dim_ordering == 'th':
+            fan_in = np.prod(shape[1:])
+            fan_out = shape[0]
+        elif dim_ordering == 'tf':
+            fan_in = np.prod(shape[:-1])
+            fan_out = shape[-1]
+        else:
+            raise Exception('Invalid dim_ordering: ' + dim_ordering)
+    else:
+        # no specific assumptions
+        fan_in = np.sqrt(np.prod(shape))
+        fan_out = np.sqrt(np.prod(shape))
     return fan_in, fan_out
 
 
@@ -19,39 +36,39 @@ def normal(shape, scale=0.05, name=None):
                       name=name)
 
 
-def lecun_uniform(shape, name=None):
+def lecun_uniform(shape, name=None, dim_ordering='th'):
     ''' Reference: LeCun 98, Efficient Backprop
         http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf
     '''
-    fan_in, fan_out = get_fans(shape)
+    fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering)
     scale = np.sqrt(3. / fan_in)
     return uniform(shape, scale, name=name)
 
 
-def glorot_normal(shape, name=None):
+def glorot_normal(shape, name=None, dim_ordering='th'):
     ''' Reference: Glorot & Bengio, AISTATS 2010
     '''
-    fan_in, fan_out = get_fans(shape)
+    fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering)
     s = np.sqrt(2. / (fan_in + fan_out))
     return normal(shape, s, name=name)
 
 
-def glorot_uniform(shape, name=None):
-    fan_in, fan_out = get_fans(shape)
+def glorot_uniform(shape, name=None, dim_ordering='th'):
+    fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering)
     s = np.sqrt(6. / (fan_in + fan_out))
     return uniform(shape, s, name=name)
 
 
-def he_normal(shape, name=None):
+def he_normal(shape, name=None, dim_ordering='th'):
     ''' Reference:  He et al., http://arxiv.org/abs/1502.01852
     '''
-    fan_in, fan_out = get_fans(shape)
+    fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering)
     s = np.sqrt(2. / fan_in)
     return normal(shape, s, name=name)
 
 
-def he_uniform(shape, name=None):
-    fan_in, fan_out = get_fans(shape)
+def he_uniform(shape, name=None, dim_ordering='th'):
+    fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering)
     s = np.sqrt(6. / fan_in)
     return uniform(shape, s, name=name)
 
@@ -85,5 +102,6 @@ def one(shape, name=None):
 
 
 from .utils.generic_utils import get_from_module
-def get(identifier):
-    return get_from_module(identifier, globals(), 'initialization')
+def get(identifier, **kwargs):
+    return get_from_module(identifier, globals(),
+                           'initialization', kwargs=kwargs)

keras/layers/convolutional.py

@@ -79,7 +79,7 @@ def __init__(self, nb_filter, filter_length,
             raise Exception('Invalid border mode for Convolution1D:', border_mode)
         self.nb_filter = nb_filter
         self.filter_length = filter_length
-        self.init = initializations.get(init)
+        self.init = initializations.get(init, dim_ordering='th')
         self.activation = activations.get(activation)
         assert border_mode in {'valid', 'same'}, 'border_mode must be in {valid, same}'
         self.border_mode = border_mode
@@ -233,7 +233,7 @@ def __init__(self, nb_filter, nb_row, nb_col,
         self.nb_filter = nb_filter
         self.nb_row = nb_row
         self.nb_col = nb_col
-        self.init = initializations.get(init)
+        self.init = initializations.get(init, dim_ordering=dim_ordering)
         self.activation = activations.get(activation)
         assert border_mode in {'valid', 'same'}, 'border_mode must be in {valid, same}'
         self.border_mode = border_mode
@@ -412,7 +412,7 @@ def __init__(self, nb_filter, kernel_dim1, kernel_dim2, kernel_dim3,
         self.kernel_dim1 = kernel_dim1
         self.kernel_dim2 = kernel_dim2
         self.kernel_dim3 = kernel_dim3
-        self.init = initializations.get(init)
+        self.init = initializations.get(init, dim_ordering=dim_ordering)
         self.activation = activations.get(activation)
         assert border_mode in {'valid', 'same'}, 'border_mode must be in {valid, same}'
         self.border_mode = border_mode

tests/integration_tests/test_image_data_tasks.py

@@ -35,7 +35,7 @@ def test_image_classification():
         Activation('relu'),
         Dense(y_test.shape[-1], activation='softmax')
     ])
-    model.compile(loss='categorical_crossentropy', optimizer='sgd')
+    model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
     history = model.fit(X_train, y_train, nb_epoch=10, batch_size=16,
                         validation_data=(X_test, y_test),
                         show_accuracy=True, verbose=0)