infer shape of depthwise op to give as input to pointwise op

theano/tensor/nnet/abstract_conv.py

@@ -545,8 +545,8 @@ def conv2d(input,
 
 
 def separable_conv2d(input,
-                     depthwise_filter,
-                     pointwise_filter,
+                     depthwise_filters,
+                     pointwise_filters,
                      num_channels,
                      input_shape=None,
                      depthwise_filter_shape=None,
@@ -556,20 +556,22 @@ def separable_conv2d(input,
                      filter_flip=True,
                      filter_dilation=(1, 1)):
 
-    depthwise_op = conv2d(input=input,
-                          filters=depthwise_filter,
-                          input_shape=input_shape,
-                          filter_shape=depthwise_filter_shape,
+    input = as_tensor_variable(input)
+    depthwise_filters = as_tensor_variable(depthwise_filters)
+    conv_op = AbstractConv2d(imshp=input_shape,
+                             kshp=depthwise_filter_shape,
                              border_mode=border_mode,
                              subsample=subsample,
                              filter_flip=filter_flip,
                              filter_dilation=filter_dilation,
                              num_groups=num_channels)
 
-    #TODO: infer shape of output pf depthwise_op
+    depthwise_op_shape = conv_op.infer_shape(None, [input_shape, depthwise_filter_shape])
+    depthwise_op = conv_op(input, depthwise_filters)
+
     pointwise_op = conv2d(input=depthwise_op,
-                          filters=pointwise_filter,
-                          input_shape=None,
+                          filters=pointwise_filters,
+                          input_shape=depthwise_op_shape[0],
                           filter_shape=pointwise_filter_shape,
                           border_mode='valid',
                           subsample=(1, 1),