added proper comparison to conv2d.grad

theano/tensor/nnet/tests/test_abstract_conv.py

@@ -1606,6 +1606,7 @@ class TestConv2dGrads(unittest.TestCase):
         self.filter_flip = [True, False]
 
         self.output_grad = T.tensor4()
+        self.output_grad_wrt = T.tensor4()
         self.filters = T.tensor4()
 
         self.x = T.tensor4('x', theano.config.floatX) #inputs
@@ -1628,20 +1629,32 @@ class TestConv2dGrads(unittest.TestCase):
         results are the same.
 
         """
-        for (in_shape, fltr_shape) in zip(inputs_shapes, filters_shapes):
-            for bm in border_modes:
-                for ss in subsamples:
-                    for ff in filter_flip:
-                        conv_out = T.nnet.conv.conv2d(x, 
-                                                    filters = filters,
+        for (in_shape, fltr_shape) in zip(self.inputs_shapes, self.filters_shapes):
+            for bm in self.border_modes:
+                for ss in self.subsamples:
+                    for ff in self.filter_flip:
+                        if filter_flip = True:
+                            fltr_shape = transpose(fltr_shape) #conv2d doesn't seem to have filter_flip
+                        conv_out = T.nnet.conv.conv2d(self.x, 
+                                                    filters = self.filters,
                                                     border_mode = bm,
                                                     subsample = ss,
                                                     image_shape = in_shape
                                                     filter_shape = fltr_shape
                                                     )
-                        conv_grad = theano.grad(wrt=[x], known_grads={conv_out: output_grad})
+                        conv_grad = theano.grad(wrt=[x], known_grads={conv_out: self.output_grad})
                         f_prime = theano.function([x, output_grad, filters], conv_grad)
-                        utt.assert_allclose(conv_grad, f_prime)
+
+                        conv_wrt_i_out = T.nnet.conv.abstract_conv.conv2d_grad_wrt_inputs(self.output_grad_wrt,
+                                                                    filters = self.filters,
+                                                                    border_mode = bm,
+                                                                    subsample = ss,
+                                                                    input_shape = in_shape,
+                                                                    filter_shape = fltr_shape,
+                                                                    filter_flip = ff
+                                                                    )
+                        f = theano.function([x, output_grad_wrt, filters], conv_wrt_i_out)
+                        utt.assert_allclose(f, f_prime)
 
     def test_conv2d_grad_wrt_weights():
         """Compares calculated abstract grads wrt weights with the fwd grads
@@ -1652,47 +1665,34 @@ class TestConv2dGrads(unittest.TestCase):
 
         """
 
-        for (in_shape, fltr_shape) in zip(inputs_shapes, filters_shapes):
-            for bm in border_modes:
-                for ss in subsamples:
-                    for ff in filter_flip:
-                        conv_out = T.nnet.conv.conv2d(w, 
-                                                    filters = filters,
+        for (in_shape, fltr_shape) in zip(self.inputs_shapes, self.filters_shapes):
+            for bm in self.border_modes:
+                for ss in self.subsamples:
+                    for ff in self.filter_flip:
+                        if filter_flip = True:
+                            fltr_shape = transpose(fltr_shape) #conv2d doesn't seem to have filter_flip
+                        conv_out = T.nnet.conv.conv2d(self.w, 
+                                                    filters = self.filters,
                                                     border_mode = bm,
                                                     subsample = ss,
-                                                    image_shape = in_shape
+                                                    image_shape = in_shape,
                                                     filter_shape = fltr_shape
                                                     )
                         conv_grad = theano.grad(wrt=[w], known_grads={conv_out: output_grad})
                         f_prime = theano.function([w, output_grad, filters], conv_grad)
-                        utt.assert_allclose(conv_grad, f_prime)
 
-    def test_conv2_grads_wrt_input_and_weights():
-        """Compares calculated abstract grads wrt [inputs, weights] with the fwd grads
+                        conv_wrt_w_out = T.nnet.conv.abstract_conv.conv2d_grad_wrt_weights(self.output_grad_wrt,
+                                                                    filters = self.filters,
+                                                                    border_mode = bm,
+                                                                    subsample = ss,
+                                                                    input_shape = in_shape,
+                                                                    filter_shape = fltr_shape,
+                                                                    filter_flip = ff
+                                                                    )
+                        f = theano.function([w, output_grad_wrt, filters], conv_wrt_w_out)
+                        utt.assert_allclose(f, f_prime)
 
-        This method checks the outputs of conv2_grad wrt inputs and weights 
-        against the outputs of T.nnet.conv forward grads to make sure the
-        results are the same.
 
-        """
-        #for (input_shape, filter_shape) in zip(inputs_shapes, filter_shapes):
-            #make_rand_inputs = theano.function([x], random_stream(input_shape))
-            #make_rand_filters = theano.function([filters], random_stream(filter_shape))
-
-        for (in_shape, fltr_shape) in zip(inputs_shapes, filters_shapes):
-            for wrt in [x,w]: #for both inputs (x) and weights (w)
-                for bm in border_modes:
-                    for ss in subsamples:
-                        for ff in filter_flip:
-                            conv_out = T.nnet.conv.conv2d(wrt, 
-                                                        filters = filters,
-                                                        border_mode = bm,
-                                                        subsample = ss,
-                                                        image_shape = in_shape
-                                                        filter_shape = fltr_shape
-                                                        )
-                            conv_grad = theano.grad(wrt=[wrt], known_grads={conv_out: output_grad})
-                            f_prime = theano.function([wrt, output_grad, filters], conv_grad)
-                            utt.assert_allclose(conv_grad, f_prime)
-            
-            
+    def test_conv2_grads_wrt_input_and_weights():
+        test_conv2d_grad_wrt_inputs()
+        test_conv2d_grad_wrt_weights()
\ No newline at end of file