Merge pull request #9 from devitocodes/zero_in_op

Move zeroing layers to inside an operator

joey/layers.py

@@ -483,8 +483,7 @@ def backprop_equations(self, prev_layer, next_layer):
         a = self._backward_tmp_constants[0]
         b = self._backward_tmp_constants[1]
 
-        return ([Eq(next_layer.result_gradients, 0),
-                 Eq(a, index // 2),
+        return ([Eq(a, index // 2),
                  Eq(b, index % 2),
                  Inc(next_layer.result_gradients[dims[0],
                                                  dims[1],

joey/net.py

@@ -68,6 +68,9 @@ def _gen_eqs(self):
 
         input_function = None
 
+        for layer in self._layers:
+            eqs.append(Eq(layer.result, 0))
+
         for layer in self._layers:
             if input_function is not None:
                 dims = input_function.dimensions
@@ -85,6 +88,19 @@ def _gen_backprop_eqs(self):
         eqs = []
         args = []
 
+        for i in range(len(self._layers)):
+            layer = self._layers[i]
+
+            if layer.kernel_gradients is not None:
+                eqs.append(Eq(layer.kernel_gradients, 0))
+
+            if layer.bias_gradients is not None:
+                eqs.append(Eq(layer.bias_gradients, 0))
+
+            if layer.result_gradients is not None \
+               and i < len(self._layers) - 1:
+                eqs.append(Eq(layer.result_gradients, 0))
+
         for i in range(len(self._layers) - 1, -1, -1):
             if i < len(self._layers) - 1:
                 prev_layer = self._layers[i + 1]
@@ -102,6 +118,17 @@ def _gen_backprop_eqs(self):
             args += layer_args
             eqs += layer_eqs
 
+        batch_size = self._layers[-1].result.shape[1]
+
+        for layer in self._layers:
+            if layer.kernel_gradients is not None:
+                eqs.append(Eq(layer.kernel_gradients,
+                              layer.kernel_gradients / batch_size))
+
+            if layer.bias_gradients is not None:
+                eqs.append(Eq(layer.bias_gradients,
+                              layer.bias_gradients / batch_size))
+
         return (eqs, args)
 
     @property
@@ -118,9 +145,6 @@ def forward(self, input_data):
         input_data : np.ndarray
             Input data for the network.
         """
-        for layer in self._layers:
-            layer.result.data[:] = 0
-
         self._layers[0].input.data[:] = input_data
         self._forward_operator.apply(**self._forward_arg_dict)
         return self._layers[-1].result.data
@@ -154,29 +178,10 @@ def backward(self, expected, loss_gradient_func, pytorch_optimizer=None):
 
             The default value is None.
         """
-        for layer in self._layers:
-            if layer.kernel_gradients is not None:
-                layer.kernel_gradients.data[:] = 0
-
-            if layer.bias_gradients is not None:
-                layer.bias_gradients.data[:] = 0
-
-            if layer.result_gradients is not None:
-                layer.result_gradients.data[:] = 0
-
-        batch_size = self._layers[-1].result.shape[1]
-
         self._layers[-1].result_gradients.data[:] = \
             np.transpose(np.array(loss_gradient_func(self._layers[-1],
                                                      expected)))
         self._backward_operator.apply(**self._backward_arg_dict)
 
-        for layer in self._layers:
-            if layer.kernel_gradients is not None:
-                layer.kernel_gradients.data[:] /= batch_size
-
-            if layer.bias_gradients is not None:
-                layer.bias_gradients.data[:] /= batch_size
-
         if pytorch_optimizer is not None:
             pytorch_optimizer.step()