Remove Avatar on gathering gradients

Dragon/python/dragon/operators/vision.py

@@ -4,6 +4,10 @@
 # Written by Ting Pan
 # --------------------------------------------------------
 
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
 import math
 from six.moves import range as xrange
 
@@ -98,7 +102,7 @@ def Conv2d(inputs, num_output, kernel_size,
                 input_size = output.shape[i + spatial_axis]
                 output_size = (input_size + s - 1) / float(s)
                 dp = int(max(0, (output_size - 1) * s + k - input_size))
-            output.shape[i + spatial_axis] = (output.shape[i + spatial_axis] + dp - dk) / s + 1
+            output.shape[i + spatial_axis] = int(output.shape[i + spatial_axis] + dp - dk / s) + 1
 
     return output
 
@@ -289,13 +293,13 @@ def Pool2d(inputs, kernel_size, stride, pad=0, padding='VALID',
             if not global_pooling:
                 if padding != 'SAME':
                     input_size = output.shape[i + spatial_axis]
-                    output_size = int(math.ceil(float(output.shape[i + spatial_axis] + 2 * p - k) / s) + 1)
+                    output_size = int(math.ceil((output.shape[i + spatial_axis] + 2 * p - k) / s) + 1)
                     if ((output_size - 1) * s >= input_size + p):
                         output_size = output_size - 1
                     output.shape[i + spatial_axis] = output_size
                 else:
                     output.shape[i + spatial_axis] = \
-                        int((output.shape[i + spatial_axis] + s - 1) / float(s))
+                        int((output.shape[i + spatial_axis] + s - 1) / s)
             else:
                 output.shape[i + spatial_axis] = 1
 

Dragon/python/dragon/vm/caffe/layers/loss.py

@@ -68,6 +68,7 @@ class SigmoidCrossEntropyLossLayer(Layer):
         super(SigmoidCrossEntropyLossLayer, self).Setup(bottom)
         loss = ops.SigmoidCrossEntropy(bottom, **self._param)
         if self._loss_weight is not None: loss *= self._loss_weight
+        return loss
 
 
 class L2LossLayer(Layer):

Dragon/python/dragon/vm/caffe/layers/vision.py

@@ -255,7 +255,7 @@ class NNResizeLayer(Layer):
         super(NNResizeLayer, self).__init__(LayerParameter)
         param = LayerParameter.resize_param
         dsize = [int(dim) for dim in param.shape.dim] \
-            if param.HasField('shape') else []
+            if param.HasField('shape') else None
         self._param = {'dsize': dsize,
                        'fx': float(param.fx),
                        'fy': float(param.fy),
@@ -287,7 +287,7 @@ class BilinearResizeLayer(Layer):
         super(BilinearResizeLayer, self).__init__(LayerParameter)
         param = LayerParameter.resize_param
         dsize = [int(dim) for dim in param.shape.dim] \
-            if param.HasField('shape') else []
+            if param.HasField('shape') else None
         self._param = {'dsize': dsize,
                        'fx': float(param.fx),
                        'fy': float(param.fy),

Dragon/src/operators/misc/gradient_op.cc

@@ -35,9 +35,11 @@ template <class Context> template <typename T>
 void GradientGatherOp<Context>::RunWithType() {
     auto* dXdata = output(0)->template mutable_data<T, Context>();
     TIndex count = output(0)->count();
-    for (int i = 1; i < indices.size(); i++) {
+    for (int i = 0; i < indices.size(); i++) {
         CHECK(output(0)->dims() == input(indices[i]).dims());
-        math::Add<T, Context>(count, dXdata, input(indices[i]).template data<T, Context>(), dXdata);
+        auto* dYdata = input(indices[i]).template data<T, Context>();
+        if (i == 0) ctx().template Copy<T, Context, Context>(count, dXdata, dYdata);
+        else math::Add<T, Context>(count, dXdata, dYdata, dXdata);
         input(indices[i]).Reset();
     }
 }
@@ -45,7 +47,7 @@ void GradientGatherOp<Context>::RunWithType() {
 template <class Context>
 void GradientGatherOp<Context>::RunOnDevice() {
     if (indices.size() == 0) return;
-    ws()->CreateAvatar(output(0), &input(indices[0]));
+    output(0)->ReshapeLike(input(indices[0]));
 
     if (input(indices[0]).template IsType<float>()) RunWithType<float>();
     else LOG(FATAL) << "Unsupported input types.";