Merge pull request #4085 from kmaehashi/avoid-np-stack

avoid using np.stack which is introduced in NumPy 1.10
chainer · Jan 10, 2018 · 5b34928 · 5b34928
2 parents e8504de + 3b0e47a
commit 5b34928
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 12 deletions.
diff --git a/chainer/functions/connection/convolution_2d.py b/chainer/functions/connection/convolution_2d.py
@@ -217,9 +217,7 @@ def _forward_grouped_convolution(self, x, W, b):
                 _y = self._forward_gpu_core(_x[g, ], _W[g, ], _bg)
             _ys.append(_y)
 
-        y = xp.stack(_ys, axis=1)  # (N, G, oCg, oH, oW)
-        _, _, _, oH, oW = y.shape
-        y = y.reshape(N, oC, oH, oW)
+        y = xp.concatenate(_ys, axis=1)  # (N, oC, oH, oW)
         return y
 
     def _forward_cudnn(self, x, W, b, y):
@@ -386,8 +384,6 @@ def _forward_grouped_convolution(self, x, gy):
         # iC, iH, iW: input channels, input height, input width
         # oC, oH, oW: output channels, output height, output width
         G = self.group
-        kH = self.kh
-        kW = self.kw
         N, iC, iH, iW = x.shape
         _, oC, oH, oW = gy.shape
         iCg = int(iC / G)
@@ -411,8 +407,7 @@ def _forward_grouped_convolution(self, x, gy):
                 _gW = self._forward_gpu_core(_x[g, ], _gy[g, ])
             _gWs.append(_gW)
 
-        gW = xp.stack(_gWs)  # (G, oCg, iCg, kH, kW)
-        gW = gW.reshape(oC, iCg, kH, kW)
+        gW = xp.concatenate(_gWs)  # (oC, iCg, kH, kW)
         return gW
 
     def _forward_cudnn(self, x, gy):

diff --git a/chainer/functions/connection/deconvolution_2d.py b/chainer/functions/connection/deconvolution_2d.py
@@ -212,9 +212,6 @@ def _forward_grouped_convolution(self, x, W, b):
         N, xC, xH, xW = x.shape
         xCg = int(xC / G)
         _, yCg, kH, kW = W.shape
-        yC = yCg * G
-        yH = self.outh
-        yW = self.outw
 
         xp = cuda.get_array_module(x)
 
@@ -233,8 +230,7 @@ def _forward_grouped_convolution(self, x, W, b):
                 _y = self._forward_gpu_core(_x[g, ], _W[g, ], _bg)
             _ys.append(_y)
 
-        y = xp.stack(_ys, axis=1)  # (N, G, yCg, yH, yW)
-        y = y.reshape(N, yC, yH, yW)
+        y = xp.concatenate(_ys, axis=1)  # (N, yC, yH, yW)
         return y
 
     def _forward_cudnn(self, x, W, b):