Reshape grad in dot

Author: apaszke

torch/autograd/_functions/blas.py

@@ -179,16 +179,17 @@ class Dot(Function):
 
  def forward(self, vector1, vector2):
  self.save_for_backward(vector1, vector2)
+ self.sizes = (vector1.size(), vector2.size())
  return vector1.new((vector1.dot(vector2),))
 
  def backward(self, grad_output):
  vector1, vector2 = self.saved_tensors
  grad_vector1 = grad_vector2 = None
 
  if self.needs_input_grad[0]:
- grad_vector1 = vector2.mul(grad_output[0])
+ grad_vector1 = vector2.mul(grad_output[0]).view(self.sizes[0])
 
  if self.needs_input_grad[1]:
- grad_vector2 = vector1.mul(grad_output[0])
+ grad_vector2 = vector1.mul(grad_output[0]).view(self.sizes[1])
 
  return grad_vector1, grad_vector2