def backward(ctx, gradoutput):
input1, input2 = ctx.saved_tensors
if input1.is_cuda:
gradinput1 = torch.cuda.FloatTensor().resize_(input1.size()).zero_()
gradinput2 = torch.cuda.FloatTensor().resize_(input2.size()).zero_()
err = my_lib.InterpolationLayer_gpu_backward(input1,input2,gradoutput,gradinput1,gradinput2)
if err != 0 :
print(err)
else:
gradinput1 = torch.FloatTensor().resize_(input1.size()).zero_()
gradinput2 = torch.FloatTensor().resize_(input2.size()).zero_()
err = my_lib.InterpolationLayer_cpu_backward(input1, input2, gradoutput, gradinput1, gradinput2)
if err != 0 :
print(err)
return gradinput1, gradinput2
def backward(ctx, gradoutput):
# print("Backward of Interpolation Layer")
# gradinput1 = input1.new().zero_()
# gradinput2 = input2.new().zero_()
# gradinput1 = torch.zeros(self.input1.size())
# gradinput2 = torch.zeros(self.input2.size())
input1, input2 = ctx.saved_tensors
if input1.is_cuda:
# print("CUDA backward")
# gradinput1 = gradinput1.cuda(self.device)
# gradinput2 = gradinput2.cuda(self.device)
gradinput1 = torch.cuda.FloatTensor().resize_(
input1.size()).zero_()
gradinput2 = torch.cuda.FloatTensor().resize_(
input2.size()).zero_()
# the input1 image should not require any gradients
# print("Does input1 requires gradients? " + str(self.input1.requires_grad))
err = my_lib.InterpolationLayer_gpu_backward(
input1, input2, gradoutput, gradinput1, gradinput2)
if err != 0:
print(err)
else:
# print("CPU backward")
# print(gradoutput)
gradinput1 = torch.FloatTensor().resize_(input1.size()).zero_()
gradinput2 = torch.FloatTensor().resize_(input2.size()).zero_()
err = my_lib.InterpolationLayer_cpu_backward(
input1, input2, gradoutput, gradinput1, gradinput2)
# print(err)
if err != 0:
print(err)
# print(gradinput1)
# print(gradinput2)
# print(gradinput1)
return gradinput1, gradinput2