def backward(ctx, gradOutput):
input, g0, g1, g2, g3, temp_out, mask = ctx.saved_tensors
# print temp_out.size()
# print mask.size()
assert (gradOutput.is_contiguous() == True)
with torch.cuda.device_of(gradOutput):
num, channels, depth, height, width = input.size()
# _, _, fsize, _, _ = g0.size()
# print fsize
gradInput = gradOutput.new().resize_(num, channels, depth, height,
width).zero_()
grad0 = gradOutput.new().resize_(g0.size()).zero_()
grad1 = gradOutput.new().resize_(g1.size()).zero_()
grad2 = gradOutput.new().resize_(g2.size()).zero_()
grad3 = gradOutput.new().resize_(g3.size()).zero_()
temp_grad = gradOutput.new().resize_(num, channels, depth, height,
width).zero_()
max_idx = gradOutput.new().resize_(num, channels, height,
width).zero_()
GANet.sga_cuda_backward(input, g0, g1, g2, g3, temp_out, mask,
max_idx, gradOutput, temp_grad, gradInput,
grad0, grad1, grad2, grad3)
# GANet.lga_cuda_backward(input, filters, gradOutput, gradInput, gradFilters, radius)
gradInput = gradInput.contiguous()
grad0 = grad0.contiguous()
grad1 = grad1.contiguous()
grad2 = grad2.contiguous()
grad3 = grad3.contiguous()
return gradInput, grad0, grad1, grad2, grad3
def forward(ctx, input, filters):
ctx.radius = radius
assert (input.is_contiguous() == True
and filters.is_contiguous() == True)
with torch.cuda.device_of(input):
num, channels, height, width = input.size()
output = input.new().resize_(num, channels, height, width).zero_()
GANet.lga_cuda_forward(input, filters, output, radius)
output = output.contiguous()
ctx.save_for_backward(input, filters)
return output
def backward(ctx, gradOutput):
input, filters = ctx.saved_tensors
assert (gradOutput.is_contiguous() == True)
with torch.cuda.device_of(gradOutput):
num, channels, height, width = input.size()
_, fsize, _, _ = filters.size()
gradInput = gradOutput.new().resize_(num, channels, height,
width).zero_()
gradFilters = gradOutput.new().resize_(num, fsize, height,
width).zero_()
GANet.lga_cuda_backward(input, filters, gradOutput, gradInput,
gradFilters, ctx.radius)
gradInput = gradInput.contiguous()
gradFilters = gradFilters.contiguous()
return gradInput, gradFilters, None
def forward(ctx, input, g0, g1, g2, g3):
assert (input.is_contiguous() == True and g0.is_contiguous() == True
and g1.is_contiguous() == True and g2.is_contiguous() == True
and g3.is_contiguous() == True)
with torch.cuda.device_of(input):
num, channels, depth, height, width = input.size()
output = input.new().resize_(num, channels, depth, height,
width).zero_()
temp_out = input.new().resize_(num, channels, depth, height,
width).zero_()
mask = input.new().resize_(num, channels, depth, height,
width).zero_()
GANet.sga_cuda_forward(input, g0, g1, g2, g3, temp_out, output,
mask)
# GANet.sga_cuda_forward(input, filters, output, radius)
output = output.contiguous()
ctx.save_for_backward(input, g0, g1, g2, g3, temp_out, mask)
return output