def backward(ctx, grad_output): grad_input = torch.zeros_like(ctx.saved_tensors[0]) saved = [grad_output] + list( ctx.saved_tensors) + [ctx.kernel, ctx.stride] + [grad_input] softpool_cuda.backward_1d(*saved) # Gradient underflow saved[-1][torch.isnan(saved[-1])] = 0 return saved[-1], None, None
def backward(ctx, grad_output): # Create contiguous tensor (if tensor is not contiguous) if (not grad_output.is_contiguous()): x = grad_output.contiguous() grad_input = torch.zeros_like(ctx.saved_tensors[0]) saved = [grad_output] + list(ctx.saved_tensors) + [ctx.kernel, ctx.stride] + [grad_input] softpool_cuda.backward_1d(*saved) # Gradient underflow saved[-1][torch.isnan(saved[-1])] = 0 return saved[-1], None, None