Hi everyone,
Do you know if it’s possible to train a model with the condition to not do the back propagation when it’s not differentiable ? For exemple, I want to throw away the input if at some point I do ReLU’(0).
class MyReLU(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
if input == 0:
ctx.mark_non_differentiable(input)
return input.clamp(min=0)
else:
ctx.save_for_backward(input)
return input.clamp(min=0)
@staticmethod
def backward(ctx, grad_output):
input, = ctx.saved_tensors
grad_input = grad_output.clone()
grad_input[input < 0] = 0
return grad_input
Do you think it’s possible to do it ?
Thanks