def backward(grad_output, input_, input_low, input_range, output,
level_low, level_high, range_sign):
mask_hi = (input_ > (input_low + input_range)).astype(input_.dtype)
mask_lo = (input_ < input_low).astype(input_.dtype)
mask_in = 1 - mask_hi - mask_lo
err = (output - input_) * np.reciprocal(input_range * range_sign)
grad_range = grad_output * (err * mask_in + range_sign *
(level_low / level_high) * mask_lo +
mask_hi)
grad_range = sum_like(grad_range, input_range)
grad_input = grad_output * mask_in
grad_low = grad_output * (mask_hi + mask_lo)
grad_low = sum_like(grad_low, input_low)
return [grad_input, grad_low, grad_range]
def backward(ctx, grad_output):
input_, scale, output = ctx.saved_tensors
level_high = ctx.level_high
level_low = ctx.level_low
alpha = float(level_low) / float(level_high)
mask_hi = (input_ > scale).type(input_.dtype)
mask_lo = (input_ < scale * alpha).type(input_.dtype)
mask_in = 1 - mask_hi - mask_lo
val_grad_out = mask_hi + alpha * mask_lo
err = (output - input_) * scale.reciprocal()
grad_scale = grad_output * (err * mask_in + val_grad_out)
grad_scale = sum_like(grad_scale, scale)
# calc gradient for input
grad_input = grad_output * mask_in
return grad_input, grad_scale, None