Hello,
I'd like to raise an issue regarding the function backward of class SFB1D(Function).
In the code, it checks whether there is a need to do backward depending on the needs_input_grad[0].
In my case, needs_input_grad = (False, True, False, False, False). So I still want to perform backward with resepct to the second input of the forward function. This is not done with the current implentation . Does it make sense to
replace the condition if ctx.needs_input_grad[0]: by
if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]: ?
def backward(ctx, dy):
print('TO compute backward SFB1D',ctx.needs_input_grad)
dlow, dhigh = None, None
if ctx.needs_input_grad[0]:
mode = ctx.mode
g0, g1, = ctx.saved_tensors
dy = dy[:, :, None, :]
dx = afb1d(dy, g0, g1, mode=mode, dim=3)
dlow = dx[:, ::2, 0].contiguous()
dhigh = dx[:, 1::2, 0].contiguous()
return dlow, dhigh, None, None, None, None, None
best,
Sixin
Hello,
I'd like to raise an issue regarding the function backward of class SFB1D(Function).
In the code, it checks whether there is a need to do backward depending on the needs_input_grad[0].
In my case, needs_input_grad = (False, True, False, False, False). So I still want to perform backward with resepct to the second input of the forward function. This is not done with the current implentation . Does it make sense to
replace the condition if ctx.needs_input_grad[0]: by
if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]: ?
def backward(ctx, dy):
print('TO compute backward SFB1D',ctx.needs_input_grad)
dlow, dhigh = None, None
if ctx.needs_input_grad[0]:
mode = ctx.mode
g0, g1, = ctx.saved_tensors
dy = dy[:, :, None, :]
best,
Sixin