Custom parameter not updating weights

1

I am trying to create a custom parameter as a scale factor to resize the tensor via bilinear interpolation

self.scl = nn.Parameter(torch.tensor(0.5, requires_grad=True))

def forward(self, inx):
    x = self.lrelu(self.part1_PW(inx))
    x0 = x
    #==========
    b, c, h, w = x.shape

    scl = self.scl.item()
    scl = round(scl, 1)
    #print(scl)
    if scl == 0:
        scl = 1.0
    elif scl <0:
        if scl > -1:
            scl = scl*(-1.0)
        else:
            scl = 1.0
    else:
        if scl > 1:
            scl = 1.0

    downscale = nn.Upsample(size=(int(h * scl), int(w * scl)), mode='bilinear', align_corners=False)

But when I checked, I realized the weights aren’t updating
is the problem from the fact that I didn’t use the weights on the input tensors?

2

Based on your code it seems you are using self.scl.item() which will return a plain Python float and will thus not update the parameter.
To get valid gradients in the parameter you would have to use it directly.
However, based on your code you are also assigning new values to it, so I’m unsure how the actual usage and update should work.