PyTorch1.0 TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect

vision
1

I’m trying to convert a model into jit trace and know where the source of my problem is. If you look at the code below

class GLU(nn.Module):
    def __init__(self):
        super(GLU, self).__init__()
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        nc = x.size(1)
        nc = int(nc/2)
        return x[:, :nc] * self.sigmoid(x[:, nc:])

The compiler tells me I’m not allowed to convert a torch value to python integer

nc = int(nc/2)

TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  nc = int(nc/2)

Yet indices in python cannot receive torch values but with integers. Is there a way to modify the forward function to avoid the tracerWarning? One solution I can think of is to modify the indices through a high level torch functions but don’t know the best place to start looking.

2

Try

nc = x.shape[1]

Works?

3

I have faced same problem. Your solution doesn’t work.