I’m using the overruling dataset is a small (n=2400, evenly split) binary classification task. I’m also using a tiny BERT model (4.4M params).

I’m getting strides like (0, 0) when writing custom Pytorch autograd functions. Here’s a minimal, reproducible example:

Im trying to classificate multilabels for sentiment analysis.

I’m trying to calculate negative log-likelihood for evaluation, ‘pred’ is already a probability vector, and by some reasons I’m trying to get log(pred) by
log_probs = F.log_softmax(pred)
but the numerical result is very wired

test = nn.Linear(1440, 1440, bias=False) hidden1 = torch.randn(100, 1440) hidden2 = torch.randn(400, 1440) output1 = test(hidden1) output2 = test(hidden2) If I test it as above, shouldn’t the output1 and output2[:100,:] parts be exactly the same? There are slightly different parts, do you know why? It should be the same as simple matmul calculation, but it […]

I’m using torch.nn.functional.interpolate recently. But I don’t know how it works.
for example, I give an input y as follows:

I have recently installed PyTorch with CUDA support on my machine, but when I run torch.cuda.is_available(), it returns False. I verified my GPU setup using nvidia-smi, and it seems that my system recognizes the GPU correctly.

This problem started a while ago when I uninstalled my CUDA version 12.4 and installed 11.2 since pytorch had some issues with 12.4 . But even after having the new version installed when I check , it shows my CUDA to be unavailable even after I checked the availability in the cmd .

This problem started a while ago when I uninstalled my CUDA version 12.4 and installed 11.2 since pytorch had some issues with 12.4 . But even after having the new version installed when I check , it shows my CUDA to be unavailable even after I checked the availability in the cmd .

How is this possible?