deeplearningdtu / 02456-deep-learning-with-pytorch Goto Github PK

Andrea:

The data generation script has some weird stuff going on, and the main notebook works around that by using minibatch-based evaluation. If the validation set is 99, this breaks unless the training batch size is at least 99. The loss is computed using 3 minibatches with size 32 and one with size 3. But actually the last one has still size 32 and the last 29 examples all have labels 0, so after some training this part of the validation set will have a huge loss. Everything works anyway, but the validation loss (NLL) is overestimated a lot.

Text in notebook:

z.sum().backward()
print(x.grad)
tensor([[27., 27.],
[27., 27.]])

It says "You should have got a matrix of 4.5." This code actually creates

tensor([[27., 27.],
[27., 27.]])

It looks like it should instead be:

z.sum()
print(x.grad)

The output of the net is softmaxed so it’s probabilities, but then we use a loss function that assumes them to be scores and applies softmax again. This is not a big problem in practice, but in principle it doesn’t make sense, and it clashes with the score given by Kaggle if you upload the results, because as far as I understand Kaggle assumes the outputs to be scores, whereas we have probabilities.

In https://github.com/DeepLearningDTU/02456-deep-learning-with-PyTorch/blob/master/1_Feedforward/1.2-automatic-differentiation.ipynb?short_path=5da64dc#L179, it says that the two statements are equivalent, however you cannot switch the two and run out.backward(torch.Tensor([1.0])).

Hi,

While I do not think that Docker is at all necessary to run any of the scripts with CUDA enabled up until now (Week3), I have spent way to much time trying to follow the guide for setting up the Docker container with current readme.md and dockerfile.gpu file.

Is it not possible for the course responsible to streamline and update the contents to actually reflect the current software available, so the students do not need to use extra time on top of the weekly tasks?

• Nvidia-docker is not available anymore, and is super-seeded by the "NVIDIA Container Toolkit"
• The dockerfile.gpu is using ubunto16 and we are at ubunto22 currently
• the link used for windows cpu, is not available anymore: https://rominirani.com/docker-on-windows-mounting-host-directories-d96f3f056a2c
• Can we have an alternative dockerfile that uses regular python rather than conda
• Maybe also a proper setup guide for windows GPU?

I think this would make it substantially easier for new students to figure out what they need to do when they get to week3.
Best Regards

Pull request Fix attribute names (creator -> grad_fn) changed creator to grad_fn.

Thanks to Léa RIERA for pointing out that

1.2-automatic-differentiation : I think the 1/4 before the sum was forgotten after the sum development. The final result stays correct