13:38 data type issues in training loop - Thanks for the great videos!

[JackXu9]

Hi, I've been struggling a bit with fixing one of the data type issues in the training loop related to the loss function (in the multi class classification model). I receive the following Runtimeerror:

RuntimeError: "log_softmax_lastdim_kernel_impl" not implemented for 'Long'

I have tried to match the input and target types for the loss function within the training loop (int64). I did this because when I changed the initial data conversion to tensors from torch.float to torch.LongTensor, I ran into more type issues prior to the training loop.
The code for the loop is as follows:

[aronvandepol]

Hi Jack,

Have you tried changing the tensor type to torch.float32? I managed to get through almost all with simply that datatype.

You can also see the error implies that you cannot call the loss function on a torch.LongTensor.

[aronvandepol]

Hi Jack,

After a careful look, I think I see where your problem lies. You are calling torch.softmax().argmax() on your forward_pass data (your y_logits). You want to use the logits to calculate your loss function, and not the prediction labels. Using torch.softmax() and .argmax() gives your prediction labels. if you calculate your predicition labels only when needed, you should be fine.

To conclude.

• Using torch.softmax() on your multi-class models raw outputs gives you the prediction probabilities for each class in tensor format.
• Calling .argmax() on your prediction probabilities tensor gets you your prediction labels (e.g class: 1 or 2).
• You want to use your models raw output (logits) in your loss function and not predictions probabilities or prediction labels.

Hope this helps,
Aron.

[JackXu9]

Ahh yes of course. I also had to change the y_blob to LongTensor as in the video, and then it worked. I just thought that the loss function required the same datatype for prediction and target. But thanks for the quick reply!

[aronvandepol]

To make sure, I believe they do require the same datatype, and if not; the main issue just lied with the fact that you called softmax() and argmax().

Also for next time, instead of screenshotting your code, copy it and put it in in between two grave icons `YOUR CODE IN BETWEEN` and renders out as: YOUR CODE IN BETWEEN making it easier for us to check your code.

Anyhow, happy it worked out and your model is training.

[JackXu9]

How does new line work in the code function? I can't seem to get it to create new lines with \n or ENTER.

[aronvandepol]

you can use

```
your code
hello world
```

renders as:

your code
hello world

[aronvandepol]

@mrdbourke this one is answered and solved :)

[mrdbourke]

Thank you so much for your help! :)