Notebook 04: torchmetrics.Accuracy error Section 4: Video 97.

[chriswilson2020]

In the above video you import torchmetrics 0.7.2 and your code works perfect (once Accuracy is moved to the device)

Your code:
from torchmetrics import Accuracy
torchmetric_accuracy = Accuracy().to(device)
torchmetric_accuracy(y_preds, y_blob_test)

You get the output:
tensor(0.9950, device='cuda:0')

However to reproduce your results using torchmetrics 0.11.0 which is the current version in pip I needed to explicitly pass task and num_classes into Accuracy() -> torchmetric_accuracy = Accuracy(task='multiclass', num_classes=5).to(device) to obtain the same result.

If I don't then I get the following error:

---

TypeError Traceback (most recent call last)
in
2
3 # Setup metric
----> 4 torchmetric_accuracy = Accuracy().to(device)
5
6 # Calculate accuracy

TypeError: new() missing 1 required positional argument: 'task'

I don't know if this is because of something I am doing or there was a change to torchmetrics.Accuracy()

[AlessandroMiola]

There was indeed a major change through these releases which can be tracked at Lightning-AI/torchmetrics#655 and the referenced issues. They added the task arg to discern which object to instantiate depending on the task itself (see the issue for the underlying reasons). For instance, in the multiclass case they're instantiating a MulticlassAccuracy object (https://github.com/Lightning-AI/metrics/blob/bc057d495703aff5c4d8d671c54060e08a677142/src/torchmetrics/classification/accuracy.py#L359).

[chriswilson2020]

Thanks Alessandro... I thought this course was pretty new but I guess I didn't factor in that release data does not equal creation date :D . So much content takes time to generate of course.

[mrdbourke]

Thank you @AlessandroMiola for fixing this in #203

I just merged the code!

Code before (errors):

from torchmetrics import Accuracy
torchmetrics_accuracy = Accuracy().to(device)

New code:

from torchmetrics import Accuracy
torchmetrics_accuracy = Accuracy(task='multiclass', num_classes=4).to(device)

See the updated MulticlassAccuracy() as well: https://torchmetrics.readthedocs.io/en/stable/classification/accuracy.html#multiclassaccuracy