Environment info

• transformers version: 4.11.2
• Platform: Linux-5.4.0-66-generic-x86_64-with-glibc2.10
• Python version: 3.8.8
• PyTorch version (GPU?): 1.9.1+cu102 (True)
• Tensorflow version (GPU?): not installed (NA)
• F

🐛 Bug

The model notebook at GColab you link when creating an issue is outdated. Tries to import metrics from pytorch lightning:

from pytorch_lightning.metrics.functional import accuracy

I believe this is because pytorch_lightning doesn't include torchmetrics anymore, and they are a completely separate library.

To Reproduce

Go to the link and try to run it:

Change tensor.data to tensor.detach() due to
pytorch/pytorch#6990 (comment)
tensor.detach() is more robust than tensor.data.

🙌🙌🙌🙌🙌

New Operator

Describe the operator

Why is this operator necessary? What does it accomplish?

This is a frequently used operator in tensorflow/keras

Can this operator be constructed using existing onnx operators?

If so, why not add it as a function?

I don't know.

Is this operator used by any model currently? Which one?

Are you willing to contribute it?

f-strings offer better readability/performance than str.format and %, so we should use them in all places in our codebase unless there is good reason to keep the older syntax.

NOTE FOR CONTRIBUTORS: To avoid large PRs and possible merge conflicts, do 1-3 modules per PR. Also, feel free to ignore the files located under datasets/*.

Is your feature request related to a problem? Please describe.
I typically used compressed datasets (e.g. gzipped) to save disk space. This works fine with AllenNLP during training because I can write my dataset reader to load the compressed data. However, the predict command opens the file and reads lines for the Predictor. This fails when it tries to load data from my compressed files.