Related to #5142, AlbertTokenizer (which uses SentencePiece) doesn't decode special tokens (like [CLS], [MASK]) properly. This issue was discovered when adding the Nystromformer model (#14659), which uses this tokenizer.

To reproduce (Transformers v4.15 or below):

!pip install -q transformers sentencepiece

from transformers import AlbertTokenizer

tokenizer = AlbertTokenizer.from

I want to train a detector based on object365 dataset, but object365 is pretty large, and caused out of memory error in my computer.
I want to split the annotation file to 10, such as ann1,ann2,...ann10, then build 10 datasets and concatenate them, but I'm not sure whether it's
gonna work or not.
Any better suggestion?

All the docs still say:

Trainer(gpus=...)

But we now support a different syntax:

Trainer(devices=4, accelerator='gpu')

Although we have backward support, all the docs (including readme) needs to reflect the new way of doing it

cc @Borda @rohitgr7 @akihironitta @ananthsub @carmocca @Borda

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

🚀 Feature

Motivation

paper "LEARNING TO REPRESENT PROGRAMS WITH GRAPHS" which encode computer programs as graphs, with rich semantic information, however, most code implementation on this dataset VarMisuse is based on TensorFlow, like [tf-gnn-samples](https://github.com/microsof

Although the results look nice and ideal in all TensorFlow plots and are consistent across all frameworks, there is a small difference (more of a consistency issue). The result training loss/accuracy plots look like they are sampling on a lesser number of points. It looks more straight and smooth and less wiggly as compared to PyTorch or MXNet.

It can be clearly seen in chapter 6([CNN Lenet](ht

Is your feature request related to a problem? Please describe.

I am uploading our dataset and models for the "Constructing interval measures" method we've developed, which uses item response theory to convert multiple discrete labels into a continuous spectrum for hate speech. Once we have this outcome our NLP models conduct regression rather than classification, so binary metrics are not r

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?

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.