We need to replay this addition that has been modelled in run_translation.py in huggingface/transformers#12351 to all other pytorch examples

The actual changes for the model example are:
https://github.com/huggingface/transformers/pull/12351/files#diff-09777f56cee1060a535a72ce99a6c96cdb7f330c8cc3f9dcca442b3f7768237a
(just run_translation.py)

Here is a time-saver

🐛 Bug

File "/home/ubuntu/vqa/GMN/mmf/mmf/datasets/builders/visual_genome/dataset.py", line 44, in init
scene_graph_file = self._get_absolute_path(scene_graph_file)
AttributeError: 'VisualGenomeDataset' object has no attribute '_get_absolute_path'

Command that i run in shell

CUDA_VISIBLE_DEVICES="0" mmf_run config=projects/gmn/configs/visual_genome/defaults.yaml model=gm

Hi
when I do classification with the pretrained vggface2 model, how do I match a probability to a name? do you have a list of labels for the trained model?

Thanks

1.希望可以把底层的api文档再完善一些，比如encoder,decoder，以便于复现一些论文
2.希望可以维护一个pytorch和paddle的api对照一览表，尽量全一些
3.错误日志能否准确一些，有时候datalaoder出的一些错误信息不好定位
4.能否增加使用梯度累加特性，进一步提高batch size

when printing these objects we do not see informative content.

The Mock object shouldn't be there..

RTD

Latest
![image](https://user-images.githubusercontent.com/18496796/114235740-e910fb80-9980-11eb