Upgrade to HF transformers 4.3.2

[vblagoje]

@lvwerra here is a proposal for HT Transformers update to 4.3.2. The only important compatibility breaking issue I found was related to a parameter name change in GPT2Model#forward function. I also upgraded simpletransformers to a version that uses Transformers 4.3.2 as well. I have tried all notebooks in nbs directory to make sure everything works as it used to. I did not upgrade any other libs as I am not that familiar with nbdev environment.

04-gpt2-sentiment-ppo-training.ipynb works as expected and trains in about 2 hours. However, for 05-gpt2-sentiment-control.ipynb I had to lower the batch size to 128 and forward_batch_size to 4 in order to make the training work and avoid CUDA out of memory errors. The estimated running time on tqdm is 5 hours.

[vblagoje]

@lvwerra I tried this branch on both imdb ppo notebooks (the basic ppo sentiment training and the controlled sentiment ppo). They both work as expected, please try it as well. Let me know if any other checks should be done.

[lvwerra]

awesome! did you also use weights and biases? in case you did, would you mind sharing the logs?

[vblagoje]

Yes, I did but I deleted the first report for 04-gpt2-sentiment-ppo-training.ipynb. Here is the report for 05-gpt2-sentiment-control.ipynb