Daisy Deng

In classification training_accuracy_vs_speed notebook, consider to add all model types and show accuracy, speed, and memory footprint result as a table for all the model types so it's much easier...

### Description abstractive_summarization_bertsumabs_cnndm notebook is outdated. Currently tested with torch==1.6.0 torchtext==0.9.1 torchvision==0.7.0 transformers==2.9.0 ### Related Issues ### Checklist: - [x] My code follows the code style of this project, as...

Daden/bartt5 add text summarization for BART and T5 models

2

### Description added bart/t5 finetuning and test functionality added a notebook example for t5-small model with performance verified ### Related Issues ### Checklist: - [x] My code follows the code...

### Description Currently, in the common.py for transformer models, a checkpoint only saves model state and the optimizer and amp state info is not saved. We can consider saving this...

### Description ### Expected behavior with the suggested feature ### Other Comments

[ASK] [tc_multi_languages_transformers.ipynb] temporary data directory and cache directory are not deleted after the notebook run

9

### Description The temporary directories should be deleted after the notebook run is finished. ### Other Comments consider https://security.openstack.org/guidelines/dg_using-temporary-files-securely.html

[Feature] add horovod and add the tests back

1

### Description Currently horovod installation is not included in the environment generation scripts and documentation of horovod is not complete. So we temporarily removed all the tests related to horovod....

### Description GPU is scare resource and we should provide users of this git repo a way to reuse their gpu machines if they already have one. ### Other Comments

[ASK] Merge preprocess_encoder_tokens and preprocess_classification_tokens

1

### Description preprocess_encoder_tokens and preprocess_classification_tokens contain large chunk of same code and we should rewrite this part of not repeat the same code. ### Other Comments

### Description currently the device selection is determined by this line device = get_device("cpu" if num_gpus == 0 else "gpu") The code can be more robust by considering torch.cuda.is_available() ###...