daniel-huang-1230 / neural_network_nlp_cmu_numerical_reasoning_in_bert

11-747-Projects

Contributors:

1. Daniel Huang, [email protected]
2. Yi Shen (Eason), [email protected]
3. Tianhao Wang (Danny), [email protected]

HW3:

The /code directory contains the source code from https://github.com/facebookresearch/SpanBERT. The only modification is on SpanBERT/code/run_squad.py, which incorporates the commit from pull request facebookresearch/SpanBERT#60 that is necessary for us to run with the apex library successfully.

We (almost) reproduced the F1 score 83.6 with SpanBERT-base model evaluated on SQuAD 2.0 dev sets. The detail steps can be found in our Colab notebook https://github.com/gn03249822/11-747-Projects/blob/main/HW3_SpanBERT_reproduced_f1.ipynb. In terms of runtime, we run in GPU + high RAM mode with my personal Colab Pro account. Notice that even with this setup, the recommended config of finetuning on SQuAD 2.0 by Facebook still cause GPU out-of-memory error. After experimentation, we decreased the train and eval batchsizes, both from 32 to 16, and was able to finish training on the downstream finetuning task in around 7.5 hours.

The final result can be found in /code/SpanBERT/squad_output/eval_results.txt

HasAns_exact = 81.66329284750337
HasAns_f1 = 88.02528196756217
HasAns_total = 5928
NoAns_exact = 70.95037846930194
NoAns_f1 = 70.95037846930194
NoAns_total = 5945
batch_size = 16
best_exact = 80.63673881916955
best_exact_thresh = -5.0234375
best_f1 = 83.34820529512551
best_f1_thresh = -5.0234375
epoch = 3
exact = 76.29916617535585
f1 = 79.47560612344888
global_step = 25311
learning_rate = 2e-05
total = 11873

The best f1 score we achieved by using the pretrained SpanBERT-base and training on SQuAD 2.0 ourself is 83.35 and Facebook Research team reported 83.6 with their same fune-tuning model, as shown above. We attribute this slightly-worse performance delta to the smaller batch_size we use due to the constraint of our computational resources.

Finally, we omitted to commit our downloaded pretrained model in this repo, as the model exceeds the 100mb size limit of a single file on GitHub. The original path should be /code/SpanBERT/squad_output/pytorch_model.bin, and one can be found it along with the model config (hyperparameters) in https://huggingface.co/SpanBERT/spanbert-base-cased/tree/main.

HW4:

It contains the source code of GenBERT with our re-implementation modification to perform pruning an quantization.
Please refer to the Google Colab notebook https://github.com/gn03249822/11-747-Projects/blob/main/Inject_prune.ipynb for the execution results.