SQuad With Transformer by HuggingFace

Setup

Run the default setup first! See READEME_bidaf

SQuAD

Based on the script run_squad.py.

Benchmarks

Model	EM	F1	NoAns_f1	HasAns_f1	batch_size	epochs
BERT-base	73.46	76.67	72.92	80.76	8	2
BERT-large	81.56	84.92	84.91	84.92	8	2
RoBERTa-base	78.82	82.16	80.93	83.49	8	2
RoBERTa-large	82.82	86.32	87.59	84.94	4	2
RoBERTa-large	83.35	86.62	86.36	86.43	6	3
ALBERT-base-v2	78.38	81.50	81.50	81.51	8	2
ALBERT-large-v2	81.16	84.22	83.24	85.29	6	2
ALBERT-xxlarge-v1	86.23	89.23	87.15	90.63	4	3

Benchmarks For CLS Model

Model	EM	F1	NoAns_f1	HasAns_f1	batch_size	epochs
Bert-base	73.74	76.69	75.98	77.47	8	2
Roberta-base	79.04	82.24	82.17	82.33	8	2
ALBERT-base-v2	78.91	82.14	81.91	82.38	8	2

Ensemble 7 models

Ensemble Method	EM	F1
Grid Search Most Confidence	87.33	89.81

Ensemble Benchmarks for ALBERT-xxlarge-v1 + ALBERT-large-v2

Ensemble Method	EM	F1	Weights
Most Confidence	85.29	87.75	x
Grid Search Most Confidence	86.80	89.49	4 : 3
Weighted Most Confidence	86.89	89.55	0.8367 : 0.6530
Ensemble Stacking	86.72	89.51	x

Ensemble albert family (2 albert-large, 1 albert-xxlarge)

Ensemble Method	EM	F1
Most Confidence	85.97	88.33
Grid Search Most Confidence	86.80	89.49

Ensemble Benchmarks for BERT-Family (BERT-large + BERT-large-batch6 + BERT-large-cls)

Ensemble Method	EM	F1	Weights
Most Confidence	83.38	86.12	x
Grid Search Most Confidence	83.60	86.30	5 : 5 : 4
Weighted Most Confidence	83.05	86.04	1 :0.99 : 1.05
Ensemble Stacking	82.81	85.83	x

Ensemble

Run ensemble weighted train

python ensemble_squad.py \
  --name ensemble-weighted \
  --model_type placeholder \
  --model_name_or_path placeholder \
  --do_train \
  --do_eval \
  --do_weighted_ensemble \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --per_gpu_train_batch_size 512 \
  --num_train_epochs 200 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --predict_file data/dev-v2.0.json \
  --saved_processed_data_dir output/saved_data_6 \
  --logging_steps 1 \
  --save_steps 1000 \
  --evaluate_during_saving  \
  --save_best_only \
  --learning_rate 5e-5

Run ensemble stack train

python ensemble_squad.py \
  --name ensemble-weighted \
  --model_type placeholder \
  --model_name_or_path placeholder \
  --do_train \
  --do_eval \
  --do_stack_ensemble \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --per_gpu_train_batch_size 512 \
  --num_train_epochs 50 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --predict_file data/dev-v2.0.json \
  --saved_processed_data_dir output/saved_data_6 \
  --logging_steps 1 \
  --save_steps 1000 \
  --evaluate_during_saving  \
  --save_best_only \
  --learning_rate 5e-5

Run ensemble voting

python ensemble_squad.py \
  --name ensemble-voting \
  --model_type placeholder \
  --model_name_or_path placeholder \
  --do_ensemble_voting \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --predict_file data/dev-v2.0.json \
  --saved_processed_data_dir output/only_dev_1

Training

For SQuAD2.0 example, you could run ./run_squad.sh

For BERT-base on 8GB RAM GPU

python run_squad.py \
  --name bert-test-1 \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 500 \
  --save_steps 5000

For BERT-large on 12GB RAM GPU

python run_squad.py \
  --name bert-large \
  --model_type bert \
  --model_name_or_path bert-large-uncased-whole-word-masking-finetuned-squad \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 4000

For RoBERTa-base on 8GB RAM GPU

python run_squad.py \
  --name roberta-base \
  --model_type roberta \
  --model_name_or_path roberta-base \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 2000

For RoBERTa-large on 12GB RAM GPU

python run_squad.py \
  --name roberta-large \
  --model_type roberta \
  --model_name_or_path roberta-large \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 6 \
  --learning_rate 3e-5 \
  --num_train_epochs 3 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --seed 91 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 3000

For ALBERT-base-v2 on 8GB RAM GPU

python run_squad \
  --name albert-base-v2 \
  --model_type albert \
  --model_name_or_path albert-base-v2 \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 50 \
  --save_steps 2000

For ALBERT-large-v2 on 12GB RAM GPU

python run_squad.py \
  --name albert-large-v2 \
  --model_type albert \
  --model_name_or_path albert-large-v2 \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 6 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 50 \
  --save_steps 2000

For ALBERT-xlarge-v2 on 12GB RAM GPU (＃TODO, takes 48 hours to train)

python run_squad.py \
  --name albert-xlarge-v2 \
  --model_type albert \
  --model_name_or_path albert-xlarge-v2 \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 4 \
  --learning_rate 3e-5 \
  --num_train_epochs 3.0 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 3000

For ALBERT-xxlarge-v1 on NC24 (22hr per epoch)

python run_squad.py   
  --name albert-xxlarge-v1 \
  --model_type albert \
  --model_name_or_path albert-xxlarge-v1 \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 4  \
  --learning_rate 3e-5  \
  --num_train_epochs 3.0 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving  \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 2000

For XLNet-base-cased on 12GB RAM K80 (currently not working)

python run_squad.py \
  --name xlnet-test \
  --model_type xlnet \
  --model_name_or_path xlnet-base-cased \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 500 \
  --save_steps 5000

Run CLS model

python run_squad_cls.py \
  --name bert-base-cls \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 4 \
  --learning_rate 3e-5 \
  --num_train_epochs 3 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative \
  --evaluate_during_saving \
  --save_best_only \
  --logging_steps 100 \
  --save_steps 3000 \
  --seed 123 \
  --gradient_accumulation_steps 2

Dev Testing

Test for run_squad.py

python run_squad.py \
  --name bert-base-test \
  --model_type bert \
  --model_name_or_path save/train/bert-base-test-01 \
  --do_eval \
  --do_lower_case \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative

Test for run_squad_cls.py

python run_squad_cls.py \
  --name bert-base-cls \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --eval_dir save/train/bert-base-cls-01/cur_best \
  --do_eval \
  --do_lower_case \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --version_2_with_negative

Generate model outputs

python run_squad.py \
  --name albert-xxlarge-v1 \
  --model_type albert \
  --model_name_or_path save/train/albert-xxlarge-v1/cur_best \
  --do_output \
  --do_lower_case \
  --train_file data/train-v2.0.json \
  --predict_file data/dev-v2.0.json \
  --per_gpu_train_batch_size 8 \
  --max_seq_length 256 \
  --doc_stride 128 \
  --version_2_with_negative \
  --per_gpu_eval_batch_size 128

Original Training script

Fine-tuning BERT on SQuAD1.0

This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a $SQUAD_DIR directory.

And for SQuAD2.0, you need to download:

Below is the example of training BERT for SQuAD1.1

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
  --model_type bert \
  --model_name_or_path bert-base-cased \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file $SQUAD_DIR/train-v1.1.json \
  --predict_file $SQUAD_DIR/dev-v1.1.json \
  --per_gpu_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir /tmp/debug_squad/

Training with the previously defined hyper-parameters yields the following results on Squad 1.1:

f1 = 88.52
exact_match = 81.22

Distributed training

Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.1:

python -m torch.distributed.launch --nproc_per_node=8 ./examples/run_squad.py \
    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking \
    --do_train \
    --do_eval \
    --do_lower_case \
    --train_file $SQUAD_DIR/train-v1.1.json \
    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
    --per_gpu_eval_batch_size=3   \
    --per_gpu_train_batch_size=3   \

Training with the previously defined hyper-parameters yields the following results:

f1 = 93.15
exact_match = 86.91

This fine-tuned model is available as a checkpoint under the reference bert-large-uncased-whole-word-masking-finetuned-squad.

Fine-tuning XLNet on SQuAD

This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .

Command for SQuAD1.0:

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
    --model_type xlnet \
    --model_name_or_path xlnet-large-cased \
    --do_train \
    --do_eval \
    --do_lower_case \
    --train_file $SQUAD_DIR/train-v1.1.json \
    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
    --per_gpu_eval_batch_size=4  \
    --per_gpu_train_batch_size=4   \
    --save_steps 5000

Command for SQuAD2.0:

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
    --model_type xlnet \
    --model_name_or_path xlnet-large-cased \
    --do_train \
    --do_eval \
    --version_2_with_negative \
    --train_file $SQUAD_DIR/train-v2.0.json \
    --predict_file $SQUAD_DIR/dev-v2.0.json \
    --learning_rate 3e-5 \
    --num_train_epochs 4 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
    --per_gpu_eval_batch_size=2  \
    --per_gpu_train_batch_size=2   \
    --save_steps 5000

Larger batch size may improve the performance while costing more memory.

Results for SQuAD1.0 with the previously defined hyper-parameters:

{
"exact": 85.45884578997162,
"f1": 92.5974600601065,
"total": 10570,
"HasAns_exact": 85.45884578997162,
"HasAns_f1": 92.59746006010651,
"HasAns_total": 10570
}

Results for SQuAD2.0 with the previously defined hyper-parameters:

{
"exact": 80.4177545691906,
"f1": 84.07154997729623,
"total": 11873,
"HasAns_exact": 76.73751686909581,
"HasAns_f1": 84.05558584352873,
"HasAns_total": 5928,
"NoAns_exact": 84.0874684608915,
"NoAns_f1": 84.0874684608915,
"NoAns_total": 5945
}

Name		Name	Last commit message	Last commit date
Latest commit History 104 Commits
data		data
models		models
output		output
save		save
.gitignore		.gitignore
CS224N__Project_Milestone.pdf		CS224N__Project_Milestone.pdf
CS224N__Project_Proposal.pdf		CS224N__Project_Proposal.pdf
LICENSE		LICENSE
README.md		README.md
README_bidaf.md		README_bidaf.md
args.py		args.py
check_data.py		check_data.py
ensemble_squad.py		ensemble_squad.py
environment.yml		environment.yml
generate_combined_output.py		generate_combined_output.py
hack.py		hack.py
notebook.ipynb		notebook.ipynb
run_squad.py		run_squad.py
run_squad.sh		run_squad.sh
run_squad_cls.py		run_squad_cls.py
setup.py		setup.py
tensorboard.sh		tensorboard.sh
test.py		test.py
train_bidaf.py		train_bidaf.py
util.py		util.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

SQuad With Transformer by HuggingFace

Setup

SQuAD

Benchmarks

Benchmarks For CLS Model

Ensemble 7 models

Ensemble Benchmarks for ALBERT-xxlarge-v1 + ALBERT-large-v2

Ensemble albert family (2 albert-large, 1 albert-xxlarge)

Ensemble Benchmarks for BERT-Family (BERT-large + BERT-large-batch6 + BERT-large-cls)

Ensemble

Run ensemble weighted train

Run ensemble stack train

Run ensemble voting

Training

Run CLS model

Dev Testing

Generate model outputs

Original Training script

Fine-tuning BERT on SQuAD1.0

Distributed training

Fine-tuning XLNet on SQuAD

Command for SQuAD1.0:

Command for SQuAD2.0:

Results for SQuAD1.0 with the previously defined hyper-parameters:

Results for SQuAD2.0 with the previously defined hyper-parameters:

About

Releases

Packages

Contributors 2

Languages

License

alvinbhou/Stanford-CS224n-QA-Project

Folders and files

Latest commit

History

Repository files navigation

SQuad With Transformer by HuggingFace

Setup

SQuAD

Benchmarks

Benchmarks For CLS Model

Ensemble 7 models

Ensemble Benchmarks for ALBERT-xxlarge-v1 + ALBERT-large-v2

Ensemble albert family (2 albert-large, 1 albert-xxlarge)

Ensemble Benchmarks for BERT-Family (BERT-large + BERT-large-batch6 + BERT-large-cls)

Ensemble

Run ensemble weighted train

Run ensemble stack train

Run ensemble voting

Training

Run CLS model

Dev Testing

Generate model outputs

Original Training script

Fine-tuning BERT on SQuAD1.0

Distributed training

Fine-tuning XLNet on SQuAD

Command for SQuAD1.0:

Command for SQuAD2.0:

Results for SQuAD1.0 with the previously defined hyper-parameters:

Results for SQuAD2.0 with the previously defined hyper-parameters:

About

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Languages

Packages