| Table of Contents |
|---|
PLEASE SEE UPDATES IN SECTION 1.6
Introduction
Language understanding is an ongoing challenge and one of the most relevant and influential areas across any industry.
...
BERT is a method of pre-training language representations, meaning that we train a general-purpose "language understanding" model on a large text corpus (like Wikipedia), and then use that model for downstream NLP tasks that we care about (like question answering). BERT outperforms previous methods because it is the first unsupervised, deeply bidirectional system for pre-training NLP.
1 SQuAD 1.1 with Tensorflow BERT-BASE
1.1 About the application and benchmarks
This guide is to be used as a starting point. It does not provide detailed guidance on optimizations and additional tuning. Please follow the guidelines in the Competition Limits section of this document.
1.1.1 About BERT-BASE
BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks.
...
For the purposes of this challenge, we will be using BERT-BASE.
1.1.2 About SQuAD 1.1
The Stanford Question Answering Dataset (SQuAD) is a popular question answering benchmark dataset. BERT (at the time of the release) obtains state-of-the-art results on SQuAD with almost no task-specific network architecture modifications or data augmentation. However, it does require semi-complex data pre-processing and post-processing to deal with (a) the variable-length nature of SQuAD context paragraphs, and (b) the character-level answer annotations which are used for SQuAD training. This processing is implemented and documented in run_squad.py.
1.2 Running SQuAD 1.1 fine tuning and inference
1.2.1 Using Docker and NVIDIA Docker Image
| Code Block |
|---|
docker pull nvcr.io/nvidia/tensorflow:20.02-tf1-py3 docker images REPOSITORY TAG IMAGE ID CREATED SIZE nvcr.io/nvidia/tensorflow 20.02-tf1-py3 0c7b70421b78 7 weeks ago 9.49GB |
...
| Code Block |
|---|
docker run -it --net=host -v bigdata:/bigdata 0c7b70421b78 |
1.2.2 Download the benchmark codes
Note: if you are using the docker container above, you already have the code and examples in /workspace/nvidia-examples/bert/ and can skip this step.
...
Some other examples include:
https://github.com/lambdal/bert - This is a fork of the original (Google's) BERT implementation, with added Multi-GPU support with Horovod.
1.2.3 Download BERT-BASE model file
The BERT-BASE, Uncased model file contains 12-layer, 768-hidden, 12-heads, 110M parameters. Its download link can be found at https://github.com/google-research/bert
...
| Code Block |
|---|
root@tessa002:/workspace/nvidia-examples/bert/data# mkdir download root@tessa002:/workspace/nvidia-examples/bert/data# cd download root@tessa002:/workspace/nvidia-examples/bert/data/download# mkdir -p download/google_pretrained_weights root@tessa002:/workspace/nvidia-examples/bert/data/download# cd download/google_pretrained_weights/ root@tessa002:/workspace/nvidia-examples/bert/data/download/google_pretrained_weights# wget https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip root@tessa002:/workspace/nvidia-examples/bert/data/download/google_pretrained_weights# unzip uncased_L-12_H-768_A-12.zip Archive: uncased_L-12_H-768_A-12.zip creating: uncased_L-12_H-768_A-12/ inflating: uncased_L-12_H-768_A-12/bert_model.ckpt.meta inflating: uncased_L-12_H-768_A-12/bert_model.ckpt.data-00000-of-00001 inflating: uncased_L-12_H-768_A-12/vocab.txt inflating: uncased_L-12_H-768_A-12/bert_model.ckpt.index inflating: uncased_L-12_H-768_A-12/bert_config.json |
1.2.4 Download the SQuAD 1.1 dataset
To run on SQuAD, you will first need to download the dataset. The SQuAD website does not seem to link to the v1.1 datasets any longer, but the necessary files can be found here:
...
| Code Block |
|---|
root@tessa002:/workspace/nvidia-examples/bert/data/download# mkdir -p squad/v1.1 root@tessa002:/workspace/nvidia-examples/bert/data/download# cd squad/v1.1 root@tessa002:/workspace/nvidia-examples/bert/data/download/squad# mkdir squad/v1.1# wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json root@tessa002:/workspace/nvidia-examples/bert/data/download/squad# cd v1.1/ root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1# wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1# wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1# wget https://github.com/allenai/bi-att-flow/archive/master.zip root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1# unzip master.zip root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1# cdcp bi-att-flow-master/ root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1/bi-att-flow-master# cd squad root@tessa002:/workspace/nvidia-examples/bert/data/download/squad/v1.1/bi-att-flow-master/squad# cp evaluate-v1.1.py /workspace/nvidia-examples/bert/data/download/squad/v1.1/squad/evaluate-v1.1.py . root@tessa002:cd /workspace/nvidia-examples/bert |
1.2.5 Start fine tuning
BERT representations can be fine tuned with just one additional output layer for a state-of-the-art Question Answering system. From within the container, you can use the following script to run fine-training for SQuAD.
Note : consider logging results with “>2&1 | tee $LOGFILE” for submissions to judges
...
| Code Block |
|---|
bash scripts/run_squad.sh 10 5e-6 fp16 true 4 384 128 base 1.1 data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/bert_model.ckpt 1.1 |
1.2.6 Verify results
| Code Block |
|---|
INFO:tensorflow:-----------------------------
I0326 01:25:43.144953 140630939256640 run_squad.py:1127] -----------------------------
INFO:tensorflow:Total Inference Time = 88.62 for Sentences = 10840
I0326 01:25:43.145423 140630939256640 run_squad.py:1129] Total Inference Time = 88.62 for Sentences = 10840
INFO:tensorflow:Total Inference Time W/O Overhead = 75.86 for Sentences = 10824
I0326 01:25:43.145554 140630939256640 run_squad.py:1131] Total Inference Time W/O Overhead = 75.86 for Sentences = 10824
INFO:tensorflow:Summary Inference Statistics
I0326 01:25:43.145649 140630939256640 run_squad.py:1132] Summary Inference Statistics
INFO:tensorflow:Batch size = 8
I0326 01:25:43.145738 140630939256640 run_squad.py:1133] Batch size = 8
INFO:tensorflow:Sequence Length = 384
I0326 01:25:43.145867 140630939256640 run_squad.py:1134] Sequence Length = 384
INFO:tensorflow:Precision = fp16
I0326 01:25:43.145962 140630939256640 run_squad.py:1135] Precision = fp16
INFO:tensorflow:Latency Confidence Level 50 (ms) = 55.79
I0326 01:25:43.146052 140630939256640 run_squad.py:1136] Latency Confidence Level 50 (ms) = 55.79
INFO:tensorflow:Latency Confidence Level 90 (ms) = 57.03
I0326 01:25:43.146145 140630939256640 run_squad.py:1137] Latency Confidence Level 90 (ms) = 57.03
INFO:tensorflow:Latency Confidence Level 95 (ms) = 57.29
I0326 01:25:43.146225 140630939256640 run_squad.py:1138] Latency Confidence Level 95 (ms) = 57.29
INFO:tensorflow:Latency Confidence Level 99 (ms) = 58.62
I0326 01:25:43.146308 140630939256640 run_squad.py:1139] Latency Confidence Level 99 (ms) = 58.62
INFO:tensorflow:Latency Confidence Level 100 (ms) = 286.80
I0326 01:25:43.146387 140630939256640 run_squad.py:1140] Latency Confidence Level 100 (ms) = 286.80
INFO:tensorflow:Latency Average (ms) = 56.07
I0326 01:25:43.146471 140630939256640 run_squad.py:1141] Latency Average (ms) = 56.07
INFO:tensorflow:Throughput Average (sentences/sec) = 142.68
I0326 01:25:43.146564 140630939256640 run_squad.py:1142] Throughput Average (sentences/sec) = 142.68
INFO:tensorflow:-----------------------------
I0326 01:25:43.146645 140630939256640 run_squad.py:1143] -----------------------------
INFO:tensorflow:Writing predictions to: /results/tf_bert_finetuning_squad_base_fp16_gbs40_200326010711/predictions.json
I0326 01:25:43.146801 140630939256640 run_squad.py:431] Writing predictions to: /results/tf_bert_finetuning_squad_base_fp16_gbs40_200326010711/predictions.json
INFO:tensorflow:Writing nbest to: /results/tf_bert_finetuning_squad_base_fp16_gbs40_200326010711/nbest_predictions.json
I0326 01:25:43.146886 140630939256640 run_squad.py:432] Writing nbest to: /results/tf_bert_finetuning_squad_base_fp16_gbs40_200326010711/nbest_predictions.json
{"exact_match": 78.0321665089877, "f1": 86.34229152935384}
|
...
{"exact_match": 78.0321665089877, "f1": 86.34229152935384}
1.2.7 (Optional) Alternative method with Lambda Labs
| Code Block |
|---|
root@tessa002:/workspace# mkdir lambdal root@tessa002:/workspace# cd lambdal root@tessa002:/workspace/lambdal# git clone https://github.com/lambdal/bert root@tessa002:/workspace/lambdal/lambdal# cd bert root@tessa002:/workspace/lambdal/bert# mpirun -np 4 -H localhost:4 -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH -mca pml ob1 -mca btl ^openib --allow-run-as-root python3 run_squad_hvd.py --vocab_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/vocab.txt --bert_config_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/bert_config.json --init_checkpoint=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/bert_model.ckpt --do_train=True --train_file=/workspace/nvidia-examples/bert/data/download/squad/v1.1/train-v1.1.json --do_predict=True --predict_file=/workspace/nvidia-examples/bert/data/download/squad/v1.1/dev-v1.1.json --train_batch_size=12 --learning_rate=3e-5 --num_train_epochs=2.0 --max_seq_length=384 --doc_stride=128 --output_dir=/results/lambdal/squad1/squad_base/ --horovod=true |
...
{"exact_match": 78.1929990539262, "f1": 86.51319484763773}
1.2.8 Example predict Q&A on real data
...
Example predict Q&A on real data is available here: github.com/google-research/bert
Note : This is the method that judges will use to score unseen data
| Code Block |
|---|
root@tessa002:/workspace/nvidia-examples/bert# cd /workspace root@tessa002:/workspace# git clone https://github.com/google-research/bert.git root@tessa002:/workspace# cd bert |
1.2.9 Create a sample input file
Create a simple input file, save as test_input.json in json format (note the "id" to reference later).
...
| Code Block |
|---|
{
"version": "v1.1",
"data": [
{
"title": "your_title",
"paragraphs": [
{
"qas": [
{
"question": "Who is current CEO?",
"id": "56ddde6b9a695914005b9628",
"is_impossible": ""
},
{
"question": "Who founded google?",
"id": "56ddde6b9a695914005b9629",
"is_impossible": ""
},
{
"question": "when did IPO take place?",
"id": "56ddde6b9a695914005b962a",
"is_impossible": ""
}
],
"context": "Google was founded in 1998 by Larry Page and Sergey Brin while they were Ph.D. students at Stanford University in California. Together they own about 14 percent of its shares and control 56 percent of the stockholder voting power through supervoting stock. They incorporated Google as a privately held company on September 4, 1998. An initial public offering (IPO) took place on August 19, 2004, and Google moved to its headquarters in Mountain View, California, nicknamed the Googleplex. In August 2015, Google announced plans to reorganize its various interests as a conglomerate called Alphabet Inc. Google is Alphabet's leading subsidiary and will continue to be the umbrella company for Alphabet's Internet interests. Sundar Pichai was appointed CEO of Google, replacing Larry Page who became the CEO of Alphabet."
}
]
}
]
}
|
1.2.10 Run run_squad.py
Run run_squad.py as do-predict=true using fine-tuned model checkpoint :
| Code Block |
|---|
root@tessa002:/workspace/bert# python3 run_squad.py --vocab_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/vocab.txt --bert_config_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/bert_config.json --init_checkpoint=/results/tf_bert_finetuning_squad_base_fp16_gbs40_200326010711/model.ckpt-2408 --do_train=False --max_query_length=30 --do_predict=True --predict_file=test_input.json --predict_batch_size=16 --max_seq_length=384 --doc_stride=128 --output_dir=/results/squad1/squad_test/ |
...
| Code Block |
|---|
root@tessa002:/workspace/lambdal/bert# python3 run_squad.py --vocab_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/vocab.txt --bert_config_file=/workspace/nvidia-examples/bert/data/download/google_pretrained_weights/uncased_L-12_H-768_A-12/bert_config.json --init_checkpoint=/results/lambdal/squad1/squad_base/model.ckpt-3649 --do_train=False --max_query_length=30 --do_predict=True --predict_file=test_input.json --predict_batch_size=16 --max_seq_length=384 --doc_stride=128 --output_dir=/results/lambdal/squad1/squad_test/ |
...
You should see similar output below
...
| Code Block |
|---|
I0326 02:11:40.096473 140685488179008 run_squad.py:1259] Processing example: 0 INFO:tensorflow:prediction_loop marked as finished I0326 02:11:40.165820 140685488179008 error_handling.py:101] prediction_loop marked as finished INFO:tensorflow:prediction_loop marked as finished I0326 02:11:40.166095 140685488179008 error_handling.py:101] prediction_loop marked as finished INFO:tensorflow:Writing predictions to: /results/squad1/squad_test/predictions.json I0326 02:11:40.166555 140685488179008 run_squad.py:745] Writing predictions to: /results/squad1/squad_test/predictions.json INFO:tensorflow:Writing nbest to: /results/squad1/squad_test/nbest_predictions.json I0326 02:11:40.166669 140685488179008 run_squad.py:746] Writing nbest to: /results/squad1/squad_test/nbest_predictions.json |
1.2.
...
11 Check correctness in file : predictions.json
| Code Block |
|---|
{
"56ddde6b9a695914005b9628": "Sundar Pichai",
"56ddde6b9a695914005b9629": "Larry Page and Sergey Brin",
"56ddde6b9a695914005b9630": "September 4, 1998",
"56ddde6b9a695914005b9631": "CEO",
"56ddde6b9a695914005b9632": "Alphabet Inc"
}
|
1.2.
...
12 Check accuracy in file: nbest_predictions.json
| Code Block |
|---|
{
"56ddde6b9a695914005b9628": [
{
"text": "Sundar Pichai",
"probability": 0.6877274611974046,
"start_logit": 7.016119003295898,
"end_logit": 6.917689323425293
},
{
"text": "Sundar Pichai was appointed CEO of Google, replacing Larry Page",
"probability": 0.27466839794889614,
"start_logit": 7.016119003295898,
"end_logit": 5.999861240386963
},
{
"text": "Larry Page",
"probability": 0.02874494871571203,
"start_logit": 4.759016513824463,
"end_logit": 5.999861240386963
},
|
...
Scores will be derived from the nbest_predictions.json output for each question on the context.
1.3 Challenge Limitation
Must stick to pre-defined model (BERT-Base, Uncased)
Teams can locally cache (on SSD) starting model weights and dataset
HuggingFace implementation (TensorFlow/PyTorch) is the official standard. Usage of other implementation, or modification to official, is subject to approval.
Teams are allowed to explore different optimizers (SGD/Adam etc.) or learning rate schedules, or any other techniques that do not modify model architecture.
Teams are not allowed to modify any model hyperparameters or add additional layers.
Entire model must be fine-tuned (cannot freeze layers)
You must provide all scripts and methodology used to achieve results
1.4 Teams must produce
Training scripts with their full training routine and command lines and output
Evaluation-only script for verification of result. Final evaluation is on a fixed sequence length (128 tokens).
Final model ckpt and inference files
Team’s training scripts and methodology, command line and logs of runs
run_squad.py predictions.json and nbest_predictions.json
1.5 Final Scoring
The judges will score with standard evaluate-v1.1.py from Squad 1.1
...
Final scores from unseen data of multiple questions; prediction from file, using standard run_squad.py
1.6 UPDATES (June 8, 2020)
In past discussion we had questions on training BERT from scratch; this is beyond the scope of this competition and is not allowed. You will need to use the BERT-BASE model file as outlined in the guidelines section 1.2.3
Change/modify the output layer and to allow additional layers
Allow for ensemble techniques
We must disallow integration of dev-set data into training dataset ; the SQUAD 1.1 datasets must remain unchanged / augmented
We must disallow additional external data integrated into training dataset for this competition because there is not enough time to be able to verify that the dev-set data might inadvertently be part of that acquired dataset augmentation
We allow any hyper-parameters ; ie. learn rate, optimizer, drop-out, etc.
We will also allow setting for random seed. This will reduce the variance between training runs
The F1 score will be used as score for team ranking.
Teams should submit their best 5 runs, please upload your runs in separate folders containing ckpt, logs, etc. - you/we will average top 3 of the 5 f1 scores for your final score
We will use the F1 as the quality metric for score / ranking. We will not round the output score computed from the output of the evaluate-v1.1.py.
The judges will score with standard evaluate-v1.1.py from Squad 1.1 as outlined in section 1.5 of the SQuAD 1.1 with Tensorflow BERT-BASE Guidelines.
We will use the probability score for unseen inference data (as test_input.json) to be provided no later than June 10th, as a secondary ranking in the event of any tie to the f1 average scoring against your top training run.