This document has advanced instructions for running BERT Large Bfloat16
training, which provides more control over the individual parameters that
are used. For more information on using /benchmarks/launch_benchmark.py,
see the launch benchmark documentation.
Prior to using these instructions, please follow the setup instructions from
the model's README and/or the
AI Kit documentation to get your environment
setup (if running on bare metal) and download the dataset, pretrained model, etc.
If you are using AI Kit, please exclude the --docker-image flag from the
commands below, since you will be running the the TensorFlow conda environment
instead of docker.
Any of the launch_benchmark.py commands below can be run on bare metal by
removing the --docker-image arg. Ensure that you have all of the
required prerequisites installed in your environment
before running without the docker container.
If you are new to docker and are running into issues with the container, see this document for troubleshooting tips.
Once your environment is setup, navigate to the benchmarks directory of
the model zoo and set environment variables for the dataset, checkpoint
directory, and an output directory where log files will be written.
cd benchmarks
export CHECKPOINT_DIR=<path to the pretrained bert model directory>
export DATASET_DIR=<path to GLUE data for classifier training or to the SQuAD data for SQuAD training>
export OUTPUT_DIR=<directory where checkpoints and log files will be saved>
BERT Large training can be run in three different modes:
-
To run SQuAD training use the following command with the
train_option=SQuAD. TheCHECKPOINT_DIRshould point to the location where you've downloaded the BERT large (whole word masking) pretrained model, and theDATASET_DIRshould point to the SQuAD data.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=24 \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ --output-dir $OUTPUT_DIR \ -- train_option=SQuAD \ vocab_file=$CHECKPOINT_DIR/vocab.txt \ config_file=$CHECKPOINT_DIR/bert_config.json \ init_checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ do_train=True \ train_file=$DATASET_DIR/train-v1.1.json \ do_predict=True \ predict_file=$DATASET_DIR/dev-v1.1.json \ learning_rate=3e-5 \ num_train_epochs=2 \ max_seq_length=384 \ doc_stride=128 \ optimized_softmax=True \ experimental_gelu=False \ do_lower_case=TrueThe dev set predictions will be saved to a file called predictions.json in the output directory.
python ${DATASET_DIR}/evaluate-v1.1.py ${DATASET_DIR}/dev-v1.1.json ${OUTPUT_DIR}/predictions.jsonAn execution with these parameters produces results in line with below scores:
Bf16: {"exact_match": 86.77388836329234, "f1": 92.98642358746287} FP32: {"exact_match": 86.72658467360453, "f1": 92.98046893150796}To run distributed training of SQuAD: for better throughput, simply specify
--mpi_num_processes=<num_of_sockets> [--mpi_num_processes_per_socket=1]. To set--mpi_num_processes=<num_of_sockets>, please runlscpuon your machine to check the available number of sockets.Note:
- the
global batch sizeismpi_num_processes * train_batch_sizeand sometimeslearning rateneeds to be adjusted for convergence. square root learning rate scalingis used by default.- for BERT fine-tuning, state-of-the-art accuracy can be achieved via parallel training without synchronizing gradients between MPI workers.
- The
--mpi_workers_sync_gradients=[True/False]controls whether the MPI workers sync gradients.By default it is set toFalsemeaning the workers are training independently and the best performing training results will be picked in the end. - To enable gradients synchronization, set the
--mpi_workers_sync_gradientstoTruein BERT-specific options. - The options
optimized_softmax=Truecan be set for better performance.
Note : Add a space after
--, for BERT-specific options.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=24 \ --mpi_num_processes=<num_of_sockets> \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --output-dir $OUTPUT_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ -- train_option=SQuAD \ vocab_file=$CHECKPOINT_DIR/vocab.txt \ config_file=$CHECKPOINT_DIR/bert_config.json \ init_checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ do_train=True \ train_file=$DATASET_DIR/train-v1.1.json \ do_predict=True \ predict_file=$DATASET_DIR/dev-v1.1.json \ learning_rate=3e-5 \ num_train_epochs=2 \ max_seq_length=384 \ doc_stride=128 \ optimized_softmax=True \ experimental_gelu=False \ do_lower_case=TrueThe results file will be written to the
${OUTPUT_DIR}directory. - the
-
To run Classifier training, use the following command with the
train-option=Classifier. For classifier training, use the GLUE data in as theDATASET_DIR. TheCHECKPOINT_DIRshould point to the BERT base uncased 12-layer, 768-hidden pretrained model that you've downloaded. This example code fine-tunes BERT-Base on the Microsoft Research Paraphrase Corpus (MRPC) corpus, which only contains 3,600 examples.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=32 \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --output-dir $OUTPUT_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ -- train-option=Classifier \ task-name=MRPC \ do-train=true \ do-eval=true \ data-dir=$DATASET_DIR/MRPC \ vocab-file=$CHECKPOINT_DIR/vocab.txt \ config-file=$CHECKPOINT_DIR/bert_config.json \ init-checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ max-seq-length=128 \ learning-rate=2e-5 \ num-train-epochs=30 \ optimized_softmax=True \ experimental_gelu=False \ do-lower-case=TrueThe results file will be written to the
${OUTPUT_DIR}directory.To run distributed training of Classifier: for better throughput, specify
--mpi_num_processes=<num_of_sockets> [--mpi_num_processes_per_socket=1], please runlscpuon your machine to check the available number of sockets. Note that theglobal batch sizeismpi_num_processes * train_batch_sizeand sometimes learning rate needs to be adjusted for convergence. By default, the script uses square root learning rate scaling.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=32 \ --mpi_num_processes=<num_of_sockets> \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --output-dir $OUTPUT_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ -- train-option=Classifier \ task-name=MRPC \ do-train=true \ do-eval=true \ data-dir=$DATASET_DIR/MRPC \ vocab-file=$CHECKPOINT_DIR/vocab.txt \ config-file=$CHECKPOINT_DIR/bert_config.json \ init-checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ max-seq-length=128 \ learning-rate=2e-5 \ num-train-epochs=30 \ optimized_softmax=True \ experimental_gelu=True \ do-lower-case=TrueThe results file will be written to the
${OUTPUT_DIR}directory. -
To run pre-training from scratch, use the following command with
train-option=Pretraining. Pre-training has two phases: In the first phase, the data is generated for sequential length 128. And in the second phase, sequential length 512 is used. Please follow instructions in google bert pre-training for data pre-processing and set theDATASET_DIRenvironment variable to the TF record directory. TheCHECKPOINT_DIRshould point to the BERT large uncased (whole word masking) directory.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=24 \ --socket-id=0 \ --num-intra-threads=24 \ --num-inter-threads=1 \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --output-dir $OUTPUT_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ -- train-option=Pretraining \ input-file=$DATASET_DIR/tf_wiki_formatted_slen512.tfrecord \ do-train=True \ do-eval=True \ config-file=$CHECKPOINT_DIR/bert_config.json \ init-checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ max-seq-length=512 \ max-predictions=76 \ num-train-steps=20 \ warmup-steps=10 \ learning-rate=2e-5 \ optimized_softmax=True \ experimental_gelu=False \ profile=FalseTo run distributed training of pretraining: for better throughput, simply specify
--mpi_num_processes=<num_of_sockets> [--mpi_num_processes_per_socket=1]. Please runlscpuon your machine to check the available number of sockets.Note that the
global batch sizeismpi_num_processes * train_batch_sizeand sometimeslearning rateneeds to be adjusted for convergence. By default, the script usessquare root learning rate scaling.python launch_benchmark.py \ --model-name=bert_large \ --precision=bfloat16 \ --mode=training \ --framework=tensorflow \ --batch-size=24 \ --num-intra-threads=22 \ --num-inter-threads=1 \ --mpi_num_processes=4 \ --volume $CHECKPOINT_DIR:$CHECKPOINT_DIR \ --volume $DATASET_DIR:$DATASET_DIR \ --output-dir $OUTPUT_DIR \ --docker-image intel/intel-optimized-tensorflow:latest \ -- train-option=Pretraining \ input-file=$DATASET_DIR/tf_wiki_formatted_slen512.tfrecord \ do-train=True \ do-eval=True \ config-file=$CHECKPOINT_DIR/bert_config.json \ init-checkpoint=$CHECKPOINT_DIR/bert_model.ckpt \ max-seq-length=512 \ max-predictions=76 \ num-train-steps=20 \ warmup-steps=10 \ learning-rate=2e-5 \ optimized_softmax=True \ experimental_gelu=False \ profile=FalseNote: for best performance, we will set
num-intra-threadas follows:- For single instance (mpi_num_processes=1) run: the value is equal to number of logical cores per socket.
- For multi-instance (mpi_num_processes>1) run: the value is equal to (num_of_logical_cores_per_socket - 2)