Benchmark OpenNMT

This repo contains scripts to benchmark the training speed of OpenSeq2Seq on different machines.

To run the benchmark, use this command

./batch_benchmark.sh <model_type> <min_num_gpu> <max_num_gpu> <batch_size_per_gpu>
cat summary-<model_type>.md

The script will spin multiple runs with number of gpus increased from min_num_gpu to max_num_gpu. Results will be saved as summary-<model_type>.md file.

model_type: Check folder example_configs/text2text/en-de. For example, use en-de-nmt-small-fp32 and en-de-nmt-small-fp16 to train a small nmt with single or mixed precision.
min_num_gpu: default to 1.
max_num_gpu: default to 8.
batch_size_per_gpu: default to 128

Name		Name	Last commit message	Last commit date
Latest commit History 1,387 Commits
ctc_decoder_with_lm		ctc_decoder_with_lm
docker		docker
docs		docs
example_configs		example_configs
open_seq2seq		open_seq2seq
scripts		scripts
.gitignore		.gitignore
.pylintrc		.pylintrc
.style.yapf		.style.yapf
AUTHORS		AUTHORS
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
Dockerfile_public		Dockerfile_public
Interactive_Infer_example.ipynb		Interactive_Infer_example.ipynb
LICENSE		LICENSE
README.md		README.md
batch_benchmark.sh		batch_benchmark.sh
clean-corpus-n.perl		clean-corpus-n.perl
clean-corpus-n.perl.1		clean-corpus-n.perl.1
clean-corpus-n.perl.2		clean-corpus-n.perl.2
requirements.txt		requirements.txt
run.py		run.py
tokenizer_wrapper.py		tokenizer_wrapper.py

License

chuanli11/benchmark_openseq2seq

Folders and files

Latest commit

History

Repository files navigation

Benchmark OpenNMT

About

Resources

License

Stars

Watchers

Forks

Languages