HAT: Hardware Aware Transformers for Efficient Natural Language Processing [paper] [website] [video]

@inproceedings{hanruiwang2020hat,
    title     = {HAT: Hardware-Aware Transformers for Efficient Natural Language Processing},
    author    = {Wang, Hanrui and Wu, Zhanghao and Liu, Zhijian and Cai, Han and Zhu, Ligeng and Gan, Chuang and Han, Song},
    booktitle = {Annual Conference of the Association for Computational Linguistics},
    year      = {2020}
} 

News

• HAT is covered by VentureBeat.
• HAT is covered by MIT News.

Overview

We release the PyTorch code and 50 pre-trained models for HAT: Hardware-Aware Transformers. Within a Transformer supernet (SuperTransformer), we efficiently search for a specialized fast model (SubTransformer) for each hardware with latency feedback. The search cost is reduced by over 10000×.

HAT Framework overview:

HAT models achieve up to 3× speedup and 3.7× smaller model size with no performance loss.

Usage

Installation

To install from source and develop locally:

git clone https://github.com/mit-han-lab/hardware-aware-transformers.git
cd hardware-aware-transformers
pip install --editable .

Data Preparation

To download and preprocess data, run:

bash configs/[task_name]/preprocess.sh

If you find preprocessing time-consuming, you can directly download the preprocessed data we provide:

bash configs/[task_name]/get_preprocessed.sh

Testing

We provide pre-trained models (SubTransformers) on the Machine Translation tasks for evaluations. The #Params and FLOPs do not count in the embedding lookup table and the last output layers because they are dependent on tasks.

Download models:

python download_model.py --model-name=[model_name]
# for example
python download_model.py --model-name=HAT_wmt14ende_raspberrypi@3.5s_bleu@25.8
# to download all models
python download_model.py --download-all

Test BLEU (SacreBLEU) score:

bash configs/[task_name]/test.sh \
    [model_file] \
    configs/[task_name]/subtransformer/[model_name].yml \
    [normal|sacre]
# for example
bash configs/wmt14.en-de/test.sh \
    ./downloaded_models/HAT_wmt14ende_raspberrypi@3.5s_bleu@25.8.pt \
    configs/wmt14.en-de/subtransformer/HAT_wmt14ende_raspberrypi@3.5s_bleu@25.8.yml \
    normal
# another example
bash configs/iwslt14.de-en/test.sh \
    ./downloaded_models/HAT_iwslt14deen_titanxp@137.8ms_bleu@34.7.pt \
    configs/iwslt14.de-en/subtransformer/HAT_iwslt14deen_titanxp@137.8ms_bleu@34.7.yml \
    sacre

Test Latency, model size and FLOPs

To profile the latency, model size and FLOPs (FLOPs profiling needs torchprofile), you can run the commands below. By default, only the model size is profiled:

python train.py \
    --configs=configs/[task_name]/subtransformer/[model_name].yml \
    --sub-configs=configs/[task_name]/subtransformer/common.yml \
    [--latgpu|--latcpu|--profile-flops]
# for example
python train.py \
    --configs=configs/wmt14.en-de/subtransformer/HAT_wmt14ende_raspberrypi@3.5s_bleu@25.8.yml \
    --sub-configs=configs/wmt14.en-de/subtransformer/common.yml --latcpu
# another example
python train.py \
    --configs=configs/iwslt14.de-en/subtransformer/HAT_iwslt14deen_titanxp@137.8ms_bleu@34.7.yml \
    --sub-configs=configs/iwslt14.de-en/subtransformer/common.yml --profile-flops

Training

1. Train a SuperTransformer

The SuperTransformer is a supernet that contains many SubTransformers with weight-sharing. By default, we train WMT tasks on 8 GPUs. Please adjust --update-freq according to GPU numbers (128/x for x GPUs). Note that for IWSLT, we only train on one GPU with --update-freq=1.

python train.py --configs=configs/[task_name]/supertransformer/[search_space].yml
# for example
python train.py --configs=configs/wmt14.en-de/supertransformer/space0.yml
# another example
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --configs=configs/wmt14.en-fr/supertransformer/space0.yml --update-freq=32

In the --configs file, SuperTransformer model architecture, SubTransformer search space and training settings are specified.

We also provide pre-trained SuperTransformers for the four tasks as below. To download, run python download_model.py --model-name=[model_name].

2. Evolutionary Search

The second step of HAT is to perform an evolutionary search in the trained SuperTransformer with a hardware latency constraint in the loop. We train a latency predictor to get fast and accurate latency feedback.

2.1 Generate a latency dataset

python latency_dataset.py --configs=configs/[task_name]/latency_dataset/[hardware_name].yml
# for example
python latency_dataset.py --configs=configs/wmt14.en-de/latency_dataset/cpu_raspberrypi.yml

hardware_name can be cpu_raspberrypi, cpu_xeon and gpu_titanxp. The --configs file contains the design space in which we sample models to get (model_architecture, real_latency) data pairs.

We provide the datasets we collect in the latency_dataset folder.

2.2 Train a latency predictor

Then train a predictor with collected dataset:

python latency_predictor.py --configs=configs/[task_name]/latency_predictor/[hardware_name].yml
# for example
python latency_predictor.py --configs=configs/wmt14.en-de/latency_predictor/cpu_raspberrypi.yml

The --configs file contains the predictor's model architecture and training settings. We provide pre-trained predictors in latency_dataset/predictors folder.

2.3 Run evolutionary search with a latency constraint

python evo_search.py --configs=[supertransformer_config_file].yml --evo-configs=[evo_settings].yml
# for example
python evo_search.py --configs=configs/wmt14.en-de/supertransformer/space0.yml --evo-configs=configs/wmt14.en-de/evo_search/wmt14ende_titanxp.yml

The --configs file points to the SuperTransformer training config file. --evo-configs file includes evolutionary search settings, and also specifies the desired latency constraint latency-constraint. Note that the feature-norm and lat-norm here should be the same as those when training the latency predictor. --write-config-path specifies the location to write out the searched SubTransformer architecture.

3. Train a Searched SubTransformer

Finally, we train the search SubTransformer from scratch:

python train.py --configs=[subtransformer_architecture].yml --sub-configs=configs/[task_name]/subtransformer/common.yml
# for example
python train.py --configs=configs/wmt14.en-de/subtransformer/wmt14ende_titanxp@200ms.yml --sub-configs=configs/wmt14.en-de/subtransformer/common.yml

--configs points to the --write-config-path in step 2.3. --sub-configs contains training settings for the SubTransformer.

After training a SubTransformer, you can test its performance with the methods in Testing section.

Dependencies

• Python >= 3.6
• PyTorch >= 1.0.0
• configargparse >= 0.14
• New model training requires NVIDIA GPUs and NCCL

Related works on efficient deep learning

MicroNet for Efficient Language Modeling

Lite Transformer with Long-Short Range Attention

AMC: AutoML for Model Compression and Acceleration on Mobile Devices

Once-for-All: Train One Network and Specialize it for Efficient Deployment

ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

Contact

If you have any questions, feel free to contact Hanrui Wang through Email (hanrui@mit.edu) or Github issues. Pull requests are highly welcomed!

Licence

This repository is released under the MIT license. See LICENSE for more information.

Acknowledgements

We are thankful to fairseq as the backbone of this repo.