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}
}
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 models achieve up to 3× speedup and 3.7× smaller model size with no performance loss.
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 .
Task | task_name | Train | Valid | Test |
---|---|---|---|---|
WMT'14 En-De | wmt14.en-de | WMT'16 | newstest2013 | newstest2014 |
WMT'14 En-Fr | wmt14.en-fr | WMT'14 | newstest2012&2013 | newstest2014 |
WMT'19 En-De | wmt19.en-de | WMT'19 | newstest2017 | newstest2018 |
IWSLT'14 De-En | iwslt14.de-en | IWSLT'14 train set | IWSLT'14 valid set | IWSLT14.TED.dev2010 IWSLT14.TEDX.dev2012 IWSLT14.TED.tst2010 IWSLT14.TED.tst2011 IWSLT14.TED.tst2012 |
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
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.
Task | Hardware | Latency | #Params (M) |
FLOPs (G) |
BLEU | Sacre BLEU |
model_name | Link |
---|---|---|---|---|---|---|---|---|
WMT'14 En-De | Raspberry Pi ARM Cortex-A72 CPU | 3.5s 4.0s 4.5s 5.0s 6.0s 6.9s |
25.22 29.42 35.72 36.77 44.13 48.33 |
1.53 1.78 2.19 2.26 2.70 3.02 |
25.8 26.9 27.6 27.8 28.2 28.4 |
25.6 26.6 27.1 27.2 27.6 27.8 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link link |
WMT'14 En-De | Intel Xeon E5-2640 CPU | 137.9ms 204.2ms 278.7ms 340.2ms 369.6ms 450.9ms |
30.47 35.72 40.97 46.23 51.48 56.73 |
1.87 2.19 2.54 2.86 3.21 3.53 |
25.8 27.6 27.9 28.1 28.2 28.5 |
25.6 27.1 27.3 27.5 27.6 27.9 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link link |
WMT'14 En-De | Nvidia TITAN Xp GPU | 57.1ms 91.2ms 126.0ms 146.7ms 208.1ms |
30.47 35.72 40.97 51.20 49.38 |
1.87 2.19 2.54 3.17 3.09 |
25.8 27.6 27.9 28.1 28.5 |
25.6 27.1 27.3 27.5 27.8 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link |
WMT'14 En-Fr | Raspberry Pi ARM Cortex-A72 CPU | 4.3s 5.3s 5.8s 6.9s 7.8s 9.1s |
25.22 35.72 36.77 44.13 49.38 56.73 |
1.53 2.23 2.26 2.70 3.09 3.57 |
38.8 40.1 40.6 41.1 41.4 41.8 |
36.0 37.3 37.8 38.3 38.5 38.9 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link link |
WMT'14 En-Fr | Intel Xeon E5-2640 CPU | 154.7ms 208.8ms 329.4ms 394.5ms 442.0ms |
30.47 35.72 44.13 51.48 56.73 |
1.84 2.23 2.70 3.28 3.57 |
39.1 40.0 41.1 41.4 41.7 |
36.3 37.2 38.2 38.5 38.8 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link |
WMT'14 En-Fr | Nvidia TITAN Xp GPU | 69.3ms 94.9ms 132.9ms 168.3ms 208.3ms |
30.47 35.72 40.97 46.23 51.48 |
1.84 2.23 2.51 2.90 3.25 |
39.1 40.0 40.7 41.1 41.7 |
36.3 37.2 37.8 38.3 38.8 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link |
WMT'19 En-De | Nvidia TITAN Xp GPU | 55.7ms 93.2ms 134.5ms 176.1ms 204.5ms 237.8ms |
36.89 42.28 40.97 46.23 51.48 56.73 |
2.27 2.63 2.54 2.86 3.18 3.53 |
42.4 44.4 45.4 46.2 46.5 46.7 |
41.9 43.9 44.7 45.6 45.7 46.0 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link link |
IWSLT'14 De-En | Nvidia TITAN Xp GPU | 45.6ms 74.5ms 109.0ms 137.8ms 168.8ms |
16.82 19.98 23.13 27.33 31.54 |
0.78 0.93 1.13 1.32 1.52 |
33.4 34.2 34.5 34.7 34.8 |
32.5 33.3 33.6 33.8 33.9 |
[email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] [email protected][email protected] |
link link link link link |
python download_model.py --model-name=[model_name]
# for example
python download_model.py [email protected][email protected]
# to download all models
python download_model.py --download-all
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/[email protected][email protected] \
configs/wmt14.en-de/subtransformer/[email protected][email protected] \
normal
# another example
bash configs/iwslt14.de-en/test.sh \
./downloaded_models/[email protected][email protected] \
configs/iwslt14.de-en/subtransformer/[email protected][email protected] \
sacre
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/[email protected][email protected] \
--sub-configs=configs/wmt14.en-de/subtransformer/common.yml --latcpu
# another example
python train.py \
--configs=configs/iwslt14.de-en/subtransformer/[email protected][email protected] \
--sub-configs=configs/iwslt14.de-en/subtransformer/common.yml --profile-flops
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]
.
Task | search_space | model_name | Link |
---|---|---|---|
WMT'14 En-De | space0 | HAT_wmt14ende_super_space0 | link |
WMT'14 En-Fr | space0 | HAT_wmt14enfr_super_space0 | link |
WMT'19 En-De | space0 | HAT_wmt19ende_super_space0 | link |
IWSLT'14 De-En | space1 | HAT_iwslt14deen_super_space1 | link |
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.
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.
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.
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.
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/[email protected] --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.
- Python >= 3.6
- PyTorch >= 1.0.0
- configargparse >= 0.14
- New model training requires NVIDIA GPUs and NCCL
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
If you have any questions, feel free to contact Hanrui Wang through Email ([email protected]) or Github issues. Pull requests are highly welcomed!
This repository is released under the MIT license. See LICENSE for more information.
We are thankful to fairseq as the backbone of this repo.