Skip to content

Code repositoy for "AOWS: Adaptive and optimal network width search with latency constraints", CVPR 2020

License

Notifications You must be signed in to change notification settings

bermanmaxim/AOWS

Repository files navigation

AOWS

AOWS: Adaptive and optimal network width search with latency constraints, Maxim Berman, Leonid Pishchulin, Ning Xu, Matthew B. Blaschko, Gérard Medioni, NAS workshop @ ICLR 2020 and CVPR 2020 (oral).

AOWS-teaser

Usage

Latency model

main file: latency.py

depends on: PyTorch, CVXPY, matplotlib, numpy, numba, scipy, torch2trt

  • Generate training and validation samples
python latency.py generate --device trt --dtype fp16 \
   --biased --count 8000 output/samples_trt16.jsonl
python latency.py generate --device trt --dtype fp16 \
   --count 200 output/val_trt16.jsonl
  • Fit the model (K controls the amount of regularization and should be set by validation)
python latency.py fit output/samples_trt16.jsonl \
   output/model_trt16_K100.0.jsonl -K 100.0
  • Validate the model (produces a plot)
python latency.py validate output/val_trt16.jsonl \
   output/model_trt16_K100.0.jsonl output/correlation_plot.png

Additionally, one can benchark a single configuration with this script using e.g.

python latency.py benchmark --device trt --dtype fp16 \
   "(16, 32, 64, 112, 360, 48, 464, 664, 152, 664, 256, 208, 816, 304)"

Network width search

main file: train_nas.py

depends on: PyTorch, numpy, numba

  • Train a slimmable network and select a configuration with OWS. See -h for optimization options.
python train_nas.py --data /imagenet --latency-target 0.04 \
   --latency-model output/model_trt16_K100.0.jsonl \
   --expname output/ows-trt16-0.04 --resume-last

In OWS, the latency target --latency-target can be changed during or after training. Using the parameter --resume-last allows to resume the last checkpoint without having to retrain, allowing for varying the latency target.

Implementation detail: for ease of implementation we here use a fixed moving average with a window of --window=100000 samples for each unary weight, while in the article we used the statistics available over one full last epoch.

  • Train a slimmable network with AOWS. See -h for optimization options. The outputs, including best configuration for each epochs, are put in the directory corresponding to the parameter --expname.
python train_nas.py --data /imagenet --latency-target 0.04 \
   --latency-model output/model_trt16_K100.0.jsonl \
   --AOWS --expname output/aows-trt16-0.04 --resume-last 

In AOWS, the latency target --latency-target should be set at the beginning of the training, since it impacts the training.

Training the final model

main file: train_final.py

depends on: mxnet

modified version of gluon-cv's train_imagenet.py for training mobilenet-v1 with varying channel numbers. Refer to gluon-cv's documentation for detailed usage.

Example command:

python train_final.py \
   --rec-train /imagenet/imagenet_train.rec \
   --rec-train-idx /imagenet/imagenet_train.idx \ 
   --rec-val /ramdisk/imagenet_val.rec \
   --rec-val-idx /ramdisk/imagenet_val.idx \
   --use-rec --mode hybrid --lr 0.4 --lr-mode cosine \
   --num-epochs 200 --batch-size 256 -j 32 --num-gpus 4 \
   --dtype float16 --warmup-epochs 5 --no-wd \
   --label-smoothing --mixup \
   --save-dir params_mymobilenet --logging-file mymobilenet.log \
   --configuration "(16, 32, 64, 112, 360, 48, 464, 664, 152, 664, 256, 208, 816, 304)"

Citation

@InProceedings{Berman2020AOWS,
  author    = {Berman, Maxim and Pishchulin, Leonid and Xu, Ning and Blaschko, Matthew B. and Medioni, Gerard},
  title     = {{AOWS}: adaptive and optimal network width search with latency constraints},
  booktitle = {Proceedings of the {IEEE} Computer Society Conference on Computer Vision and Pattern Recognition},
  month     = jun,
  year      = {2020},
}

Disclaimer

The code was re-implemented and is not fully tested at this point.

About

Code repositoy for "AOWS: Adaptive and optimal network width search with latency constraints", CVPR 2020

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages