uber-research / permute-quantize-finetune

Using ideas from product quantization for state-of-the-art neural network compression.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Permute, Quantize, and Fine-tune

This repository contains the source code and compressed models for the paper Permute, Quantize, and Fine-tune: Efficient Compression of Neural Networks: https://arxiv.org/abs/2010.15703

Permutation optimization

Our method compresses the weight matrices of the network layers by

  1. Finding permutations of the weights that result in a functionally-equivalent, yet easier-to-compress network,
  2. Compressing the weights using product quantization [1],
  3. Fine-tuning the codebooks via stochastic gradient descent.

We provide code for compressing and evaluating ResNet-18, ResNet-50 and Mask R-CNN.

Contents

Requirements

Our code requires Python 3.6 or later. You also need these additional packages:

Additionally, if you have installed Horovod, you may train ResNet with multiple GPUs, but the code will work with a single GPU even without Horovod.

Data

Our experiments require either ImageNet (for classification) or COCO (for detection/segmentation). You should set up a data directory with the datasets.

<your_data_path>
├── coco
│   ├── annotations   (contains      6 json files)
│   ├── train2017     (contains 118287 images)
│   └── val2017       (contains   5000 images)
└── imagenet
    ├── train         (contains   1000 folders with images)
    └── val           (contains   1000 folders with images)

Then, make sure to update the imagenet_path or coco_path field in the config files to point them to your data.

Training ResNet

Besides making sure your ImageNet path is set up, make sure to also set up your output_path in the config file, or pass them via the command line:

python -m src.train_resnet --config ../config/train_resnet50.yaml

The output_path key inside the config file must specify a directory where all the training output should be saved. This script will create 2 subdirectories, called tensorboard and trained_models, inside of the output_path directory.

Launching a tensorboard with the tensorboard directory will allow you observe the training state and behavior over time.

tensorboard --logdir <your_tensorboard_path> --bind_all --port 6006

The trained_models directory will be populated with checkpoints of the saved model after initialization, and then after every epoch. It will also separately store the "best" of these models (the one that attains the highest validation accuracy).

Training Mask R-CNN

Mask R-CNN (with a ResNet-50 backbone) can be trained by running the command:

python -m src.train_maskrcnn --config ../config/train_maskrcnn.yaml

Once again, you need to specify the output_path and the dataset path in the config file before running this.

Pretrained models

We provide the compressed models we learned from running our code at

../compressed_models

All models provided have been compressed with k = 256 centroids

Model (original top-1) Regime Comp. ratio Model size Top-1 accuracy (%)
ResNet-18 (69.76%) Small blocks
Large blocks
29x
43x
1.54MB
1.03MB
66.74
63.33
ResNet-50 (76.15%) Small blocks
Large blocks
19x
31x
5.09MB
3.19MB
75.04
72.18
ResNet-50 Semi-Supervised (78.72%) Small blocks 19x 5.09MB 77.19

We also provide a compressed Mask R-CNN model that attains the following results compared to the uncompressed model:

Model Size Comp. Ratio Box AP Mask AP
Original Mask R-CNN 169.4 MB - 37.9 34.6
Compressed Mask R-CNN 6.65 MB 25.5x 36.3 33.5

which you may use as given for evaluation.

Evaluating ResNet

To evaluate ResNet architectures run the following command from the project root:

python -m src.evaluate_resnet

This will evaluate a ResNet-18 with small blocks by default. To evaluate a ResNet-18 with large blocks, use

python -m src.evaluate_resnet \
    --model.compression_parameters.large_subvectors True \
    --model.state_dict_compressed ../compressed_models/resnet18_large.pth

For ResNet-50 with small blocks, use

python -m src.evaluate_resnet \
    --model.arch resnet50 \
    --model.compression_parameters.layer_specs.fc.k 1024 \
    --model.state_dict_compressed ../compressed_models/resnet50(_ssl).pth

You may load the resnet50_ssl.pth model, which has been pretrained on an unsupervised dataset as well.

And for ResNet-50 with large blocks, use

python -m src.evaluate_resnet \
    --model.arch resnet50 \
    --model.compression_parameters.pw_subvector_size 8 \
    --model.compression_parameters.large_subvectors True \
    --model.compression_parameters.layer_specs.fc.k 1024 \
    --model.state_dict_compressed ../compressed_models/resnet50_large.pth

Evaluating Mask R-CNN

Simply run the command:

python -m src.evaluate_maskrcnn

to load and evaluate the appropriate model.

Citation

If you use our code, please cite our work:

@article{martinez_2020_pqf,
  title={Permute, Quantize, and Fine-tune: Efficient Compression of Neural Networks},
  author={Martinez, Julieta and Shewakramani, Jashan and Liu, Ting Wei and B{\^a}rsan, Ioan Andrei and Zeng, Wenyuan and Urtasun, Raquel},
  journal={arXiv preprint arXiv:2010.15703},
  year={2020}
}

References

[1] Product quantization for nearest neighbor search
[2] And the bit goes down: Revisiting the quantization of neural networks

About

Using ideas from product quantization for state-of-the-art neural network compression.

License:Other


Languages

Language:Python 100.0%