MetaQuant

Codes for Accepted Paper : "MetaQuant: Learning to Quantize by Learning to Penetrate Non-differentiable Quantization" in NeurIPS 2019.

Camera-ready is now released here!

Poster is released here!

Slides is released here!

About MetaQuant

check meta-quantize-tutorial.ipynb for description.

How to use it

Prepare pre-trained model

Please check here.

The following command train a ResNet20 using CIFAR10:

python train_base_model.py -m ResNet20 -d CIFAR10

Or users can use the default pretrained model provided by us. Uploaded in Results/model-dataset/model-dataset-pretrain.pth

Run MetaQuant

The following commands run MetaQuant on ResNet20 using CIFAR10 dataset with dorefa as forward quantization method and SGD as optimization.

The resulting quantized model is quantized using 1 bits: {+1, -1} for all layers (conv, fc).

Initial learning rate is set as 1e-3 and decreases by a factor of 0.1 every 30 epochs: 1e-3->1e-4->1e-5:

CUDA_VISIBLE_DEVICES='0' python meta-quantize.py -m ResNet20 -d CIFAR10 -q dorefa -bw 1 -o SGD -meta MultiFC -hidden 100 -ad 30

Experiments

Model	Dataset	Forward	Backward	Optimizer	Best Acc	Last 10 Acc	FP Acc
ResNet20	CIFAR10	dorefa	STE	SGD	82.704(0.044)	80.745(2.113)	91.5
			MultiFC		89.666(0.06)	88.942(0.466)
			FC-Grad		89.618(0.007)	88.840(0.291)
			LSTMFC		89.278(0.09)	88.305(0.81)
			STE	Adam	90.222(0.014)	89.782(0.172)
			MultiFC		90.308(0.012)	89.941(0.068)
			FC-Grad		90.407(0.02)	89.979(0.103)
			LSTMFC		90.310(0.011)	89.962(0.068)
		BWN	STE	SGD	78.550(0.035)	75.913(3.495)
			FC-Grad		89.500(0.008)	88.949(0.231)
			LSTMFC		89.890(0.018)	89.289(0.212)
			STE	Adam	90.470(0.004)	89.896(0.182)
			FC-Grad		90.426(0.042)	90.036(0.109)
			LSTMFC		90.494(0.030)	90.042(0.098)

Best accuracy refers to the best test accuracy recorded during training, mean and variance (in parentheses) for 5 times experiments are reported.

Last 10 acc refers to the test accuracy in the last 10 epochs for 5 times experiments

Raw experimental data can be found in Results/model-dataset/runs-Quant-Multiple

Comparison using dorefa as forward and SGD as optimizer

Comparison using dorefa as forward and Adam as optimizer

Comparison using BWN as forward and SGD as optimizer

Comparison using BWN as forward and Adam as optimizer

Shadow around line is the max/min value during the multiple trial experiments.

Required

pytorch > 0.4

Customization

For using MetaQuant in your own model, the following steps are required:

Replace all quantized layer with meta_utils.meta_quantized_module.MetaConv and meta_utils.meta_quantized_module.Linear
Construct layer_name_list as in models.quantized_meta_resnet.py, which is used to access layer in MetaQuant.
Write your code as in meta-quantize.py

Support

Leave an issue if there is any bug and email me if any concerns about paper.

Citation

Cite the paper if anything helps you:

@article{chen2019metaquant,
  title={MetaQuant: Learning to Quantize by Learning to Penetrate Non-differentiable Quantization},
  author={Chen, Shangyu and Wang, Wenya and Pan, Sinno Jialin},
  journal={Conference on Neural Information Processing Systems},
  year={2019}
}

jawae / metaquant Goto Github PK

metaquant's Introduction