MLE Quantization of CNNs

We perform non-uniform quantization according to analytical analysis assuming a Laplace distribution. Neural networks quantized by our method show better or comparable performance to other SOTA 4-bit post training quantization methods. A more detailed explanation can be found in mle_quantization.pdf

Requirements

1. python3 (>=3.7)
2. pytorch 1.6
3. torchvision 0.7
4. tqdm 4.49
5. matplotlib 3.2
6. numpy 1.19
7. scipy 1.5

Data

• To run this code you need the validation set from ILSVRC2012 data
• To get the ILSVRC2012 data, you should register on their site for access: http://www.image-net.org/
• Place the images in the following path: ./imagenet/ILSVRC2012/validation_data/ILSVRC2012_imgs/
• Place the labels text files (those in the data_labels folder) in the following path: ./imagenet/ILSVRC2012/validation_data/ILSVRC2012_labels/
• The torchvision pretrained models have different id classes than the ground truth labels of ILSVRC2012. Check pytorch/vision#484 for more information.

Quantization Simulation

An experiment can be carried out by running the following command:

python test_model.py --model --quantize --w_bits --a_bits

• model expects VGG16, VGG16_BN, ResNet18, ResNet50, ResNet101 or Inception_V3.
• quantize is a boolean flag to quantize the model.
• w_bits and a_bits are the number of bits allocated for weights and activations respectively (if quantized).

Results

The following table shows the accuracies of all the different deep learning models we tested:

Running MLE analysis

Run the following command:

python mle_analysis.py

A similar graph to the one below should be generated. These graphs are created with the hyper parameters suggested in mle_quantization.pdf