wgao9 / weight-quant

Weight Quantization

This is the source code for the paper [Information-Theoretic Understanding of Population Risk Improvement with Model Compression] (https://arxiv.org/abs/1901.09421)

To use this code, it is required to install Python3, with latest NumPy, PyTorch, TorchVision.

Two steps for running experiments:

Step 1: Retraining the model by subsampling

python3 retrain.py --type=mnist --subsample_rate=0.1

Then the retrained model will be stored in sub_models/mnist/ssr=100/model

Note that if you set --compute_gradient=True or --compute_hessian=True, the gradient and hessian of the weights are also computed and stored in sub_models/mnist/ssr=100/importances

Step 2: load pretrained model and run

python3 deep_compress.py --type=mnist --mode=normal --fix_bit=2 --subsample_rate=0.1

and the results are

Before quantization, type=mnist, training acc1=0.9956+-0.0023, acc5=1.0000+-0.0001, loss=0.015394+-0.007845
Before quantization, type=mnist, validation acc1=0.9468+-0.0046, acc5=0.9959+-0.0007, loss=0.320736+-0.029296
Compression ratio = 0.0585+-0.0004
After quantization, type=mnist, training acc1=0.9839+-0.0068, acc5=0.9999+-0.0001, loss=0.060263+-0.015268
After quantization, type=mnist, validation acc1=0.9286+-0.0075, acc5=0.9941+-0.0009, loss=0.259629+-0.018100

Important arguments

1. --type denotes the dataset and model architecture. Supported: mnist, cifar10, cifar100

2. --mode denotes the weight quantization algorithms. Supported:

   • Typical K-means: Set --mode=normal
   • Hessian-weighted K-means: Set --compute_hessian=True for retraining and --mode=hessian for quantization
   • Diameter-regularized Hessian-weighted (DRHW) K-means: Set --compute_hessian=True for retraining and --mode=hessian and --diameter_reg>0 for quantization

3. --subsample_rate denotes the sub sample rate for retraining data. We use 0.1 for MNIST and 0.2 for CIFAT

4. --fix_bits denotes the bits for model compression (same for all layer). If you want different ratio for different layer, use --bits=[3,4,5] for example

5. Please see the helps in the code for other arguments.

Contact [email protected]