This code is intended to be used as a supplemental material for submission to NeurIPS 2021.

DO NOT DISTRIBUTE

This code is tested on Ubuntu 20.04 with Python 3 and CUDA 10.1. Other cuda versions can be used by modifying the cupy version in requirements.txt, provided that CuDNN is installed.

# Set up environment
python3 -m venv
source venv/bin/activate
pip3 install -r requirements.txt

Configurations are provided for CIFAR10/ResNet50 in the acgc/configs folder.

source venv/bin/activate
cd acgc
./configs/rn50_baseline.sh

To replicate GridQuantZ results from the paper, you additionally need to:

• Run quantz with bitwidths of 2, 4, 6, 8, 10, 12, 14, and 16 bits, and run each 5 times
• Select the result with the lowest bitwidth and average accuracy no less than the baseline - 0.1%

Evaluation with the CIFAR10 test dataset is run during training. The 'validation/main/accuracy' entry in the report.txt or log contains test accuracy throughout training.

You can download pre-trained snapshots for each config from acgc/configs.

• rn50_baseline
• rn50_quant_8bit
• rn50_quantz_8bit
• rn50_autoquant
• rn50_autoquantz

These snapshots can be run using

python3 train_cifar_act_error.py ... --resume <snapshot_file>

We have added reports and logs for each configuration under acgc/results. The logs are associated with each snapshot, above.

A summarized output from these runs is:

Argument parsing and model initialization are handled in acgc/cifar.py and acgc/train_cifar_act_error.py

Modifications to the training loop are in acgc/common/compression/compressed_momentum_sgd.py.

The baseline fixpoint implementation is in acgc/common/compression/quant.py.

The AutoQuant implementation, and error bound calculation are in acgc/common/compression/autoquant.py.

Gradient and parameter estimation are performed in acgc/common/compression/grad_approx.py