eDQA

This repo is the offical repo of paper eDQA: Efficient Deep Quantization of DNN Activations on Edge Devices.

Introduction

This repo is composed by two parts: algorithmic part (ml_exps) and hardware part (hw_exps). The hardware part corresponds to the experiments in the section 4.4 of the paper and the algorithmic part corresponds to the rest experiments.

Algorithmic Part

In this part, all the algorithms are embedded in the model files (ml_exps/experiment_files/models_with_algorithm).

To run the main experiments in section 4.2, refer to the scripts:
ml_exps/experiment_files/run_tmlr.sh,
ml_exps/experiment_files/run_ext.sh,
ml_exps/experiment_files/run_sript.sh,
ml_exps/experiment_files/run_sript_six.sh
ml_exps/experiment_files/run_easy_quant.sh.

For the NoisyQuant, refer to scripts: ml_exps/experiment_files/noisy_quant/experiment_files/run_tmlr.sh. (It contains part of the cases in our paper, but can be an example for how to implement and run the NoisyQuant in our research.)

To run the ablation study in section 4.3, refer to the script: ml_exps/experiment_files/run_ablation_ext.sh.

To run the experiments in section 3.2.2, refer to the script:
ml_exps/experiment_files/run_ablation_compresson.sh,
ml_exps/experiment_files/run_ablation_compression_ext.sh.

Get the channel ranks

Note that for DQA, we need to get the ranks of activation channels first. To get the channels, refer to the scripts:
ml_exps/path/experiment_files/run_ext.sh,
ml_exps/path/experiment_files/run_im.sh,
ml_exps/path/experiment_files/run_sript.sh
ml_exps/path/experiment_files/run_sript_six.sh.

Namings

Note that 'im' means the ImageNet and 'pot' means power-of-two quantization.

Hardware Part

The hardware part (hw_exps/) covers the FPGA experiments in Section 4.4 of the paper. It holds two accelerators, written in SystemC and synthesised with HLS using the SECDA methodology, for the PYNQ-Z1 board:

• v1 (eDQA): hw_exps/dqa_exp/v1/accelerator quantizes an activation channel and compresses its low-order bits with an on-chip canonical Huffman encoder, then decodes and dequantizes it.
• v2 (NoisyQuant baseline): hw_exps/dqa_exp/v2/accelerator searches for the activation scale and noise bias, then quantizes and dequantizes the channel.

Each accelerator comes with a host driver (accelerator/driver/) and an experiment program (experiment.cc). The program runs channel sizes from 7×7 to 56×56 and records quantization and dequantization cycles, data transfer times and the compression ratio.

Other contents:

• hw_exps/hw_files/: prebuilt bitstreams (.bit + .hwh) for both designs, their hardware generation configs and the Vivado block-design template.
• hw_exps/data/: input activation data and the golden outputs used to check the results (dqa_exp/gen_data.ipynb generates the inputs).

Note that the experiments are built inside a SECDA framework workspace (Bazel), not as a standalone project.