amitpant7 / Quantizing-Efficientnetv2-using-Vitis-AI-Pytorch

Quantization of efficientnet model that i pruned 81% using Vitis AI and compiling it for deployment on FPAG

• Main only contains the scripts to run quantization on efficientnetv2, check other branches for quantized models
• check the replaced branch for efficientnetV2 with the replaced activation function, hard swish.
• Compiled branch contains code for the compiled version of efficientnet v2 after sturctured pruning, quantization and compilation for DPU 1024, Zynq Mpsoc FPGA

Quantized and Compiled Model For deploying on FPGA

• Branch : Compiled

Creating blocks to Deploy in FPGA for Latency Table Construction

• Branch: qt_for_latency_blocks

Installation

1. Environment requirement - pytorch, opencv, tqdm ... - vai_q_pytorch(Optional, required by quantization) - XIR Python frontend (Optional, required by quantization)

2. Installation with GPU Docker

• Please refer to vitis-ai for how to obtain the GPU docker image.

3. Installation without GPU Docker

• Create virtual envrionment and activate it:

conda activate vitis-ai-pytorch

Preparation

Dataset

Donload CIFAR-10 Dataset, or you can set down-true inside the load_data in efficientnetv3_quant.py

Quantization

python ./code/test/efficientnetv2_quant.py --device "cpu" --quant_mode calib --subset_len 1000

To deploy model and export it in xmodel format:

sudo /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/bin/python ./code/test/efficientnetv2_quant.py --quant_mode test --subset_len 1 --batch_size=1 --deploy 

Similarly for quantization-aware training
Fast fintune models 2000 images, calib 2000 images

export W_QUANT=1

sudo /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/bin/python code/efficientnetv2_QAT.py --device "cpu" --quant_mode calib --fast_finetune --subset_len 50

Test

python code/efficientnetv2_QAT.py --device "cpu" --quant_mode test --fast_finetune

--deploy

python code/efficientnetv2_QAT.py --device "cpu" --quant_mode test --fast_finetune --subset_len 1 --batch_size 1 --deploy