FAB: FPGA-Accelerated Fully-Pipelined Bottleneck Architecture with Batching for High-Performance MobileNetv2 Inference

This repository contains PyTorch, C reference code, some FPGA constraints for FAB project.

Pretrained & C Reference Data

You can download the pretrained weights and the extracted data from the following URL.

https://drive.google.com/drive/folders/1cTyFCJMDTP-DIKNH75waVvRhpdTzzI-m?usp=sharing

---

Pytorch Evaluation

Modify the dataset root directly in the config

If the train dataset root is not set, the ImageNet task cannot be recognized, resulting in an FC layer size mismatch error.

python main_eval.py --common.config-file {config_url} --model.classification.pretrained ./base_weight/mobilenetv2-1.00.pt

Configuration File

W Quant

./config/classification/imagenet/mobilenetv2_ptq.yaml

W/O Quant

./config/classification/imagenet/mobilenetv2.yaml

Quantization configuration Detail

quant.quant : Determining whether to perform quantization
quant.quant_method : meaningless argument
quant.weight_bit : Weight quantization bit
quant.activation_bit : Activation quantization bit
quant.calibration_a : Layer wise calibration
quant.calibration_w : Layer wise calibration
quant.calibration_c : Channel wise calibration
quant.calib_iter : Calibration iter 

Model Detail

Overall Model file

Build The MobileNetV2 Block

/cvnets/modules/mobilenetv2.py

Quantization Module

Build The Quantization Module

/cvnets/ptq

C Reference

Overall code explaination

main.c :
qact.c :
bottleneck.c :
quantizer.c :
conv.c :
utils.c :
fc.c:

FPGA Constraints

Will be updated