Fast_Yolov7_Pytorch🎉️🎉️🎉️

Guide

• Installation
• Quick Start
• References
• Contact Me

Installation：

git clone https://github.com/ChristianYang37/fast_yolov7_pytorch.git

pip install -r requirements.txt

If you can't install torch_pruning, please do as follow

git clone https://github.com/VainF/Torch-Pruning.git
cd Torch-Pruning
python setup.py install

For pytorch yolov7 state dicts, click here to download.

Quick Start：

Pruning & Training

python train.py --workers 8 --device 0 --batch-size 32 --data data/custom.yaml --img 640 640 --cfg cfg/training/yolov7-custom.yaml --weights 'yolov7_training.pt' --name yolov7-custom --hyp data/hyp.scratch.custom.yaml --sparsity 0.3 --num_epoch_to_prune 4 --prune_nore L2

if you want prune model without training, you can just set epochs = 0

Opts

sparsity: the sparsity of pruning

num_epoch_to_prune: prune model after num_epoch_to_prune times finetune

prune_norm: L1 or L2

the code actually do prune as follows

    for idx, epoch in enumerate(range(start_epoch, epochs)):
        if (idx + 1) % opt.num_epochs_to_prune:
            yolo_pruner.step(model, device)

So, for more efficient pruning, we suggest you set num_batch_to_prune big enough to make sure the model has fitted the data before you prune it, and also set epochs optimally.

Some Results

Prune yolov7_training.pt On COCO128.yaml (without finetune)

Speed test on GPU=A5000, batch_size=32

Quantization

python train.py --workers 8 --device 0 --batch-size 32 --data data/custom.yaml --img 640 640 --cfg cfg/training/yolov7-custom.yaml --weights 'yolov7_training.pt' --name yolov7-custom --hyp data/hyp.scratch.custom.yaml --method static

Opts

method: algorithm to quantify model, static or dynamic

deploy_device: pytorch now support x86 and arm, is enabled for method == static only

When you set method = dynamic, it require train data to make quantified model fit the distribution.

References:

1. WongKinYiu/yolov7: Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors (github.com)
2. VainF/Torch-Pruning: [CVPR-2023] Towards Any Structural Pruning; LLaMA / CNNs / Transformers (github.com)
3. PyTorch
4. ultralytics/yolov5: YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite (github.com)

Contact：

This repository is for AIRS's project, the author is an undergraduate student at Sun Yat sen University.

Email: [email protected]