jack16888 / yolov4_prune

20210106 updates:

add block prune and soft prune in channels at the same time, in order to get the maximum compression rate, but it not work.

In my experiment, every csp blcok is very important and cannot be deleted. Every time soft-prune is used, the precision drops a lot, and the final precision loss is very serious.

    paper reference: Soft Filter Pruning for Accelerating Deep Convolutional Neural Networks

    paper reference: Data-Driven Sparse Structure Selection for Deep Neural Networks

all this paper is auto prune, maybe i can use iterative prune to use this way,

*********Recently, I found a bug that if you use sliming method to prune, once you quantify the model, it is very likely that there will be a big loss of accuracy

*********Note that this is related to the pruning principle of bn layer. It is best not to use this method when quantization is needed. If it is used, it is best to truncate the parameters of bn

20201128 updates:

I want to duplicate the yolo algorithm based on my own ideas, and structed prune yolo use pytorch, if you agree with my work, thank u to give me a star to support

20201209 updates:

add prune code ,finetune code,  knowledge distill code.240Mb

paper: Learning Efficient Convolutional Networks through Network Slimming

reference code: https://github.com/tanluren/yolov3-channel-and-layer-pruning  (in this project, yolov4 voc dataset map only 0.83, but my project best map is 0.87 )

how to prune: see ths last in readme.md

next stage: i will use dynamic prune yolov4 and commit code

voc datasets:

                    map                        size

        yolov4      0.87(after 3 epoch)        240Mb

        yolov3      0.83(after 10 epoch)       240Mb

  yolov4_prune      0.832                       75Mb

    u can easily obtain this performance

how to train your own dataset:

1. create dataset, make your dataset become txt file- just like this:

   img_id img_dir width height label xmin ymin xmax ymax

   1 dir/img1.jpg width height label1 xmin ymin xmax ymax label2 xmin ymin xmax ymax .......

   2 dir/img2.jpg width height label1 xmin ymin xmax ymax

2.create label.names just like data/voc.names in my code

3.train command:

    python train.py --cfg cfg/yolov4.cfg --weights weights/yolov4.weights --device 0,1,2,3 --train_path data/train.txt --val_path data/val.txt --names_classes data/coco.names

yolov4_prune

how to prune:

1. use my code train a best ap model

2. set best model as pretrained_weights in train_prune.py,                                then:   python train_prune.py    -> get a sparse model

3. after second stage, u will get a sparse model, set this model in slim_prune.py         then:   python slim_prune.py     -> get a structed prune model

4. u will obtain new yolov4.cfg and yolov4.weights, finetune model                        then:   python train.py