porphy/dcci

Content-Aware Adaptive Device-Cloud Collaborative Inference for Object Detection

DCCI

Content-Aware Adaptive Device-Cloud Collaborative Inference for Object Detection

Dependencies

There are some dependencies for running this

1. python3
2. pytorch
3. torchvision
4. zeromq
5. fvcore

How to compress large models

1. Load initialization weights for basic training

python train.py --cfg cfg/yolov4.cfg --data data/voc.data --weights weights/yolov4.weights --epochs 100 --batch-size 16

2. Based on the optimal weights trained in step 1, perform sparse training

python train.py --cfg cfg/yolov4.cfg --data data/voc.data --weights weights/best.pt --epochs 299 --batch-size 64 -sr --s 0.0001 --prune 1

3. Channel-level pruning using sparsely trained models

python channel_pruning.py --cfg cfg/yolov4.cfg --data data/voc.data --weights weights/best.pt --global_percent 0.9 --layer_keep 0.01

4. Layer-level pruning using channel-pruned models

python layer_prune.py --cfg cfg/light_weight_channel.cfg --data data/voc.data --weights weights/light_weight_channel.weight --rl_epochs 80 --batch-size 32

5. Fine-tune recovery accuracy

python train.py --cfg cfg/light_weight_layer.cfg --data data/voc.data --weights weights/light_weight_layer.weights --epochs 100 --batch-size 32

Run end-to-end

We give an example of an end-to-end test under examples. The end_to_end_test file is used to test the end-to-end accuracy. The partition_datasets file is used to partition datasets into hard and simple cases. detect_client is deployed to IoT devices and detect_server is deployed to the cloud.

The weight model can be obtained here

link：https://pan.baidu.com/s/19gVlej3ZkRVxn-n6uVMpPQ

password：lar5