yolov3-channel-and-layer-pruning
yolov3-channel-and-layer-pruning copied to clipboard
tanluren
fine-tune 時，mAP下降很多
您好，我想請教一下
我使用 prune.py 做剪枝後的 mAP 是 0.72
但是我將他 fine-tune 時，mAP 下降到幾乎為 0 了，請問是哪邊有錯誤嗎? 謝謝。 我執行的指令如下: python train.py --cfg ../prune_0.1_Person_detection/prune_0.1_cfg/prune_0.1_yolov4-tiny-obj.cfg --weights ../prune_0.1_Person_detection/prune_0.1_weights/prune_0.1_yolov4-tiny-obj_final.weights --data ../Person_detection/cfg/person-prune.data --epochs 50 --batch-size 32
以下是訓練時的訊息: learning rate: 0.0011625 /home/joy/.local/lib/python3.6/site-packages/torch/cuda/memory.py:346: FutureWarning: torch.cuda.memory_cached has been renamed to torch.cuda.memory_reserved FutureWarning) 1/49 0.367G 1.38 6.37 0 7.76 0 0 4 160: Class Images Targets P R mAP F1: 100%|█| 69/69 [03:4 all 2.18e+03 3.88e+03 0.0288 0.295 0.0927 0.0525
Epoch gpu_mem GIoU obj cls total soft rratio targets img_size
0%| | 0/274 [00:00<?, ?it/s]learning rate: 0.002324 2/49 0.367G 1.12 5.61 0 6.73 0 0 8 160: Class Images Targets P R mAP F1: 100%|█| 69/69 [01:4 all 2.18e+03 3.88e+03 0.00237 0.00954 2.99e-05 0.00379
Epoch gpu_mem GIoU obj cls total soft rratio targets img_size
0%| | 0/274 [00:00<?, ?it/s]learning rate: 0.0034855 Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 1024.0 3/49 0.367G 1.04 5.2 0 6.24 0 0 4 160: Class Images Targets P R mAP F1: 90%|▉| 62/69 [04:
@chingi071 Can you pls tell the folder structure you have used for sparsity Training ? I mean .data file and the .txt file. Since for me during the sparsity training only 1 Img is getting detected and its corresponding label .
I have added my query in detail in https://github.com/tanluren/yolov3-channel-and-layer-pruning/issues/135
Thanks in advance :)
