tusharsangam
commented
2 weeks ago Make sure NPS data is preprocessed using the NPS reproduce branch (this is a condensed two-step of the original multi-step preprocessing routine.) (step 1. videos_to_frames_nps.py, step2. nps_to_yolo.py, optional step3. plot videos) yolo5l.pt is publicly available yolo5l model checkpoint. Following code also commented in submit-train.slurm file
python -m torch.distributed.launch --nproc_per_node=2 --master_port 47769 \
--use_env train.py --img 1280 --adam --batch 4 --epochs 80 --data ./data/NPS.yaml \
--hy ./data/hyps/hyp.VisDrone.yaml --cfg ./models/yolov5l.yaml \
--name image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0 --project ./runs/train/NPS \
--weights ./pretrained/yolo5l.pt --exist-ok
Hello, thank you very much for your contribution. I have a question for you.
When I use these models (such as yolov5s, yolov5s, yolov5l, yolov5l xs tph) to train NPS, my accuracy, recall, and map do not increase but decrease, as shown in the following figure.
Can you give me some advice on how to solve this problem?
My data distribution is as follows, labels from left to right are category, center point coordinates x, y, box width, and height in order.
