Open wer751 opened 3 months ago
@jameslahm Thanks. I used 1*3090 to train YOLOv10n (cache ram), each epoch costs 2min 54s, while YOLOv8n (cache ram) costs 30s. The cache can't speed up the training. I called the cache and didn't get accelerated training like v8. Then I used your ultralytics to train v8 (cache), and the speed slowed down, compared with v8 (official), the speed was five times slower. To sum up, although cache code caches data, it cannot fully use the performance of the GPU.
too
Does the codebase share the same python environment with the v8 (official)?
@jameslahm Thanks. I used 1*3090 to train YOLOv10n (cache ram), each epoch costs 2min 54s, while YOLOv8n (cache ram) costs 30s. The cache can't speed up the training. I called the cache and didn't get accelerated training like v8. Then I used your ultralytics to train v8 (cache), and the speed slowed down, compared with v8 (official), the speed was five times slower. To sum up, although cache code caches data, it cannot fully use the performance of the GPU.
Yeah. @jameslahm
Thanks. Could you please provide the details of how you launch the training in this codebase and the v8 (official)?
yolo detect train data=data.yaml model=yolov8n.yaml epochs=2 batch=32 cache=ram In this codebase, I didn't pip install ultralytics, used pip install e. Then, I used v8 (official) to train my data. In the v8 (official), to train model in this environment (pip install ultralytics).
@wer751 @woshicaiji250 Thanks! Could you please check if it works after applying these changes? Thank you!
Thanks, but it didn't work.
Thanks! Could you please apply the below patch and try again? Thank you! diff.patch
Thanks!The problem has been solved by replacing v8's base.py, and so has your patch. Thank you!
Thanks for your interest! Could you please provide more details about your setup? In our local environment with 8*3090, each epoch of training YOLOv10-N costs ~1min54s, which is similar to that of YOLOv8-N (~1min54s).