Feng Wang
@LZHgrla could you post a code snippt to reproduce such an error ? It's helpful for me to check. Thanks!
dataloader使用torch没有问题的，不影响训练。如果你有bug可以report issue，这种承诺会fix。除此之外的更新看需求。
Maybe it's caused by updating of ncnn. If your output is correct, it's ok.
plz check all the link and contribute them in a single PR, thanks.
Could you please share more details with me ? What did you do and how your benchmark came out ?
Apex is not needed anymore. For pycocotools, how about [pycocotools windows](https://pypi.org/project/pycocotools-windows/) ?
Could you plz provide more environment info like your torch version, git commit hash and so on? Are you training COCO dataset?
It's unexpected here. However, I'm not sure what's the true reason of performance drop. Have you tried small input like 320 x 320 or 192x 192 ?
I'm not a matlab user, however, I googled [this doc](https://www.mathworks.com/help/deeplearning/ref/importonnxnetwork.html) for you, may it help.
Plz check your image path.