Yi Lu
Yi Lu
@jnulzl 精度会受影响么
请问下ignite是不支持dataparallel么，为啥使用了dataparallel以后其实还是只有某一张卡分到显存
@netaz I think after set padding=4 the image become 40 x 40, random crop operator can get more result. If we do not set padding=4, random crop only get origin...
@lsy17096535 输入图像分辨率是多大
@rainofmine finally i get the result, on the hard set of WIDER FACE, this code get about 40(mAP).
how many iterations do you train? Can I have a look on your solver.prototxt?@nyyznyyz1991
@nyyznyyz1991 ok, i see it. Some more question, have you follow the author's instruction one by one? Using the Matlab script to align the face, then train them? I ask...
> ubuntu@ubuntu:~/mmdeploy$ python3 tools/deploy.py configs/mmpose/pose-detection_rtmo_tensorrt-fp16_dynamic-640x640.py /home/ubuntu/mmpose/configs/body_2d_keypoint/rtmo/body7/rtmo-t_8xb32-600e_body7-416x416.py https://download.openmmlab.com/mmpose/v1/projects/rtmo/rtmo-t_8xb32-600e_body7-416x416-f48f75cb_20231219.pth /home/ubuntu/mmpose/tests/data/coco/000000000785.jpg --work-dir onnx_dir --dump-info --show --device cuda:0 > > face the problem > > RuntimeError: Unsupported: ONNX export of transpose for tensor of...
> 可以尝试下 [rtmlib](https://github.com/Tau-J/rtmlib/blob/main/rtmo_demo.py) 能问下你的pytorch是什么版本的么，转rtmo的op_set是多少
@kl456123 the author mentioned it that use ROI-Pooling can bring extra computation. Meanwhile, I think discarding Fully Connection Layer also can accelerate the speed of train and inference.