Built mmhuman3d from zero, here's problems I have encountered.

[ykk648]

I have no human3d training exp before , most are datasets preprocess problems.

1. Inference demo render fail. solved by #136
2. MPII official dataset missing train.h5, see #135
3. MPI-INF-3DHP & Human3.6m needs extract frame first, I add extract_img=True in convert_datasets.py, it costs about 16 & 6 hours with low cpu consumation on one hard disk, no one had told me that :(
4. According human36m-mosh, the file structure is /datasets/h36m_mosh, but the h36m_spin convertor acturally needs /datasets/h36m_extras/mosh_data
5. Where is h36m datasets itself? I even have thinked about h36m datasets came from Adobe :(. I found h36m here and moshed date from HybrIK, kindness Jeff btw.
6. Why remove frame_ in MPI-INF-3DHP image path ref here? It raise FileNotFoundError: [Errno 2] No such file or directory: 'data/datasets/mpi_inf_3dhp/S5/Seq2/imageFrames/video_7/frame_000368.jpg' when you start your first SPIN training.

Finally, in case #53 occur, run python -m torch.distributed.launch --nproc_per_node=4 --master_port=29500 tools/train.py ./configs/spin/resnet50_spin_pw3d.py --no-validate --work-dir ./work_dir --gpus 4 --launcher pytorch`

2022-03-30 07:17:02,817 - mmhuman3d - INFO - Epoch [1][50/782] lr_backbone: 3.000e-05 lr_head: 3.000e-05, eta: 5:46:10, time: 2.673, data_time: 0.355, memory: 3296, keypoints3d_loss: 0.9700, keypoints2d_loss: 0.1061, vertex_loss: 0.1511, smpl_pose_loss: 0.0978, smpl_betas_loss: 0.0287, camera_loss: 0.0000, loss: 1.3538

The world is bright again.

[caizhongang]

Hi @ykk648 , thank you very much for your feedback.

2. MPII official dataset missing train.h5, see #135 We are looking into this, will update here if we have any progress.
3. MPI-INF-3DHP & Human3.6m needs extract frame first, I add extract_img=True in convert_datasets.py, it costs about 16 & 6 hours with low cpu consumation on one hard disk, no one had told me that :(
4. According human36m-mosh, the file structure is /datasets/h36m_mosh, but the h36m_spin convertor acturally needs /datasets/h36m_extras/mosh_data
5. Why remove frame_ in MPI-INF-3DHP image path ref here? It raise FileNotFoundError: [Errno 2] No such file or directory: 'data/datasets/mpi_inf_3dhp/S5/Seq2/imageFrames/video_7/frame_000368.jpg' when you start your first SPIN training.

These problems will be fixed in #143 .

5. Where is h36m datasets itself? I even have thinked about h36m datasets came from Adobe :(. I found h36m here and moshed date from HybrIK, kindness Jeff btw.

To download H36M, you need to make an account at the official website first and wait for approval. For H36M mosh, the original license does not allow us to redistribute it, but it is indeed available online.

[ykk648]

thx, issue close