Custom data processing "weights.npy" missing

[Yi-Cheng0101]

Hi, I followed the tutorial for training custom data and faced a problem in this step: python tools/prepare_zjumocap.py --data_root ${data_root} --output_root ${data_root} --smpl_model_path ${smpl_model_path} --smpl_uv_path ${smpl_uv_path} --ranges 0 200 1

The issue is prepare_zjumocap.py doesn't generate "weights.npy", but training code: train_net.py requires "weights.npy". I have checked weight output files that generated from prepare_zjumocap.py, it only generates bweights directory, I thought this directory just contains the weights of each frame.

Could you provide the detailed information about how to get weights.npy from custom data? Thanks!

[gcgeng]

Hi, the smpl-meta folder can be found at the dataset request link at: https://github.com/zju3dv/instant-nvr/blob/master/docs/install.md#zju-mocap-dataset.

Could you download the folder there and let us know if it works? Thanks a lot.

Chen

[Yi-Cheng0101]

Hi, the smpl-meta folder can be found at the dataset request link at: https://github.com/zju3dv/instant-nvr/blob/master/docs/install.md#zju-mocap-dataset.

Could you download the folder there and let us know if it works? Thanks a lot.

Chen

Hi, There is no problem when using the datasets you provided, which has weights.npy in smpl-meta folder. The issue I mentioned is that I used my own video and followed "Custom Dataset" tutorial link and run the 3.2. Convert smpl, it doesn't generate weights.npy. So, I would like to know how to generate weights.npy using our own custom data not use the datasets you provided. Thanks for your reply!

[gcgeng]

Hi, I think the smpl-meta folder should be also used in your custom dataset.

Thanks, Chen

[Yi-Cheng0101]

Hi, So I need to use the same smpl-meta folder from the dataset you provide when I process my custom data, right? Thanks!

[gcgeng]

Yes, please try this way and let us know if it works for you.

Thanks, Chen

Sent from Gmail Mobile

On Tue, Jul 23, 2024 at 10:01 Yi-Cheng Hsiao @.***> wrote:

Hi, so I need to use the same smpl-meta from the dataset you provide in my custom datas right? Thanks!

— Reply to this email directly, view it on GitHub https://github.com/zju3dv/instant-nvr/issues/24#issuecomment-2245770728, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACENDOF7Y4CGXOSRLQ6BBO3ZN2D7PAVCNFSM6AAAAABLGK5QYWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENBVG43TANZSHA . You are receiving this because you commented.Message ID: @.***>

[Yi-Cheng0101]

Hi, Ok.

I will try it and let you know it works or not. Thanks!

And this makes me a little confused because when I used my own video and processed it as my custom dataset, I have to generate my own faces.npy , parents.npy, joints.npy,  weights.npy, bigpose_vertices.npy,  bigpose_bw.npy files? Why do I need to use the same files as original datasets? If I got something wrong, please let me know. Thanks!

[gcgeng]

Basically the smpl-meta folder stores shared information across different datasets, and the files generated for different datasets are for those specific datasets. Please let me know if you have further questions.

Thanks, Chen

Sent from Gmail Mobile

On Tue, Jul 23, 2024 at 10:26 Yi-Cheng Hsiao @.***> wrote:

Hi, Ok.

I will try it and let you know it works or not. Thanks!

And this makes me a little confused because when I used my own video and processed it as my custom dataset, I have to generate my own faces.npy , parents.npy, joints.npy, weights.npy, bigpose_vertices.npy, bigpose_bw.npy files? Why do I need to use the same files as original datasets? If I got something wrong, please let me know. Thanks!

— Reply to this email directly, view it on GitHub https://github.com/zju3dv/instant-nvr/issues/24#issuecomment-2245825421, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACENDOGKBFL67U3UYHRAZW3ZN2G2XAVCNFSM6AAAAABLGK5QYWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENBVHAZDKNBSGE . You are receiving this because you commented.Message ID: @.***>