'data/static_fits/h36m_fits.npy' not found

[QuteSaltyFish]

When I run the train.py code in the docker, it said this file not found. How can I get this file?

[ikvision]

Indeed human3.6 has licensing issues #20 (comment)
You may try to train without it #32 (comment)

[GuiZhaoyang]

I have met your problem.I have produced the h36m_train.npz,but I did not find the code to produce h36m_fits.npy. Can you tell me how you solved the problem? @QuteSaltyFish

[ikvision]

h36m_fits.npy contains the best pseudo ground truth for h36m, it is used to populate fits_dict, and it is not published due to licensing issues and I think it is infeasible to perfectly reproduce it.
But as you train your model fits_dict gets updated and saved periodically.
That mean that if you fine-tune using the latest model and use --run_smplify it will at some point generate good fits for h36m.
Therefore, you can simply initialize the h36m fits_dict using with the mean parameters

SPIN/models/hmr.py

Lines 86 to 89 in b95a00a

	mean_params = np.load(smpl_mean_params)
	init_pose = torch.from_numpy(mean_params['pose'][:]).unsqueeze(0)
	init_shape = torch.from_numpy(mean_params['shape'][:].astype('float32')).unsqueeze(0)
	init_cam = torch.from_numpy(mean_params['cam']).unsqueeze(0)

Once h36m fits_dict doesn't get updated you can assume the best pseudo ground truth was reached

[GuiZhaoyang]

@ikvision Thanks for your advice.I think I can understand what you mean in general.But it seemed to modify much code in the training programming.May I could konw how to modify the code.In addition,is it possible to produce the h36m_fit.npy by running eval.py on human3.6m dataset.What is the format of the file, if applicable

[yanran11]

I have met your problem.I have produced the h36m_train.npz,but I did not find the code to produce h36m_fits.npy. Can you tell me how you solved the problem? @QuteSaltyFish

I had a similar problem.How did you generate h36m_fits.npy to train finally?Thank you！！

[yqq-jj]

How did you generate h36m_train.npz？@GuiZhaoyang