CUDA out of memory about 4k4d HOT 2 OPEN

Hi @miu200521358 thanks for bringing attention to this. I see that you ran out of GPU memory during the precomputation step (load_state_dict_post_hook of super_charged_r4dv).

1. The very large memory usage 17179869184.00 GiB might be a bug on the PyTorch side and it looks like only a logging bug.
2. It is possible to render 4K4D on a 8GB machine. The memory required for the rendering itself is very small. The OOM error you encountered happened during pre-computation. And I assume it's because I set the batch size too high (same as the number of views). I tuned the memory usage for a 12GB machine. You can try modifying this invokation

    # Compute projected color of every image, using the original size image
    ibrs_rgbs = sample_geometry_feature_image(
        xyz,
        src_feat_inps,
        src_exts,
        src_ixts,
        src_inps.new_ones(2, 1),
    )  # B, S, N, 3

to wrap it around a for-loop on the second (the S) dimension like how I computed the image features later:

    # Compute blending weights from the image features
    bws = torch.cat([sampler.ibr_regressor.rgb_mlp(xyz_ibr_rgbs[:, j:j + 1]) for j in range(xyz_ibr_rgbs.shape[-3])], dim=-3)  # B, S, N, 1
    timer.record('compute blending weights')

from 4k4d.

@miu200521358 I updated the rendering code to load the DNA-Rendering dataset more efficiently in some recent commits. Could you please try whether this fixed the OOM issue?

from 4k4d.