cuda outof memory with 5min audio input

[lokvke]

gpu: A100-80G

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 3.17 GiB (GPU 0; 79.19 GiB total capacity; 62.80 GiB already allocated; 1.50 GiB free; 64.06 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[jinqiupeter]

It's not optimized obviously. I had a 4:30min long audio and it took ~115G memory to generate the blendshape, I had to use CPU and increase swap memory:

[jinqiupeter]

[hungho77]

Can I ask why when running with aud.wav from May.mp4, my output shape is different from the shape of the bs.npy file in the author's download. I am (7062, 52) compared to the author's (6073, 52).

[xiaoniuchushi]

请问为什么当使用 May.mp4 中的 aud.wav 运行时，我的输出形状与作者下载中的 bs.npy 文件的形状不同。与作者的 (6073, 52) 相比，我是 (7062, 52)。

Perhaps due to the default fps of 30 in the code

[hungho77]

请问为什么当使用 May.mp4 中的 aud.wav 运行时，我的输出形状与作者下载中的 bs.npy 文件的形状不同。与作者的 (6073, 52) 相比，我是 (7062, 52)。

Perhaps due to the default fps of 30 in the code

thank for your advice, i have changed three lines in code and it worked. I don't know if it is riight but the shape output is correct (6073, 52)

[Tiandishihua]

Can I ask why when running with aud.wav from May.mp4, my output shape is different from the shape of the bs.npy file in the author's download. I am (7062, 52) compared to the author's (6073, 52).

请问你的GPU没有爆显存吗？我的A100都扛不住

[2hiTee]

请问为什么当使用 May.mp4 中的 aud.wav 运行时，我的输出形状与作者下载中的 bs.npy 文件的形状不同。与作者的 (6073, 52) 相比，我是 (7062, 52)。

Perhaps due to the default fps of 30 in the code

thank for your advice, i have changed three lines in code and it worked. I don't know if it is riight but the shape output is correct (6073, 52)

Have you tried it? I change the code as you did, but the output video is all black. Thank you for your help!