Closed Vinnieassaulter closed 1 month ago
I just found out that the problem reason! It is that I changed the unit_length in checkpoints/kit/Comp_v6_KLD005/opt.txt and the MovementConvEncoder in the eval is fixed to do two convolutions with filter=4, stride=2, so in method getco embeddings(), m_lens should always be equal to m_lens // 4
I tried to set the VQ's down_t to 1 for training so that the unit_length of motion sequence is 2. But I got a RuntimeError: cuDNN error: CUDNN_STATUS_MAPPING_ERROR when I trained residual_trasn further and the problem seems to be in eval_t2m.py evaluation_res_transformer() function, when pose does not contain outliers in tensor before inputting into eval_wrapper.get_co_embedding(), but produces inf and nan values after passing it in even as a clone (which is strange), do you have any insights on this please!