Reproduce Problem on HM3D Objectnav with HSSD BackBone （Ver + Clip） #21
Description
Hello @mukulkhanna !
I have recently been able to reproduce the pretrained results on HSSD successfully. However, I have encountered some discrepancies with the results on HM3D, which seem to deviate significantly from those reported in the paper. For instance, the success rate after 160M steps (utilizing 4 GPUs and 24 envs) is less than 5%, whereas in your paper it is stated to be nearly 50%. I have diligently followed the setup file provided in this repository, and have attempted both single-GPU and multi-GPU setups multiple times, yet the results remain inconsistent. Additionally, I have observed that the training fps for HSSD can exceed 1000, whereas for HM3D it hovers around 200. Could this discrepancy in training fps contribute to the poor performance of HM3D?
I would greatly appreciate your time and assistance in resolving these issues. The following tensorboard displays my reproduction results for HM3D.
