Closed cmh1027 closed 6 months ago
Perhaps it is due to the learning rate warmup to avoid Adam spikes? When running two-pass optimization, we use a LR warmup over 100 steps in the beginning of the second pass: https://github.com/NVlabs/nvdiffrecmc/blob/main/train.py#L666, which is not enabled when only running only the second pass (using a base_mesh). https://github.com/NVlabs/nvdiffrecmc/blob/main/train.py#L682
These are metrics of Chair dataset of NeRF-Synthetic obtained by performing stage 1 and 2 from scratch.
PSNR 28.989 / SSIM 0.944 / LPIPS 0.060
And these are metrics obtained by performing only stage 1, and restart the program with base_mesh flag.PSNR 26.268 / SSIM 0.912 / LPIPS 0.091
Not only Chair, but also every other scenes show the same degenerated performance. I can't see the reason why this happens. Aren't they supposed to show the same performance?