about the geo_feat_dim
Opened this issue · 4 comments
Hi ，what an excellent work
我简单的将TriMipRF 里的geo_feat_dim和feature_dim提升到了31和32，在训练过程中没有出现问题，但是到达eval_step执行eval_img报错：
将geo_feat_dim重新设置为15不会出现该问题，您是否有头绪？万分感谢
It may be an issue of tinycudann, you may set a smaller value for Trainer.test_chunk_size to try it
I'm experiencing a similar situation, but it's caused by a sample point of 0. I'm not sure how you solved it? Any ideas on how to fix it please?
Referen
I also meet this problem, have you solved it?
Not a network issue, it's because during testing, the synthetic data is at the center of the scene, causing the initial chunksize ray sampling to be invalid, as they might just be the first few lines of pixels, all of which are blank. You can simply skip these intervals. You can refer to:
nerfstudio-project/nerfacc#222 (comment)
nerfstudio-project/nerfacc#207 (comment)