Something wrong in `lseg_encoder`.

ShijieZhou-UCLA / feature-3dgs

[CVPR 2024 Highlight] Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields

Other

374 stars 23 forks source link

Something wrong in `lseg_encoder`. #37

Closed ZhiyeTang closed 3 months ago

ZhiyeTang commented 3 months ago

This results in different lengths of input and kwargs. I checked the original code and found that you've modified something. Can you please fix this problem so that we can use multiple GPUs at a time? Thx!

IQRAASGHAR1999 commented 3 months ago

hey @ZhiyeTang I faced the same issue specifying CUDA devices helped me, you can try this

CUDA_VISIBLE_DEVICES=0 python -u encode_images.py --backbone clip_vitl16_384 --weights demo_e200.ckpt --widehead --no-scaleinv --outdir ../../data/corroded_section/rgb_feature_langseg --test-rgb-dir ../../data/corroded_section/images --workers 4

ZhiyeTang commented 3 months ago

@IQRAASGHAR1999 Hey there, thanks for your discussion. Specifying CUDA devices like that would help because only device cuda:0 is visible to Python, while all other CUDA devices are unused. Therefore, the issue remains unsolved.

ShijieZhou-UCLA commented 3 months ago

Hi @ZhiyeTang! In most cases one single GPU is enough for LSeg feature encoding and we didn't investigate too much on multi-GPU for this part. To the best of my knowledge, strictly following the instruction and using the command provided in readme will not cause any bugs.