search

NVlabs / nvdiffrec

Official code for the CVPR 2022 (oral) paper "Extracting Triangular 3D Models, Materials, and Lighting From Images".
Other
2.09k stars 222 forks source link

Image resizing makes Img losses on batches not converge #101

Open iraj465 opened 1 year ago

iraj465 commented 1 year ago

Hi, I found that resizing images and then training makes the convergence much slower as compared to keeping the original resolutions, but for higher resolutions i get CUDA OOM errors. Is there a tradeoff that hits the sweet spot or is there an alternative for this?

Any help is appreciated

jmunkberg commented 1 year ago

I found that resizing images and then training makes the convergence much slower as compared to keeping the original resolutions

This is not something we have observed. For example, try changing the training resolution in the bob config, from 512x512 to 1024x1024.

Also, we rescale the NeRD datasets before training, as shown here: https://github.com/NVlabs/nvdiffrec/blob/main/data/download_datasets.py

In general, we recommend training at as large as possible resolution (given your GPU) as that will improve texture and geometry quality.