search

sled-group / CycleNet

Official Code for NeurIPS 2023 Paper: CycleNet: Rethinking Cycle Consistent in Text‑Guided Diffusion for Image Manipulation
https://cyclenetweb.github.io/
Apache License 2.0
76 stars 6 forks source link

Too much VRAM cost and ckpt file volume compared to the paper #6

Open yryun opened 8 months ago

yryun commented 8 months ago

Thanks for great work!

I followed the example provided with cfill dataset. -> https://github.com/sled-group/CycleNet/blob/main/docs/train.md

I found 2 differences with the original works.

  1. Training with batch size = 1 cost over 30GB VRAM, which is different in paper (batch size = 4 cost 27.9GB VRAM).

  2. My ckpt file (saved in training) is 13.7 GB, which is different in provided ckpt (summer2winter.ckpt = 6.7 GB)

Would you explain how to reproduce the original works?