search

sjtuplayer / few-shot-diffusion

[ICCV 2023] Phasic Content Fusing Diffusion Model with Directional Distribution Consistency for Few-Shot Model Adaption
53 stars 2 forks source link

Question with train-recon.py #2

Closed JubSteven closed 10 months ago

JubSteven commented 11 months ago

Hi, I have an question with train-recon.py. In the README file, it is said that we need to provide a path for checkpoints (presumably the result of stage 1 using train.py), but I don't seem to find the code in train-recon.py for actually using the checkpoints, and directly running the scripts appears to train from the beginning. Can you help clarify the issue? Thanks!

sjtuplayer commented 10 months ago

Hi, I have rewritten the code for stages 2 and 3, you can now easily train the model.

JubSteven commented 10 months ago

Hi, I have rewritten the code for stages 2 and 3, you can now easily train the model.

Thanks! May I ask what is the GPU requirement for this task?

sjtuplayer commented 10 months ago

Hi, I have rewritten the code for stages 2 and 3, you can now easily train the model.

Thanks! May I ask what is the GPU requirement for this task?

If you use the pre-trained diffusion model on the source domain, a single 24G GPU is enough to adapt the model from source domain to target domain (step 2 and 3). For step 1, we have used 8 V100 GPUs to train.

JubSteven commented 10 months ago

OK, thanks for your patience!