search

krisrs1128 / clouds_dist

Simulation of low-clouds, from weather measures.
4 stars 1 forks source link

Initialize from pretrained weights #117

Closed krisrs1128 closed 4 years ago

krisrs1128 commented 4 years ago

This creates an init_chckpt_dir and init_chkpt_step which let's you initialize with a previous checkpoint, using the existing resume() method.

In theory, you could fake the checkpointing and place arbitrary pretrained models in the directory.

vict0rsch commented 4 years ago

@krisrs1128 did you see I added a --resume flag some time ago and a parallel_resume.py file?

krisrs1128 commented 4 years ago

Yep, I'm basically just using the same code, but letting you specify the model, rather than looking up what's in the written conf. I'm guessing we may want to initialize using a model from a different run.

krisrs1128 commented 4 years ago

The changes related to extrapolation / step in extragradient are because, upon resume, I think it might try to take a step before extrapolation (depending on the number of steps in the checkpoint).

vict0rsch commented 4 years ago

makes sense sorry :)