Question with inferencing stage

[Michael-H777]

Hi,

I was able to get the code to run and train on cifar 10. However, I am struggling to understand how would I inference with the model.

Lets say I have a trained model, and a random noise of size (64, 3, 32, 32) (B,C,H,W). how would I use the existing model and the restore_checkpoint('from utils import restore_checkpoint') to create 64 new images?

feel free to refer to any function in the repo.

Thanks,
Michael

[Newbeeer]

Hi Michael,

Please use the --config.eval.save_images flag in the command line, as well as setting the eval batch size --config.eval.batch_size to 64. Generated images would be saved to eval_dir/ode_images_{ckpt}.png. Please refer to

Poisson_flow/run_lib.py

Lines 472 to 476 in bb51603

	if config.eval.save_images:
	# Saving a few generated images for debugging / visualization
	image_grid = make_grid(samples_torch, nrow=int(np.sqrt(len(samples_torch))))
	save_image(image_grid, os.path.join(eval_dir, f'ode_images_{ckpt}.png'))
	exit(0)

Please consider using the more advanced PFGM++ model 😃 : https://github.com/Newbeeer/pfgmpp

Best,
Yilun

[Newbeeer]

For command line of restoration of existing models, please see the Checkpoint section in README.