JuliaWolleb / diffusion-anomaly

Anomaly detection with diffusion models
MIT License
119 stars 23 forks source link

How many. steps did you use to train cheapest? #20

Closed Junyiliu0 closed 1 year ago

Junyiliu0 commented 1 year ago

Hi Can you tell me how iterations did you use for training the chexpert_dataset? I found maybe the batch size of the cheapest_dataset is 1. And how long cost? Besides, I found there will be two .pt file during training.

image

Which one should I choose during testing? What is meaning of ema? Hope your reply! Thanks!

JuliaWolleb commented 1 year ago

Good morning

As written in the paper, we train our models for 20,000 iterations (but you can always train longer...). It takes about one day. "ema" means exponential moving average of the model parameters. I chose the one without ema, but choosing ema is alsow fine.

Best

Junyiliu0 commented 1 year ago

Thank a lot!