search

JuliaWolleb / diffusion-anomaly

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

How to make an abnormal picture be a normal one? #21

Open Junyiliu0 opened 1 year ago

Junyiliu0 commented 1 year ago

I am using the chexpert data to transform the picture. And I found a mistake (maybe). During the dataload in chexpert dataset. The abnormal label is 0, and the normal label is 1. But I found the code maybe just using label(0) to do. Maybe because it is suitable for "BRATSDataset".

image

I change it to force the label is 1. Did I change it right? I think the code make the normal picture label is 1 for chexpert. Hope your reply! Thanks!

JuliaWolleb commented 1 year ago

Hi yes you are right, in both datasets the class labels should be 0 for healthy, and 1 for diseased. I suggest to adapt your dataloader such that the labels are consistent.

best, Julia

Junyiliu0 commented 1 year ago

I'm sorry to bother you again. When I use the code to make the normal picture. May be I want to generate about 5k pictures. I use batch_size =32 for infer. But I found It's getting slower and slower. Why this appear? I found "the time for 1000" become larger and larger.But the GPU memory hasn't changed.

image