PEAL shortcut analysis

Upload an ONNX image classifier and a zip with one folder per class. DiDAE distils the classifier into a sparse concept space, renders counterfactuals with a pretrained representation autoencoder, and asks you which edits are spurious. (Handing back a corrected classifier is switched off in this demo for now.)

queue loading…

1. Files

Drop the ONNX classifier here or click to choose
single input [batch, 3, H, W]; up to MB

Drop the dataset zip here or click to choose
<class_a>/*.jpg, <class_b>/*.jpg, …; up to MB

2. What the model expects (all fields required)

Class A folder *

Class B folder *

Class A name (as shown to you when judging) *

Class B name *

Model output index for class A *

Model output index for class B *

Input normalization *

Input height *

Input width *

Images per class (0 = all)

Directions to renderevery atom is swept in latent space; only the top N directions are rendered as images (each with all its flips). More = longer job.

Renders per direction (0 = all)cap on the latent flips rendered per direction (the deepest ones first); 0 = all, which can be thousands of renders and take hours.

Direction type pairs found the background shortcut on Waterbirds more often; single concepts are easier to read.

Correction after your verdicts DFR becomes available once the model's final linear layer is found in the uploaded graph.

I accept that this is a non-commercial research demo (the generator weights are CC BY-NC 4.0), that my files are stored on the server until the job is cleaned up, and that I upload only data I am allowed to share. *

Job

log

PEAL shortcut analysis

1. Files

2. What the model expects (all fields required)

Job

Your verdicts are needed

Results