Closed yangchaohua closed 11 months ago
Hi @yangchaohua. We haven't ran experiments with segmentation models. Based on our preliminary work with diffusion models (where the output is also high-dimensional), my guess is that the sum of per-pixel outputs should be a good start.
Hi! Thank you for sharing this work. I want to use TRAK for my segmentation task (each pixel of the model is represented as a binary classification of whether it is a foreground or not), I don't know which method is more suitable for calculating f(z; θ), sum of the output or average of the output. Is there any other better method?