Training process about coin HOT 2 OPEN

One more, if you fed images to a single model one by one, i think the single model will properly reconstruct in terms of only the lastly trained image. What do you think about that?

from coin.

Hi! We do indeed train a different model for each image independently. As can be seen in line 59 in main.py, we initialize the model inside the for loop, so we will have a different model for every image. As you mention, this does indeed lead to more overfitting and better reconstruction performance, since we optimize a different model for each image. Hope this makes sense!

from coin.