Closed lanterno closed 1 year ago
To train our LR or MLP over PatchCore (let's call it the transfer model), we should also use data from MVTec test set that also contains anomalous data. So I suggest to create two splits from the test set, one to train the transfer model and one to evaluate it. The big train set with normal image remains the one used to prepare the memory bank.
Moving to this issue https://github.com/c4dt/predictive-maintenance/issues/11
After having tested transfer learning with Linear Regression from PatchCore, we saw that there was no obvious linear correlation between the patch embedding, and its anomaly score!
In this issue, we attempt a similar experiment, using MLP Regressor instead of a Linear Regressor.