LR-Finder on ResNet 50

implementation help
1

I am trying to train a binary classification model for x-rays based on a pretrained ResNet50 model (ImageNet). As soon as I adjust the last FC-Layer of the ResNet down to 2 outputs (from originally 1000 for ImageNet1k, the learning rate finder just shows a graph that looks like this:

Bildschirm­foto 2023-03-09 um 18.32.01
Bildschirm­foto 2023-03-09 um 18.32.011150×882 47.3 KB

I have rebuilt the model from scratch and tested the LR-Finder along the way, so I am fairly sure that changing the FC-Layer is what breaks it.
Has anyone ever experienced a problem like this? Thanks!

2

It looks like none of the chosen learning rates is leading to a decrease in loss (note the graph doesn’t have any negative slope). I suggest first doing the sanity check that you can overfit on a small set of samples (that the model is actually trainable).