Huge negative validation loss

I’m currently running a multiclass RNN to identify 160 classes using complaints text from CFPB. After fine tuning the language model, I dropped the encoding into the classification model, and the first cycle generated an incredibly unusual number - massive negative validation loss. I’m currently trying to investigate, but can anyone help me and point out what might be going wrong here?
image

1 Like

May be that max_lr is higher. Try something like 0.01?

1 Like

On the second cycle of execution this problem just magically went away. Not sure what this actually is. I will come back and post if I ever get to figure it out.

The problem is what @arunoda said. You need to pick the part of the lr finder graph with the strongest negative slope, not the lowest value.

1 Like

Yes of course - biggest slope! thanks for the reminder, i was using lr find the wrong way