MS2 transfer learning broken

[GeorgWa]

Description

Performing transfer learning using the alphadia FinetuneManager fails for MS2 with zero PCC already on the forward pass.
I think this might be related to peptdeep expecting a different library order.

2024-12-14 09:47:46>  Model tested on validation dataset with the following metrics:
2024-12-14 09:47:46>  l1_loss                       : 0.0353
2024-12-14 09:47:46>  PCC-mean                      : 0.0000
2024-12-14 09:47:46>  COS-mean                      : 0.0000
2024-12-14 09:47:46>  SA-mean                       : 0.0000
2024-12-14 09:47:46>  SPC-mean                      : 0.2157
2024-12-14 09:47:46>  Fine-tuning MS2 model with the following settings:
2024-12-14 09:47:46>  Train fraction:      0.70     Train size:      65216     
2024-12-14 09:47:46>  Validation fraction: 0.20     Validation size: 18633     
2024-12-14 09:47:46>  Test fraction:       0.10     Test size:       9317      
2024-12-14 09:48:25>  Epoch 0   Lr: 0.00020   Training loss: 0.0243   validation loss: 0.0353
2024-12-14 09:48:59>  Epoch 3   Lr: 0.00050   Training loss: 0.0219   validation loss: 0.0353
2024-12-14 09:49:38>  Epoch 6   Lr: 0.00060   Training loss: 0.0227   validation loss: 0.0353
2024-12-14 09:50:17>  Epoch 9   Lr: 0.00030   Training loss: 0.0226   validation loss: 0.0353

Steps to reproduce

I've uploaded some validated test data here:
https://datashare.biochem.mpg.de/s/4mdf5PMHsjLXGZe

Using the finetuning notebook and the current dev version of the alphax packages should reproduce the behaviour.