TODOS.md

TODOS

• Rerun model with current best parameters
  • default MS2 loss, predicting target
  • default MS2 loss + MS1 loss, positive only constraint
• multichannel conditioning signal
• Look into the data to figure out why there are weird MS1s

1. ms2 mse loss predicting "noise"
2. ms2 mse loss predicting target directly
3. ms2 + ms1 mse loss using best version of 1/2?
4. Multichannel conditioning signal (repeat 1/2 with new signal?)

• pseudo extracted spectra?

• Optimize inference and data types with torchao

• Add a raw mzML/tdf parser (timsrust_py03) - Josh

• Obtain another dataset for testing - Josh

  • Potentially use HeLa
  • Orbitrap from 2018

• Benchmark if time

In Progress

Completed

• - Dataloaders for MS1+MS2 data
• Implement base diffusion model from PyTorch
• Adapter layer for MS data to input dimensions of above
• Maybe custom training loop
• Eval code
• Update dataloader to perform grid split of data - Justin
• Move sampling.py into codebase - Justin
• Add learning rate scheduler (apopt from AlphaPeptDeep) - Justin
• Update Transformer model - Leon
• Update/Test different loss functions - Leon
• continue from checkpoint - Saksham
• Add eval metrics to WandB logging, separate from training