Add utilities for training on a single-entry dataset.
Allow validation skipping.
WIP AF3 Non-equivariant structure encoder/decoder
Add flag to force training from scratch
Force training from scratch in debug config
All modules in diffusion module implemented
Document behavior of dropout with test
Finish majority of model trunk
Convert some ModuleLists to nn.Sequential
Add RelativePositionEncoding and WIP af3_repro config
Fix ref_space_uid embedding in AtomEncoder
Put Model together with fake MSAModule and TemplateEmbedder
AF3 repro loads model.
WIP af3 data-adaptor, AF3_structure fixes
Feature initializer working
Standardize S_inputs_I
Fix pairformer stack
Forward pass working, WIP: backward pass stale reference fixing
Add dataloader_adaptor_af3.py
Backward pass working, WIP: still some unused params
Backprop working
Training runs
Add pytorch lightning training and some wandb logging
Training converging for single example.
Run:
/home/ahern/reclone/rf_diffusion_staging/rf_diffusion/exec/rf_diffusion_aa_2.sif
trainer_lightning.py --config-name af3_repro_single_example_small
logger.use_wandb=True af3_data_prep.D=6
Log loss
Training working for single example.
Run: /home/ahern/reclone/rf_diffusion_staging/rf_diffusion/exec/rf_diffusion_aa_2.sif
trainer_lightning.py --config-name
af3_repro_single_example_small_working_4 logger.use_wandb=True
on an a4000
Add test_diffusion_module.py
