parallel branch length optimization

[matsen]

• implementing parallel branch length optimization
• storing datasets rather than loaders (makes more sense; pickleable)
• facility to split data set
• using AdamW optimizer
• dropping verbose due to lr_on_plateau no longer supporting it
• walltime since start of execution time
• ignore UserWarning from torch.nn.modules.transformer

[mmjohn]

Cool! Is this the reason for the switch from torch.optim.Adam() to torch.optim.AdamW()? https://stackoverflow.com/questions/64621585/adamw-and-adam-with-weight-decay

[matsen]

Yes, exactly.