Matthew Aitchison

MD5s for ROMs differ from https://github.com/mgbellemare/Arcade-Learning-Environment/blob/master/md5.txt on some games (when importing via gym[atari]. Is this intentional? I can provide a list of non-matching MD5s if that helps.

The classifier should include some summary about the novelty of a track in the JSON output. It's not clear how to properly summerise the novelty of a track, but here...

Update novelty presentation when generating MPEGs via classify.py so that it only shows above some threshold. Possibly modify colors to indicate high levels of novelty.

Currently, checkpoints are saved to both a shared checkpoints folder and a unique log folder. This causes problems if two models train at the same time, as both try to...

If a training job is started but doesn't finish it should be able to pick up from where it left off. Checkpoints are created in both the checkpoints and log...

I'm adding data to the ProcGen benchmark, and new results require a task to be selected. I'd like to make this task "ProcGen Hard (100M)" but I don't seem to...

# Motivation Tinker 0.5.1 supports new loss functions, however the types in cookbook are limited to ['Importance_sampling", "ppo"]. I've resolved this by using the LossFnType from `tinker.types` rather than specifying...

### Motivation Some of the [loss functions](https://tinker-docs.thinkingmachines.ai/losses) have configuration settings that are currently not accessible. This PR adds an optional `loss_fn_config` the RL and distillation training steps and forward_backward calls....