Skip to content

Latest commit

 

History

History
39 lines (27 loc) · 1.58 KB

workflow.md

File metadata and controls

39 lines (27 loc) · 1.58 KB
Raw

Workflow

For each task we will (1) acquire data either by:

  • (a) Generating training data from scratch with scripted oracles (via policy_eval.py), OR
  • (b) Downloading training data from the web.

And then (2) run a train+eval by:

  • Running both training and evaluation in one script (via train_eval.py)

Note that each train+eval will spend a minute or two computing normalization statistics, then start training with example printouts:

I1013 22:26:42.807687 139814213846848 triggers.py:223] Step: 100, 11.514 steps/sec
I1013 22:26:48.352215 139814213846848 triggers.py:223] Step: 200, 18.036 steps/sec

And at certain intervals (set in the configs), run evaluations:

I1013 22:19:30.002617 140341789730624 train_eval.py:343] Evaluating policy.
...
I1013 22:21:11.054836 140341789730624 actor.py:196]
		 AverageReturn = 21.162763595581055
		 AverageEpisodeLength = 48.79999923706055
		 AverageFinalGoalDistance = 0.016136236488819122
		 AverageSuccessMetric = 1.0

There is Tensorboard support which can be obtained (for default configs) by running the following (and then going to localhost:6006 in a browser. (Might be slightly different for you to set up -- let us know if there are any issues.)

tensorboard --logdir /tmp/ibc_logs

And several chunks of useful information can be found in the train+eval log dirs for each experiment, which will end up for example at /tmp/ibc_logs/mlp_ebm after running the first suggested training. For example operative-gin-config.txt will save out all the hyperparameters used for that training.