timaeus/jaxgmg2_3phase_fast

UNUSED

A run to dial in suitable values of alpha, learning rate, and discount_rate after the learning rate bug was fixed. All trained with the same seed, so not especially useful for experiments, but was useful for dialing in hyperparams.

Trained with now obsolete code on commit hash 8b73ab2736efca9a494348d2e68a62442e219aab Wandb https://wandb.ai/devinterp/jaxgmg2_3phase_fast

Ran with

wandb sweep sweep.yaml

Contents of sweep.yaml

command:
- env
- WANDB_AGENT_MAX_INITIAL_FAILURES=1
- /root/timaeus/.venv/bin/python
- ${program}
- ${args}
- --use-wandb
- --use-hf
entity: devinterp
method: random
parameters:
  alpha:
    distribution: uniform
    max: 0.9
    min: 0.2
  cheese-loc:
    value: any
  ckpt-dir:
    value: jaxgmg2_3phase_fast
  discount-rate:
    values:
    - 0.99
    - 0.975
  eval-schedule:
    value: 0:1,250:2,500:5
  f-str-ckpt:
    value: al_{alpha:.3f}_g_{discount_rate}_seed_{seed}_pa_{num_prev_actions}_lr_{lr:.0e}
  grad-acc-per-chunk:
    value: 5
  lr:
    values:
    - 5e-05
    - 0.0001
    - 0.0002
    - 0.0005
    - 0.001
    - 0.002
    - 0.005
  mask-type:
    value: first_episode
  ntfy:
    value: david_jaxgmg
  num-levels:
    value: 9600
  num-prev-actions:
    value: 1
  num-total-env-steps:
    value: 2000000000
  seed:
    value: 100
  wandb-project:
    value: jaxgmg2_3phase_fast
program: /root/timaeus/projects/rl/main_train.py
project: jaxgmg2_3phase_fast