Code

Eval mode rework; refactoring#279

Code

Merged

kengz merged 29 commits intomasterkengz/SLM-Lab:masterfrom eval-envkengz/SLM-Lab:eval-envCopy head branch name to clipboardFeb 4, 2019

+752-673

Commits

Commits on Feb 2, 2019

• add contextual lab mode in util

  kengz

  committed
  a2fbea2

  Copy full SHA for a2fbea2

• store action_pd in body, group all such updates under agent.update()

  kengz

  committed
  cd75ce7

  Copy full SHA for cd75ce7

Commits on Feb 3, 2019

• refactor calc_df_row and add body.eval_update

  kengz

  committed
  9b3d6c4

  Copy full SHA for 9b3d6c4

• make calc_df_row use t to reflect consistency with reward

  kengz

  committed
  06f2d75

  Copy full SHA for 06f2d75

• remove adhoc mean_entropy compute. delegate to body df update

  kengz

  committed
  a3ac629

  Copy full SHA for a3ac629

• use eval_df to compute body.current_reward_ma for ckpt

  kengz

  committed
  1e98816

  Copy full SHA for 1e98816

• simply to body.grad_norm, unify under action_pd_update efficiently

  kengz

  committed
  56a4864

  Copy full SHA for 56a4864

• update body debugging variables to use mean; move compute to flush

  kengz

  committed
  8c41437

  Copy full SHA for 8c41437

• retire body.last_loss

  kengz

  committed
  442eb92

  Copy full SHA for 442eb92

• guard ctx_lab_mode against None value

  kengz

  committed
  b871d40

  Copy full SHA for b871d40

• make NoOpLRScheduler API consistent, simplify get_mean_lr

  kengz

  committed
  46333b4

  Copy full SHA for 46333b4

• improve logging for both train and eval using df directly

  kengz

  committed
  38b4847

  Copy full SHA for 38b4847

• format numerical logging to use :g instead of :.f

  kengz

  committed
  dc60e43

  Copy full SHA for dc60e43

• purge useless memory total_reward history

  kengz

  committed
  d7987a7

  Copy full SHA for d7987a7

• run_eval_episode in eval context

  kengz

  committed
  0ec465e

  Copy full SHA for 0ec465e

• update changed key in openai gym max_timestep

  kengz

  committed
  250c195

  Copy full SHA for 250c195

Commits on Feb 4, 2019

• set_rand_seed and close eval_env

  kengz

  committed
  83baa73

  Copy full SHA for 83baa73

• guard openai env re-register

  kengz

  committed
  5907f60

  Copy full SHA for 5907f60

• refactor parallel_eval method into retro_analysis, activate via metaspec

  kengz

  committed
  fb3858b

  Copy full SHA for fb3858b

• move max_tick_unit from env to meta_spec

  kengz

  committed
  9c68b28

  Copy full SHA for 9c68b28

• generalize save_frequency to meta.eval_frequency

  kengz

  committed
  c1005f2

  Copy full SHA for c1005f2

• expand session analysis to work for both kinds of body df

  kengz

  committed
  ca9e07d

  Copy full SHA for ca9e07d

• remove reindex_session_data

  kengz

  committed
  9a4fc03

  Copy full SHA for 9a4fc03

• refactor in_eval_lab_modes

  kengz

  committed
  377e7de

  Copy full SHA for 377e7de

• fix retro analysis wrong prefix in reading from file

  kengz

  committed
  df31c72

  Copy full SHA for df31c72

• mute SessionSpace ckpt and eval

  kengz

  committed
  184f1a5

  Copy full SHA for 184f1a5

• add tmp SpaceSession substitution for eval_df

  kengz

  committed
  e33a388

  Copy full SHA for e33a388

• set vars to end_val for eval ctx

  kengz

  committed
  7a8cbce

  Copy full SHA for 7a8cbce

• fix codestyle

  kengz

  committed
  ec9756f

  Copy full SHA for ec9756f