Persist Adam optimizer state and round counter across restarts #35

New Issue

2026-08-17T19:49:02+02:00

SirStone commented

2026-08-17 19:49:02 +02:00

Parent: #29

Question

PPO_Bot currently loses Adam optimizer momentum (m, v, t vectors) and round counter on crash/restart. This means:

Optimizer restarts cold after every crash — training slows until momentum rebuilds
Round counter resets to 0 — stats lose continuity

Fix: Save Adam state (ACAdamStates) alongside network weights in weights/latest/ and checkpoint dirs. Save round counter to a simple file. Load both in loadBestAvailable.

Independent of other tickets — improves crash recovery quality regardless of runner design.

Parent: #29 ## Question PPO_Bot currently loses Adam optimizer momentum (m, v, t vectors) and round counter on crash/restart. This means: - Optimizer restarts cold after every crash — training slows until momentum rebuilds - Round counter resets to 0 — stats lose continuity **Fix:** Save Adam state (`ACAdamStates`) alongside network weights in `weights/latest/` and checkpoint dirs. Save round counter to a simple file. Load both in `loadBestAvailable`. **Independent of other tickets** — improves crash recovery quality regardless of runner design.

SirStone added the wayfinder:task label 2026-08-17 19:49:02 +02:00

SirStone commented

2026-08-18 15:41:09 +02:00

Resolved in commit 0b17430 on branch research/goto-controller.

What was done:

Exported AdamState type and its m, v, t fields, plus all ACAdamStates fields, from training.nim so weights.nim can access them.
Added saveAdamStates / loadAdamStates procs in weights.nim: each Adam state's m and v tensors are written as .npy files (same format as weights), t counters go to adam_t.txt (one int per line).
Updated saveCheckpoint signature to (ac, adam, weightsRoot, roundNum) — saves Adam state alongside weights in latest/ and checkpoint dirs; writes round_counter.txt to weightsRoot/.
Updated loadBestAvailable signature to (ac, adam, weightsRoot) → (loaded, roundNum) — loads Adam state if files are present, cold-starts if absent (backward compatible).
Updated PPO_Bot.nim call sites accordingly; roundCounter is now seeded from the persisted value on startup.
All existing tests pass; no new dependencies added.

Resolved in commit `0b17430` on branch `research/goto-controller`. **What was done:** - Exported `AdamState` type and its `m`, `v`, `t` fields, plus all `ACAdamStates` fields, from `training.nim` so `weights.nim` can access them. - Added `saveAdamStates` / `loadAdamStates` procs in `weights.nim`: each Adam state's `m` and `v` tensors are written as `.npy` files (same format as weights), `t` counters go to `adam_t.txt` (one int per line). - Updated `saveCheckpoint` signature to `(ac, adam, weightsRoot, roundNum)` — saves Adam state alongside weights in `latest/` and checkpoint dirs; writes `round_counter.txt` to `weightsRoot/`. - Updated `loadBestAvailable` signature to `(ac, adam, weightsRoot) → (loaded, roundNum)` — loads Adam state if files are present, cold-starts if absent (backward compatible). - Updated `PPO_Bot.nim` call sites accordingly; `roundCounter` is now seeded from the persisted value on startup. - All existing tests pass; no new dependencies added.

SirStone closed this issue

2026-08-18 15:41:13 +02:00

SirStone referenced this issue

2026-08-18 15:41:22 +02:00

[Map] Autonomous LLM-supervised PPO_Bot training pipeline #29

SirStone referenced this issue

2026-08-18 15:57:00 +02:00

Build the training runner tool #34

SirStone referenced this issue

2026-08-18 16:16:22 +02:00

[Map] Autonomous LLM-supervised PPO_Bot training pipeline #29

SirStone referenced this issue from a commit