reinforce_training_loop
Run bounded REINFORCE stochastic-policy training loops with sampled actions, scalar rewards, baselines, and optimizer traces.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- reinforce_training_loop
- description
- Run bounded REINFORCE stochastic-policy training loops with sampled actions, scalar rewards, baselines, and optimizer traces.
02 资源文件 2
- 脚本scripts/train_bandit.py4.5 KB
- 脚本tests/test_train_bandit.py1.4 KB
03 SKILL.md 正文
REINFORCE Training Loop
Use this skill to recover or test Williams-style REINFORCE behavior on a small stochastic environment. It is appropriate when the evidence must show sampled stochastic actions, scalar reinforcement, score-function updates, and changed policy parameters. Do not use it for deterministic policy-gradient, Q-learning, or supervised imitation-only experiments.
Inputs
- Initial Bernoulli logit or equivalent stochastic policy parameters.
- Reward function mapping sampled actions to scalar rewards.
- Episode count, learning rate, random seed, and baseline value.
- Optional path to the
score_function_estimatorscript for cross-checking the local score term.
Outputs
- Episode-level trace containing action probability, sampled action, reward, advantage, update, and parameter value.
- Summary metrics before and after training.
params_beforeandparams_afterfields suitable for Distiller recovery validation.- Mechanism booleans for stochastic sampling and optimizer execution.
Workflow
- Initialize the stochastic policy and compute its expected reward before training.
- For every episode, sample an action from the current policy distribution.
- Observe scalar reward and compute the REINFORCE score-function update.
- Apply
theta += learning_rate * updateand record the parameter transition. - Report expected reward after training and verify that parameters changed.
- When used in recovery, save the trace under
recovery/logs/training_trace.json.
Validation
Run:
python scripts/train_bandit.py --episodes 128 --seed 7 --output /tmp/reinforce_trace.json
python -m pytest tests
The tests assert that the better action's probability and expected reward improve in a deterministic seeded two-action bandit.
Limitations
This skill is a reduced mechanism-faithful recovery harness. It does not claim to reproduce every historical example from the paper or delayed-reinforcement variants.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).