training
Guides Tacotron training command construction, hyperparameter overrides, data-feeder contracts, logs, checkpoints, TensorBoard, restore runs, and loss troubleshooting.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides Tacotron training command construction, hyperparameter overrides, data-feeder contracts, logs, checkpoints, TensorBoard, restore runs, and loss troubleshooting.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/cli-reference.md1.1 KB
- 参考文档references/hparams.md1020 B
- 参考文档references/troubleshooting.md1.3 KB
- 参考文档references/workflows.md1.6 KB
- 脚本scripts/build_train_command.py2.3 KB
03 SKILL.md 正文
Training
Use this route for training a Tacotron model, preparing a restore run, changing hparams, interpreting logs/checkpoints, or planning resource requirements. Full training is long-running and data-dependent; build and validate the command before launching it.
Workflow
- Complete the data-preparation route and validate
training/train.txt. - Choose a stable
--base_dir,--input, and--name. The default log directory is<base_dir>/logs-<name-or-model>. - Pass hparam overrides as comma-separated
name=valuepairs. Preserve the same audio/model hparams for evaluation and serving. - Use
--restore_step Nonly when<log_dir>/model.ckpt-Nexists. Use--gitonly in a clean source checkout; it intentionally fails on dirty state. - Monitor summaries, generated audio, and alignments. If loss explodes or attention is lost, stop safely and consider restoring before the spike.
Command roots and training boundary
Build commands from the skill root; execute the printed command only from the
Tacotron checkout root. The input and output paths remain relative to the
selected --base-dir, not to the skill directory.
SKILL_ROOT=/path/to/tacotron-skill
CHECKOUT_ROOT=/path/to/tacotron-checkout
cd "$SKILL_ROOT" && python sub-skills/training/scripts/build_train_command.py --checkout-root "$CHECKOUT_ROOT" --base-dir /data/tacotron --input training/train.txt --hparams batch_size=16,max_iters=300
The builder is dry-run only. Actual training requires validated metadata, decoded/preprocessed audio arrays, the compatible TensorFlow 1.x stack, disk, memory, and potentially GPU resources; it may write checkpoints, WAV samples, alignments, and logs. No training convergence or checkpoint restore is implied by command construction or metadata validation.
Read references/cli-reference.md for flags,
references/hparams.md for defaults and duration
calculations, references/workflows.md for a safe
sequence, and references/troubleshooting.md
for data, loss, and restore failures. Use
scripts/build_train_command.py to construct
commands without starting a training loop.
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).