inference
Routes HiFi-GAN wav-to-wav and mel-to-wav inference, checkpoint/config pairing, output naming, and smoke checks.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- inference
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes HiFi-GAN wav-to-wav and mel-to-wav inference, checkpoint/config pairing, output naming, and smoke checks.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/root-material.md1.8 KB
- 参考文档references/troubleshooting.md3.3 KB
- 参考文档references/workflows.md3.0 KB
03 SKILL.md 正文
HiFi-GAN Inference
Use this sub-skill when you need to synthesize waveforms with a trained HiFi-GAN generator.
Use this route for
- Self-contained
scripts/infer_hifigan.py --mode wavsynthesis from a directory of input wavs. - Self-contained
scripts/infer_hifigan.py --mode melsynthesis from a directory of mel.npyfiles. - Loading a generator checkpoint that sits beside its matching
config.json. - Choosing input and output directories and understanding generated filename suffixes.
- CPU/GPU selection, checkpoint/config mismatch recovery, and synthetic smoke checks.
Do not use this route for
- Training, fine-tuning, or checkpoint creation/resume. Use the training sub-skill instead.
- Upstream text-to-mel generation. Prepare those mel files first, then return here.
Read first
Bundled helpers
- scripts/infer_hifigan.py — self-contained wav/mel inference entrypoint using bundled HiFi-GAN runtime source.
- scripts/make_dummy_checkpoint.py — create a synthetic generator checkpoint from bundled source and copy or intentionally mismatch the paired
config.json. - scripts/make_tiny_inference_fixtures.py — create tiny wav and mel fixtures for smoke tests and negative cases.
- scripts/run_inference_smoke.py — build the synthetic inputs and run both inference modes end to end.
Key contracts
scripts/infer_hifigan.pywraps the bundledinference.pyandinference_e2e.py; both underlying paths readconfig.jsonfrom the checkpoint directory and do not accept a separate config flag.- The checkpoint must be a generator checkpoint dict with a
generatorentry. Training checkpoints nameddo_########are not valid inference inputs. scripts/infer_hifigan.py --mode wavexpects mono wav files, computes mels with the loaded config, and writes<stem>_generated.wav.scripts/infer_hifigan.py --mode melexpects mel.npyfiles and writes<stem>_generated_e2e.wav.- Both scripts auto-select CUDA when
torch.cuda.is_available()is true and fall back to CPU otherwise. - The bundled inference entrypoint ultimately runs code that iterates
os.listdir(...)directly, so keep the input directory clean and populated only with the intended files. - Use the matching config family for the checkpoint. The skill bundles
config_v1.json,config_v2.json, andconfig_v3.jsonunder../../scripts/hifigan_runtime/configs/.
Known-good inspection environment
- Verified inspection facts:
torch 2.3.1+cu121imported on an NVIDIA A100-SXM4-40GB host;torch.cuda.is_available()returned true;torch.cuda.get_device_name(0)worked;torch.utils.tensorboardimported;librosa.util.normalizewas present inlibrosa 0.10.2.post1.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.