Skip to content

weights.md

Latest commit

 

History

History
101 lines (73 loc) · 3.8 KB

weights.md

File metadata and controls

101 lines (73 loc) · 3.8 KB
Raw

Weights

The timbre weights are in pretrain/

They are committed to the repository (~17 MB each) so that a clone gives you a working tool with no download step.

models/registry.json is the record. For each weight it states:

field meaning
id the name you pass to --weights
file filename in pretrain/ (or the cache, for downloaded weights)
purpose timbre or speaker — they answer different questions
default which one --encoder timbre picks with no --weights
step the training step the checkpoint itself reports
sha256 hash of the file, verified on load
urls download locations (only for weights that are not committed)
source provenance: where the weight came from, including the git blob id for archived ones

The weights

timbre-v1 — default

The weight this tool has always been used with. It separates timbre styles well, which is the entire point.

Recovered from git history: an earlier commit overwrote it, and the registry records the blob id (e8560833…) along with the SHA-256 of the original bytes so that the recovery is verifiable rather than hopeful.

timbre-alt-v1

An alternative checkpoint. It separates noticeably worse, which is why it is not the default. The file is named for the step counter inside the checkpoint (165000), because the previous filename claimed 1570000 and was simply wrong.

speaker-upstream-v1 — speaker identity

The upstream Resemblyzer encoder. This one separates singers, not registers — a different question. Use it when you need to tell performers apart, for instance when auditing a mixed dataset. Same architecture, so it drops into the same encoder. Downloaded on first use.

The emotion model

Not an encoder checkpoint: a wav2vec2 model downloaded from the HuggingFace hub, declared under downloads.emotion in the registry. Fetched as safetensors rather than pytorch_model.bin, because it loads by tensor name with no unpickling of a file we did not produce.

Where they live

weight location
timbre-v1, timbre-alt-v1 pretrain/ in the repository
speaker-upstream-v1, emotion model per-user cache, downloaded on demand

Cache directory:

platform default
Windows %LOCALAPPDATA%\colorsplitter\weights
Linux / macOS $XDG_CACHE_HOME/colorsplitter/weights, or ~/.cache/colorsplitter/weights
either override with COLORSPLITTER_HOME

A file placed in the cache directory by hand is used as-is, so an offline or air-gapped setup works: drop the files in, and no network is touched.

Getting them

The timbre weights are already in pretrain/ — no action needed. The other two are fetched on demand:

cs weights list                     # what exists
cs weights fetch                    # download speaker + emotion
cs weights fetch --only emotion     # just the emotion model

Mirrors

huggingface.co is unreachable from some networks — through a blocking proxy it fails outright rather than being slow. Every hub asset therefore has an ordered candidate list:

  1. HF_ENDPOINT, if set,
  2. the mirrors in registry.json (https://hf-mirror.com by default),
  3. huggingface.co last.

The host that last worked is written to .hf_host in the cache and tried first afterwards, so an unreachable primary costs one timeout rather than one per file. The first candidate is given a deliberately short leash for the same reason.

Downloads resume from a partial file where the server supports range requests, and a completed file is hashed before it is accepted. A hash mismatch deletes the file and reports it — a quiet substitution of a different weight is the failure mode worth guarding against.

If sha256 is null, the asset has not been verified. That means "unchecked", not "fine".