Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
35 lines
1.4 KiB
Python
35 lines
1.4 KiB
Python
#!/usr/bin/env python3
|
|
"""The ONNX graph reads the `tools/rknn/prepare_*` converters share.
|
|
|
|
The Piper vocoder and the Inflect v2 decoder converters each carried a private copy of the same
|
|
two reads: the input-name list of the sliced graph, and the latent channel count read back from
|
|
its `z`. One copy lives here now. The same-named function in the cosyvoice3 converter is not a
|
|
reader -- it validates the export's feeder set against the expected ladder -- and it stays there.
|
|
|
|
`onnx` is imported at module level: the only importers are the converters, and a converter only
|
|
runs in the toolkit venv, which pins `onnx==1.16.1`.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from pathlib import Path
|
|
|
|
import onnx
|
|
|
|
|
|
def graph_input_names(onnx_path: Path) -> list[str]:
|
|
return [value.name for value in onnx.load(str(onnx_path)).graph.input]
|
|
|
|
|
|
def decoder_channels_of(graph, *, where: str = "graph") -> int:
|
|
"""Latent channel count read back from an already-loaded graph."""
|
|
for value in graph.input:
|
|
if value.name == "z":
|
|
dims = [d.dim_value for d in value.type.tensor_type.shape.dim]
|
|
if len(dims) == 3 and dims[1] > 0:
|
|
return int(dims[1])
|
|
raise SystemExit(f"{where}: input z must be [1, channels, frames]")
|
|
|
|
|
|
def decoder_channels(onnx_path: Path) -> int:
|
|
return decoder_channels_of(onnx.load(str(onnx_path)).graph, where=str(onnx_path))
|