150 lines
5.2 KiB
Python
150 lines
5.2 KiB
Python
# Copyright (c) 2025 Zhipu AI Inc (authors: CogAudio Group Members)
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import os
|
|
import torch
|
|
import pathlib
|
|
from typing import Union
|
|
from utils.audio import mel_spectrogram
|
|
from cosyvoice.hifigan_cosy2.f0_predictor import ConvRNNF0Predictor
|
|
from cosyvoice.hifigan_cosy2.generator import HiFTGenerator
|
|
|
|
class HiFTInference:
|
|
def __init__(self,
|
|
ckpt_path: Union[str, pathlib.Path],
|
|
device: Union[str, torch.device] = 'cuda',
|
|
load_only_nsf: bool = False):
|
|
"""
|
|
Wrapper class for HiFTGenerator inference.
|
|
"""
|
|
if isinstance(ckpt_path, str):
|
|
ckpt_path = pathlib.Path(ckpt_path)
|
|
|
|
self.device = device
|
|
self.sample_rate = 24000
|
|
|
|
self.model = self._build_model()
|
|
|
|
self._load_weights(ckpt_path, load_only_nsf)
|
|
|
|
self.model.eval()
|
|
|
|
def _build_model(self) -> HiFTGenerator:
|
|
"""Instantiate the F0 Predictor and HiFTGenerator."""
|
|
f0_predictor = ConvRNNF0Predictor(
|
|
num_class=1,
|
|
in_channels=80,
|
|
cond_channels=512
|
|
)
|
|
|
|
# Instantiate HiFTGenerator with CosyVoice 2.0 parameters
|
|
hift = HiFTGenerator(
|
|
in_channels=80,
|
|
base_channels=512,
|
|
nb_harmonics=8,
|
|
sampling_rate=self.sample_rate,
|
|
nsf_alpha=0.1,
|
|
nsf_sigma=0.003,
|
|
nsf_voiced_threshold=10,
|
|
upsample_rates=[8, 5, 3],
|
|
upsample_kernel_sizes=[16, 11, 7],
|
|
istft_params={
|
|
"n_fft": 16,
|
|
"hop_len": 4
|
|
},
|
|
resblock_kernel_sizes=[3, 7, 11],
|
|
resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
|
source_resblock_kernel_sizes=[7, 7, 11],
|
|
source_resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
|
lrelu_slope=0.1,
|
|
audio_limit=0.99,
|
|
f0_predictor=f0_predictor
|
|
)
|
|
return hift.to(self.device)
|
|
|
|
def _load_weights(self, ckpt_path: pathlib.Path, load_only_nsf: bool):
|
|
"""Load state dictionary into the model."""
|
|
if not ckpt_path.exists():
|
|
raise FileNotFoundError(f"Checkpoint not found at {ckpt_path}")
|
|
|
|
# Clear CUDA cache to free up memory before loading
|
|
if torch.cuda.is_available():
|
|
torch.cuda.empty_cache()
|
|
|
|
state_dict = torch.load(ckpt_path, map_location=self.device)
|
|
|
|
if not load_only_nsf:
|
|
self.model.load_state_dict(state_dict)
|
|
else:
|
|
# Load only specific parts: m_source and f0_predictor
|
|
m_source_dict = {}
|
|
f0_predictor_dict = {}
|
|
for key, value in state_dict.items():
|
|
if 'm_source' in key:
|
|
m_source_dict[key.replace('m_source.', '')] = value
|
|
if 'f0_predictor' in key:
|
|
f0_predictor_dict[key.replace('f0_predictor.', '')] = value
|
|
|
|
self.model.m_source.load_state_dict(m_source_dict, strict=True)
|
|
self.model.f0_predictor.load_state_dict(f0_predictor_dict, strict=True)
|
|
|
|
def __call__(self, mel: torch.Tensor) -> torch.Tensor:
|
|
"""
|
|
Inference entry point.
|
|
Input: Mel Spectrogram (B, C, T)
|
|
Output: Audio Waveform (B, 1, T_out)
|
|
"""
|
|
mel = mel.to(self.device)
|
|
with torch.no_grad():
|
|
audio, _ = self.model.inference(mel)
|
|
return audio
|
|
|
|
@staticmethod
|
|
def extract_mel(wav: torch.Tensor) -> torch.Tensor:
|
|
"""
|
|
Static method to extract mel spectrogram consistent with training.
|
|
"""
|
|
return mel_spectrogram(
|
|
wav,
|
|
n_fft=1920,
|
|
num_mels=80,
|
|
sampling_rate=24000,
|
|
hop_size=480,
|
|
win_size=1920,
|
|
fmin=0,
|
|
fmax=8000,
|
|
center=False
|
|
)
|
|
|
|
def load_hift(device: str = "cuda", load_only_nsf: bool = False) -> HiFTInference:
|
|
"""Factory function to load HiFT model."""
|
|
base_dir = pathlib.Path(__file__).resolve().parent.parent
|
|
repo_root = base_dir.parent.parent.parent
|
|
candidates = [
|
|
base_dir / "ckpt" / "hift" / "hift.pt",
|
|
base_dir / "hift" / "hift.pt",
|
|
repo_root / "models" / "GLM-TTS" / "hift" / "hift.pt",
|
|
pathlib.Path("ckpt") / "hift" / "hift.pt",
|
|
]
|
|
ckpt_path = None
|
|
for p in candidates:
|
|
if p.exists():
|
|
ckpt_path = p
|
|
break
|
|
if ckpt_path is None:
|
|
ckpt_path = candidates[0]
|
|
ckpt_path = str(ckpt_path)
|
|
print(f"Loading HiFT model from {ckpt_path} on {device}...")
|
|
return HiFTInference(ckpt_path, device=device, load_only_nsf=load_only_nsf)
|