mirror of
https://github.com/FunAudioLLM/CosyVoice.git
synced 2026-02-04 17:39:25 +08:00
90 lines
3.5 KiB
Python
90 lines
3.5 KiB
Python
# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
|
|
# 2024 Alibaba Inc (authors: Xiang Lyu, Zetao Hu)
|
|
#
|
|
# 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 json
|
|
import tensorrt as trt
|
|
import torchaudio
|
|
import logging
|
|
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
|
logging.basicConfig(level=logging.DEBUG,
|
|
format='%(asctime)s %(levelname)s %(message)s')
|
|
|
|
|
|
def read_lists(list_file):
|
|
lists = []
|
|
with open(list_file, 'r', encoding='utf8') as fin:
|
|
for line in fin:
|
|
lists.append(line.strip())
|
|
return lists
|
|
|
|
|
|
def read_json_lists(list_file):
|
|
lists = read_lists(list_file)
|
|
results = {}
|
|
for fn in lists:
|
|
with open(fn, 'r', encoding='utf8') as fin:
|
|
results.update(json.load(fin))
|
|
return results
|
|
|
|
|
|
def load_wav(wav, target_sr):
|
|
speech, sample_rate = torchaudio.load(wav, backend='soundfile')
|
|
speech = speech.mean(dim=0, keepdim=True)
|
|
if sample_rate != target_sr:
|
|
assert sample_rate > target_sr, 'wav sample rate {} must be greater than {}'.format(sample_rate, target_sr)
|
|
speech = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sr)(speech)
|
|
return speech
|
|
|
|
|
|
def convert_onnx_to_trt(trt_model, onnx_model, fp16):
|
|
_min_shape = [(2, 80, 4), (2, 1, 4), (2, 80, 4), (2,), (2, 80), (2, 80, 4)]
|
|
_opt_shape = [(2, 80, 193), (2, 1, 193), (2, 80, 193), (2,), (2, 80), (2, 80, 193)]
|
|
_max_shape = [(2, 80, 6800), (2, 1, 6800), (2, 80, 6800), (2,), (2, 80), (2, 80, 6800)]
|
|
input_names = ["x", "mask", "mu", "t", "spks", "cond"]
|
|
|
|
logging.info("Converting onnx to trt...")
|
|
network_flags = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
|
|
logger = trt.Logger(trt.Logger.INFO)
|
|
builder = trt.Builder(logger)
|
|
network = builder.create_network(network_flags)
|
|
parser = trt.OnnxParser(network, logger)
|
|
config = builder.create_builder_config()
|
|
config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 33) # 8GB
|
|
if fp16:
|
|
config.set_flag(trt.BuilderFlag.FP16)
|
|
profile = builder.create_optimization_profile()
|
|
# load onnx model
|
|
with open(onnx_model, "rb") as f:
|
|
if not parser.parse(f.read()):
|
|
for error in range(parser.num_errors):
|
|
print(parser.get_error(error))
|
|
raise ValueError('failed to parse {}'.format(onnx_model))
|
|
# set input shapes
|
|
for i in range(len(input_names)):
|
|
profile.set_shape(input_names[i], _min_shape[i], _opt_shape[i], _max_shape[i])
|
|
tensor_dtype = trt.DataType.HALF if fp16 else trt.DataType.FLOAT
|
|
# set input and output data type
|
|
for i in range(network.num_inputs):
|
|
input_tensor = network.get_input(i)
|
|
input_tensor.dtype = tensor_dtype
|
|
for i in range(network.num_outputs):
|
|
output_tensor = network.get_output(i)
|
|
output_tensor.dtype = tensor_dtype
|
|
config.add_optimization_profile(profile)
|
|
engine_bytes = builder.build_serialized_network(network, config)
|
|
# save trt engine
|
|
with open(trt_model, "wb") as f:
|
|
f.write(engine_bytes)
|