File size: 3,607 Bytes
384dc5c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 |
# 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 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):
import tensorrt as trt
_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)
|