deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 18:09:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

add hifigan train

Browse Source
This commit is contained in:
lyuxiang.lx
2024-10-16 11:37:32 +08:00
parent cb200b21c5
commit 789ee9e5e7
13 changed files with 314 additions and 477 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
cosyvoice/bin/average_model.py Normal file
View File

@@ -0,0 +1,93 @@
# Copyright (c) 2020 Mobvoi Inc (Di Wu)
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# 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 argparse
import glob
import sys
import yaml
import torch
def get_args():
parser = argparse.ArgumentParser(description='average model')
parser.add_argument('--dst_model', required=True, help='averaged model')
parser.add_argument('--src_path',
required=True,
help='src model path for average')
parser.add_argument('--val_best',
action="store_true",
help='averaged model')
parser.add_argument('--num',
default=5,
type=int,
help='nums for averaged model')
args = parser.parse_args()
print(args)
return args
def main():
args = get_args()
val_scores = []
if args.val_best:
yamls = glob.glob('{}/*.yaml'.format(args.src_path))
yamls = [
f for f in yamls
if not (os.path.basename(f).startswith('train')
or os.path.basename(f).startswith('init'))
]
for y in yamls:
with open(y, 'r') as f:
dic_yaml = yaml.load(f, Loader=yaml.BaseLoader)
loss = float(dic_yaml['loss_dict']['loss'])
epoch = int(dic_yaml['epoch'])
step = int(dic_yaml['step'])
tag = dic_yaml['tag']
val_scores += [[epoch, step, loss, tag]]
sorted_val_scores = sorted(val_scores,
key=lambda x: x[2],
reverse=False)
print("best val (epoch, step, loss, tag) = " +
str(sorted_val_scores[:args.num]))
path_list = [
args.src_path + '/epoch_{}_whole.pt'.format(score[0])
for score in sorted_val_scores[:args.num]
]
print(path_list)
avg = {}
num = args.num
assert num == len(path_list)
for path in path_list:
print('Processing {}'.format(path))
states = torch.load(path, map_location=torch.device('cpu'))
for k in states.keys():
if k not in avg.keys():
avg[k] = states[k].clone()
else:
avg[k] += states[k]
# average
for k in avg.keys():
if avg[k] is not None:
# pytorch 1.6 use true_divide instead of /=
avg[k] = torch.true_divide(avg[k], num)
print('Saving to {}'.format(args.dst_model))
torch.save(avg, args.dst_model)
if __name__ == '__main__':
main()

17
cosyvoice/bin/train.py
View File

@@ -86,8 +86,12 @@ def main():
args = get_args()
logging.basicConfig(level=logging.DEBUG,
format='%(asctime)s %(levelname)s %(message)s')
# gan train has some special initialization logic
gan = True if args.model == 'hifigan' else False
override_dict = {k: None for k in ['llm', 'flow', 'hifigan'] if k != args.model}
override_dict = {k: None for k in ['llm', 'flow', 'hift', 'hifigan'] if k != args.model}
if gan is True:
override_dict.pop('hift')
with open(args.config, 'r') as f:
configs = load_hyperpyyaml(f, overrides=override_dict)
configs['train_conf'].update(vars(args))
@@ -97,7 +101,7 @@ def main():
# Get dataset & dataloader
train_dataset, cv_dataset, train_data_loader, cv_data_loader = \
init_dataset_and_dataloader(args, configs)
init_dataset_and_dataloader(args, configs, gan)
# Do some sanity checks and save config to arsg.model_dir
configs = check_modify_and_save_config(args, configs)
@@ -108,13 +112,13 @@ def main():
# load checkpoint
model = configs[args.model]
if args.checkpoint is not None:
model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'))
model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'), strict=False)
# Dispatch model from cpu to gpu
model = wrap_cuda_model(args, model)
# Get optimizer & scheduler
model, optimizer, scheduler = init_optimizer_and_scheduler(args, configs, model)
model, optimizer, scheduler, optimizer_d, scheduler_d = init_optimizer_and_scheduler(args, configs, model, gan)
# Save init checkpoints
info_dict = deepcopy(configs['train_conf'])
@@ -129,7 +133,10 @@ def main():
train_dataset.set_epoch(epoch)
dist.barrier()
group_join = dist.new_group(backend="gloo", timeout=datetime.timedelta(seconds=args.timeout))
executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join)
if gan is True:
executor.train_one_epoc_gan(model, optimizer, scheduler, optimizer_d, scheduler_d, train_data_loader, cv_data_loader, writer, info_dict, group_join)
else:
executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join)
dist.destroy_process_group(group_join)

137
cosyvoice/bin/train_gan.py
View File

@@ -1,137 +0,0 @@
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# 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.
from __future__ import print_function
import argparse
import datetime
import logging
logging.getLogger('matplotlib').setLevel(logging.WARNING)
from copy import deepcopy
import torch
import torch.distributed as dist
import deepspeed
from hyperpyyaml import load_hyperpyyaml
from torch.distributed.elastic.multiprocessing.errors import record
from cosyvoice.utils.executor_gan import Executor
from cosyvoice.utils.train_utils import (
init_distributed,
init_dataset_and_dataloader,
init_optimizer_and_scheduler_gan,
init_summarywriter, save_model,
wrap_cuda_model, check_modify_and_save_config)
def get_args():
parser = argparse.ArgumentParser(description='training your network')
parser.add_argument('--train_engine',
default='torch_ddp',
choices=['torch_ddp', 'deepspeed'],
help='Engine for paralleled training')
parser.add_argument('--model', required=True, help='model which will be trained')
parser.add_argument('--config', required=True, help='config file')
parser.add_argument('--train_data', required=True, help='train data file')
parser.add_argument('--cv_data', required=True, help='cv data file')
parser.add_argument('--checkpoint', help='checkpoint model')
parser.add_argument('--model_dir', required=True, help='save model dir')
parser.add_argument('--tensorboard_dir',
default='tensorboard',
help='tensorboard log dir')
parser.add_argument('--ddp.dist_backend',
dest='dist_backend',
default='nccl',
choices=['nccl', 'gloo'],
help='distributed backend')
parser.add_argument('--num_workers',
default=0,
type=int,
help='num of subprocess workers for reading')
parser.add_argument('--prefetch',
default=100,
type=int,
help='prefetch number')
parser.add_argument('--pin_memory',
action='store_true',
default=False,
help='Use pinned memory buffers used for reading')
parser.add_argument('--deepspeed.save_states',
dest='save_states',
default='model_only',
choices=['model_only', 'model+optimizer'],
help='save model/optimizer states')
parser.add_argument('--timeout',
default=30,
type=int,
help='timeout (in seconds) of cosyvoice_join.')
parser = deepspeed.add_config_arguments(parser)
args = parser.parse_args()
return args
@record
def main():
args = get_args()
logging.basicConfig(level=logging.DEBUG,
format='%(asctime)s %(levelname)s %(message)s')
override_dict = {k: None for k in ['llm', 'flow', 'hifigan'] if k != args.model}
with open(args.config, 'r') as f:
configs = load_hyperpyyaml(f, overrides=override_dict, overrides_must_match=False)
configs['train_conf'].update(vars(args))
# Init env for ddp
init_distributed(args)
# Get dataset & dataloader
train_dataset, cv_dataset, train_data_loader, cv_data_loader = \
init_dataset_and_dataloader(args, configs)
# Do some sanity checks and save config to arsg.model_dir
configs = check_modify_and_save_config(args, configs)
# Tensorboard summary
writer = init_summarywriter(args)
# load checkpoint
model = configs[args.model]
if args.checkpoint is not None:
model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'))
# Dispatch model from cpu to gpu
model = wrap_cuda_model(args, model)
# Get optimizer & scheduler
model, optimizer, scheduler, optimizer_d, scheduler_d = init_optimizer_and_scheduler_gan(args, configs, model)
# Save init checkpoints
info_dict = deepcopy(configs['train_conf'])
save_model(model, 'init', info_dict)
# Get executor
executor = Executor()
# Start training loop
for epoch in range(info_dict['max_epoch']):
executor.epoch = epoch
train_dataset.set_epoch(epoch)
dist.barrier()
group_join = dist.new_group(backend="gloo", timeout=datetime.timedelta(seconds=args.timeout))
executor.train_one_epoc(model, optimizer, scheduler, optimizer_d, scheduler_d, train_data_loader, cv_data_loader, writer, info_dict, group_join)
dist.destroy_process_group(group_join)
if __name__ == '__main__':
main()