Spaces:
Running
on
L4
Running
on
L4
File size: 8,559 Bytes
5b4c852 |
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 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 |
# Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
# 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 logging
from contextlib import nullcontext
import os
import torch
import torch.distributed as dist
from cosyvoice.utils.train_utils import update_parameter_and_lr, log_per_step, log_per_save, batch_forward, batch_backward, save_model, cosyvoice_join
class Executor:
def __init__(self, gan: bool = False):
self.gan = gan
self.step = 0
self.epoch = 0
self.rank = int(os.environ.get('RANK', 0))
self.device = torch.device('cuda:{}'.format(self.rank))
def train_one_epoc(self, model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, scaler, group_join):
''' Train one epoch
'''
lr = optimizer.param_groups[0]['lr']
logging.info('Epoch {} TRAIN info lr {} rank {}'.format(self.epoch, lr, self.rank))
logging.info('using accumulate grad, new batch size is {} times'
' larger than before'.format(info_dict['accum_grad']))
# A context manager to be used in conjunction with an instance of
# torch.nn.parallel.DistributedDataParallel to be able to train
# with uneven inputs across participating processes.
model.train()
model_context = model.join if info_dict['train_engine'] == 'torch_ddp' else nullcontext
with model_context():
for batch_idx, batch_dict in enumerate(train_data_loader):
info_dict["tag"] = "TRAIN"
info_dict["step"] = self.step
info_dict["epoch"] = self.epoch
info_dict["batch_idx"] = batch_idx
if cosyvoice_join(group_join, info_dict):
break
# Disable gradient synchronizations across DDP processes.
# Within this context, gradients will be accumulated on module
# variables, which will later be synchronized.
if info_dict['train_engine'] == 'torch_ddp' and (batch_idx + 1) % info_dict["accum_grad"] != 0:
context = model.no_sync
# Used for single gpu training and DDP gradient synchronization
# processes.
else:
context = nullcontext
with context():
info_dict = batch_forward(model, batch_dict, scaler, info_dict)
info_dict = batch_backward(model, scaler, info_dict)
info_dict = update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict)
log_per_step(writer, info_dict)
# NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and \
(batch_idx + 1) % info_dict["accum_grad"] == 0:
dist.barrier()
self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=False)
model.train()
if (batch_idx + 1) % info_dict["accum_grad"] == 0:
self.step += 1
dist.barrier()
self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=True)
def train_one_epoc_gan(self, model, optimizer, scheduler, optimizer_d, scheduler_d, train_data_loader, cv_data_loader,
writer, info_dict, scaler, group_join):
''' Train one epoch
'''
lr = optimizer.param_groups[0]['lr']
logging.info('Epoch {} TRAIN info lr {} rank {}'.format(self.epoch, lr, self.rank))
logging.info('using accumulate grad, new batch size is {} times'
' larger than before'.format(info_dict['accum_grad']))
# A context manager to be used in conjunction with an instance of
# torch.nn.parallel.DistributedDataParallel to be able to train
# with uneven inputs across participating processes.
model.train()
model_context = model.join if info_dict['train_engine'] == 'torch_ddp' else nullcontext
with model_context():
for batch_idx, batch_dict in enumerate(train_data_loader):
info_dict["tag"] = "TRAIN"
info_dict["step"] = self.step
info_dict["epoch"] = self.epoch
info_dict["batch_idx"] = batch_idx
if cosyvoice_join(group_join, info_dict):
break
# Disable gradient synchronizations across DDP processes.
# Within this context, gradients will be accumulated on module
# variables, which will later be synchronized.
if info_dict['train_engine'] == 'torch_ddp' and (batch_idx + 1) % info_dict["accum_grad"] != 0:
context = model.no_sync
# Used for single gpu training and DDP gradient synchronization
# processes.
else:
context = nullcontext
with context():
batch_dict['turn'] = 'discriminator'
info_dict = batch_forward(model, batch_dict, scaler, info_dict)
info_dict = batch_backward(model, scaler, info_dict)
info_dict = update_parameter_and_lr(model, optimizer_d, scheduler_d, scaler, info_dict)
optimizer.zero_grad()
log_per_step(writer, info_dict)
with context():
batch_dict['turn'] = 'generator'
info_dict = batch_forward(model, batch_dict, scaler, info_dict)
info_dict = batch_backward(model, scaler, info_dict)
info_dict = update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict)
optimizer_d.zero_grad()
log_per_step(writer, info_dict)
# NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and \
(batch_idx + 1) % info_dict["accum_grad"] == 0:
dist.barrier()
self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=False)
model.train()
if (batch_idx + 1) % info_dict["accum_grad"] == 0:
self.step += 1
dist.barrier()
self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=True)
@torch.inference_mode()
def cv(self, model, cv_data_loader, writer, info_dict, on_batch_end=True):
''' Cross validation on
'''
logging.info('Epoch {} Step {} on_batch_end {} CV rank {}'.format(self.epoch, self.step + 1, on_batch_end, self.rank))
model.eval()
total_num_utts, total_loss_dict = 0, {} # avoid division by 0
for batch_idx, batch_dict in enumerate(cv_data_loader):
info_dict["tag"] = "CV"
info_dict["step"] = self.step
info_dict["epoch"] = self.epoch
info_dict["batch_idx"] = batch_idx
num_utts = len(batch_dict["utts"])
total_num_utts += num_utts
if self.gan is True:
batch_dict['turn'] = 'generator'
info_dict = batch_forward(model, batch_dict, None, info_dict)
for k, v in info_dict['loss_dict'].items():
if k not in total_loss_dict:
total_loss_dict[k] = []
total_loss_dict[k].append(v.item() * num_utts)
log_per_step(None, info_dict)
for k, v in total_loss_dict.items():
total_loss_dict[k] = sum(v) / total_num_utts
info_dict['loss_dict'] = total_loss_dict
log_per_save(writer, info_dict)
model_name = 'epoch_{}_whole'.format(self.epoch) if on_batch_end else 'epoch_{}_step_{}'.format(self.epoch, self.step + 1)
save_model(model, model_name, info_dict)
|