-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
170 lines (151 loc) · 7.65 KB
/
train.py
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
"""
# !/usr/bin/env python
-*- coding: utf-8 -*-
@Time : 2022/3/15 下午4:57
@Author : Yang "Jan" Xiao
@Description : train
"""
import logging
from sklearn.metrics import f1_score
from tqdm import tqdm
import torch.nn as nn
from torch.utils.data import DataLoader
from utils.data_loader import *
from utils.utils import *
from torch.utils.data import DataLoader, Subset, random_split
from audiomentations import Compose, AddGaussianNoise, TimeStretch, PitchShift, Gain
logger = logging.getLogger()
def get_dataloader_keyword(data_path, class_list, class_encoding, parameters, noise_aug=False):
"""
CL task protocol: keyword split.
To get the GSC data and build the data loader from a list of keywords.
Args:
data_path: Path to the data root, expected to contain JSON files.
class_list: A list of class names (keywords) to include.
class_encoding: A dictionary mapping class names to numeric labels.
batch_size: Batch size for the dataloaders.
noise_aug: Whether to apply noise augmentation.
Returns:
A tuple of DataLoader instances for training, validation, and testing.
"""
if len(class_list) == 0:
raise ValueError("The class list is empty!")
batch_size = parameters.batch
# Specify JSON file paths
train_json = f"{data_path}/train_manifest.json"
valid_json = f"{data_path}/validation_manifest.json"
test_json = f"{data_path}/test_manifest.json"
# Initialize datasets
train_dataset = SpeechCommandDataset(data_path, train_json, True, class_list, class_encoding, noise_aug=noise_aug)
valid_dataset = SpeechCommandDataset(data_path, valid_json, False, class_list, class_encoding)
test_dataset = SpeechCommandDataset(data_path, test_json, False, class_list, class_encoding)
# Initialize dataloaders
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False, drop_last=True)
test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, drop_last=True)
return train_dataloader, valid_dataloader, test_dataloader
class Trainer:
"""
The KWS model training class.
"""
def __init__(self, opt, model):
self.opt = opt
self.lr = opt.lr
self.step = opt.step
self.epoch = opt.epoch
self.batch = opt.batch
self.model = model
self.device, self.device_list = prepare_device(opt.gpu)
# map the model weight to the device.
self.model.to(self.device)
# enable multi GPU training.
if len(self.device_list) > 1:
print(f">>> Available GPU device: {self.device_list}")
self.model = nn.DataParallel(self.model)
self.best_acc = 0.0
self.best_model = None
self.criterion = nn.CrossEntropyLoss()
self.loss_name = {
"train_loss": 0.0, "train_accuracy": 0.0, "train_total": 0, "train_correct": 0,
"valid_loss": 0.0, "valid_accuracy": 0.0, "valid_total": 0, "valid_correct": 0}
def model_save(self):
save_directory = os.path.join("./model_save", self.opt.save)
if not os.path.isdir(save_directory):
os.makedirs(save_directory)
if self.loss_name["valid_accuracy"] > self.best_acc:
self.best_acc = self.loss_name["valid_accuracy"]
self.best_model = self.model
logger.info(f"Saving the best model with accuracy {self.best_acc:.4f}")
torch.save(self.model.state_dict(), os.path.join(save_directory, f"best.pt"))
if (self.epo + 1) == self.epoch:
torch.save(self.model.state_dict(), os.path.join(save_directory, "last.pt"))
def model_train(self, optimizer, scheduler, train_dataloader, valid_dataloader):
"""
Normal model training process, without modifying the loss function.
"""
train_length, valid_length = len(train_dataloader), len(valid_dataloader)
logger.info(f"[3] Training for {self.epoch} epochs...")
for self.epo in range(self.epoch):
self.loss_name.update({key: 0 for key in self.loss_name})
self.model.cuda(self.device)
self.model.train()
for batch_idx, (waveform, labels) in tqdm(enumerate(train_dataloader), position=0, total=len(train_dataloader)):
waveform, labels = waveform.to(self.device), labels.to(self.device)
"""
waveform:(B,C,16000)
MFCC:(B,C,F,T)
labels:(B,)
"""
# print((waveform.size,labels))
optimizer.zero_grad()
logits = self.model(waveform)
loss = self.criterion(logits, labels)
loss.backward()
optimizer.step()
self.loss_name["train_loss"] += loss.item() / train_length
_, predict = torch.max(logits.data, 1)
self.loss_name["train_total"] += labels.size(0)
self.loss_name["train_correct"] += (predict == labels).sum().item()
self.loss_name["train_accuracy"] = self.loss_name["train_correct"] / self.loss_name["train_total"]
self.model.eval()
for batch_idx, (waveform, labels) in enumerate(valid_dataloader):
with torch.no_grad():
waveform, labels = waveform.to(self.device), labels.to(self.device)
logits = self.model(waveform)
loss = self.criterion(logits, labels)
self.loss_name["valid_loss"] += loss.item() / valid_length
_, predict = torch.max(logits.data, 1)
self.loss_name["valid_total"] += labels.size(0)
self.loss_name["valid_correct"] += (predict == labels).sum().item()
self.loss_name["valid_accuracy"] = self.loss_name["valid_correct"] / self.loss_name["valid_total"]
scheduler.step()
self.model_save()
logger.info(
f"Epoch {self.epo + 1}/{self.epoch} | train_loss {self.loss_name['train_loss']:.4f} "
f"| train_acc {100 * self.loss_name['train_accuracy']:.4f} | "
f"valid_loss {self.loss_name['valid_loss']:.4f} "
f"| valid_accuracy {100 * self.loss_name['valid_accuracy']:.4f} | "
f"lr {optimizer.param_groups[0]['lr']:.4f}"
)
return self.loss_name
def model_test(self, test_dataloader):
self.best_model.eval()
test_length = len(test_dataloader)
self.loss_name.update({key: 0 for key in self.loss_name})
for batch_idx, (waveform, labels) in enumerate(test_dataloader):
with torch.no_grad():
waveform, labels = waveform.to(self.device), labels.to(self.device)
logits = self.best_model(waveform)
loss = self.criterion(logits, labels)
self.loss_name["valid_loss"] += loss.item() / test_length
_, predict = torch.max(logits.data, 1)
self.loss_name["valid_total"] += labels.size(0)
self.loss_name["valid_correct"] += (predict == labels).sum().item()
self.loss_name["valid_accuracy"] = self.loss_name["valid_correct"] / self.loss_name["valid_total"]
self.loss_name["f1_score"] = f1_score(labels.cpu().numpy(), predict.cpu().numpy(), average='macro')
logger.info(
f"test_loss {self.loss_name['valid_loss']:.4f} "
f"| test_acc {self.loss_name['valid_accuracy']:.4f}"
f"| f1_score {self.loss_name['f1_score']:.4f}"
)
return self.loss_name