-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
232 lines (205 loc) · 7.42 KB
/
main.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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
import os
import argparse
import wandb
import torch
from torch import nn
from torch.utils.data import DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR
from torchvision import transforms
from torch.nn import DataParallel
from x_transformers import TransformerWrapper, Encoder, Decoder
import warnings
from dataset.dataset import RDropDataset
from dataset.dataset_finetune import RDropDataset_finetune
from pretrain import train, validate
from finetune import train_finetune, validate_finetune
from utils import *
from models import *
def main():
CONSTANTS = load_config("CONSTANTS")
parser = argparse.ArgumentParser(
description="Yonsei Final Project : Tokenized Lip Reading"
)
parser.add_argument("--run", type=str, choices=["pretrain", "finetune"])
running_args = parser.parse_args()
# path definition
ROOT_DIR = CONSTANTS["ROOT_DIR"]
LABEL_DIR = CONSTANTS["LABEL_DIR"]
LOAD_PATH = CONSTANTS["LOAD_PATH"]
SAVE_GPT_PATH = CONSTANTS["SAVE_GPT_PATH"]
SAVE_PATH = CONSTANTS["SAVE_PATH"]
if not os.path.exists(SAVE_PATH):
os.makedirs(SAVE_PATH)
# constant definition
SEED = CONSTANTS["SEED"]
NUM_CLASS = CONSTANTS["NUM_CLASS"]
CHANNEL = CONSTANTS["CHANNEL"]
IMG_RES = CONSTANTS["IMG_RES"]
# hyperparameter definition
seed_everything(SEED)
HYPERPARAMS = load_config("MLP_HYPERPARAMS") # load hyperparams
wandb.init(project=HYPERPARAMS.wandb_project_name, entity=CONSTANTS.WANDB_USER)
wandb.config.update(HYPERPARAMS) # add hyperparams to wandb
print("current parameters: ", HYPERPARAMS)
# image transformation
mean = [0.5037278, 0.503253, -7.063131e-05]
std = [0.11415035, 0.13153046, 0.07024033]
vanilla_transform = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std),
]
)
if running_args == "pretrain":
train_dataset = RDropDataset(
label_dir=LABEL_DIR, split="train", transforms=True
)
train_dataset.set_transform(vanilla_transform)
valid_dataset = RDropDataset(label_dir=LABEL_DIR, split="val", transforms=True)
valid_dataset.set_transform(vanilla_transform)
test_dataset = RDropDataset(label_dir=LABEL_DIR, split="test", transforms=True)
test_dataset.set_transform(vanilla_transform)
elif running_args == "finetune":
train_dataset = RDropDataset_finetune(
root_dir=ROOT_DIR, label_dir=LABEL_DIR, split="train", transforms=True
)
train_dataset.set_transform(vanilla_transform)
valid_dataset = RDropDataset_finetune(
root_dir=ROOT_DIR, label_dir=LABEL_DIR, split="val", transforms=True
)
valid_dataset.set_transform(vanilla_transform)
test_dataset = RDropDataset_finetune(
root_dir=ROOT_DIR, label_dir=LABEL_DIR, split="test", transforms=True
)
test_dataset.set_transform(vanilla_transform)
else:
raise ValueError("Invalid run argument: run with --run [pretrain|finetune]")
device, NUM_WORKERS = check_device()
train_loader = DataLoader(
train_dataset,
batch_size=HYPERPARAMS.batch_size,
shuffle=True,
num_workers=NUM_WORKERS,
drop_last=False,
)
valid_loader = DataLoader(
valid_dataset,
batch_size=HYPERPARAMS.batch_size,
shuffle=False,
num_workers=NUM_WORKERS,
drop_last=False,
)
test_loader = DataLoader(
test_dataset,
batch_size=HYPERPARAMS.batch_size,
shuffle=False,
num_workers=NUM_WORKERS,
drop_last=False,
)
model = VideoImageModel(
num_tokens=NUM_CLASS,
max_seq_len=58,
attn_layers=Encoder(
dim=HYPERPARAMS.working_dim,
depth=HYPERPARAMS.num_layers,
heads=HYPERPARAMS.heads,
layer_dropout=HYPERPARAMS.layer_dropout,
ff_dropout=HYPERPARAMS.ff_dropout, # Let's set this as 0.3 or higher
use_rmsnorm=True,
ff_glu=True,
rotary_pos_emb=True,
),
channels=CHANNEL,
image_size=IMG_RES,
patch_size=HYPERPARAMS.patch_size,
dim=HYPERPARAMS.working_dim,
post_emb_norm=False,
emb_dropout=HYPERPARAMS.emp_dropout, # Let's set this as 0.15 (lower than 0.25)
)
model = DataParallel(model)
model.to(device)
if running_args == "pretrain":
pass
elif running_args == "finetune":
model.load_state_dict(torch.load(LOAD_PATH))
del model.module.decoder
del model.module.gpt
clf_model = VideoImageModelForClassification(
num_tokens=500,
max_seq_len=58,
attn_layers=model.module.attn_layers,
channels=CHANNEL,
image_size=IMG_RES,
patch_size=HYPERPARAMS.patch_size,
dim=HYPERPARAMS.working_dim,
post_emb_norm=False,
emb_dropout=HYPERPARAMS.emp_dropout, # Let's set this as 0.15 (lower than 0.25)
num_classes=NUM_CLASS,
)
clf_model = DataParallel(clf_model)
clf_model.to(device)
else:
raise ValueError("Invalid run argument: run with --run [pretrain|finetune]")
if running_args == "pretrain":
criterion = nn.CrossEntropyLoss()
elif running_args == "finetune":
# labelsmoothing loss
criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
optimizer = AdamW(
model.parameters(),
lr=HYPERPARAMS.learning_rate,
weight_decay=HYPERPARAMS.weight_decay,
)
scheduler = CosineAnnealingLR(optimizer, T_max=HYPERPARAMS.num_epochs, verbose=True)
if running_args == "pretrain":
best_model = train(
CONFIG=HYPERPARAMS,
save_path=SAVE_GPT_PATH,
model=model,
train_loader=train_loader,
valid_loader=valid_loader,
optimizer=optimizer,
scheduler=scheduler,
criterion=criterion,
device=device,
)
# run test
test_loss, _, _ = validate(
HYPERPARAMS, test_loader, best_model, criterion, device
)
print(f"Test loss: {test_loss:4.2}")
wandb.log(
{
"test/loss": test_loss,
}
)
elif running_args == "finetune":
best_clf_model = train_finetune(
CONFIG=HYPERPARAMS,
save_path=SAVE_PATH,
model=clf_model,
train_loader=train_loader,
valid_loader=valid_loader,
optimizer=optimizer,
scheduler=scheduler,
criterion=criterion,
device=device,
)
# run test
test_loss, test_acc1, test_acc5 = validate_finetune(
HYPERPARAMS, test_loader, best_clf_model, criterion, device
)
print(
f"Test loss: {test_loss:4.2}, test acc1: {test_acc1:4.2}, test acc5: {test_acc5:4.2}"
)
wandb.log(
{
"test/loss": test_loss,
"test/top5_accuracy": test_acc5,
"test/top1_accuracy": test_acc1,
}
)
wandb.finish()
if __name__ == "__main__":
main()