-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
133 lines (112 loc) · 6.37 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
import argparse
from matchvoice_dataset import MatchVoice_Dataset
from models.matchvoice_model import matchvoice_model
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import AdamW
import torch
import numpy as np
import random
import os
from pycocoevalcap.cider.cider import Cider
# Use CIDEr score to do validation
def eval_cider(predicted_captions, gt_captions):
cider_evaluator = Cider()
predicted_captions_dict =dict()
gt_captions_dict = dict()
for i, caption in enumerate(predicted_captions):
predicted_captions_dict[i] = [caption]
for i, caption in enumerate(gt_captions):
gt_captions_dict[i] = [caption]
_, cider_scores = cider_evaluator.compute_score(predicted_captions_dict, gt_captions_dict)
return cider_scores.tolist()
def train(args):
train_dataset = MatchVoice_Dataset(feature_root=args.feature_root, ann_root=args.train_ann_root,
window=args.window, fps=args.fps, tokenizer_name=args.tokenizer_name, timestamp_key=args.train_timestamp_key)
val_dataset = MatchVoice_Dataset(feature_root=args.feature_root, ann_root=args.val_ann_root,
window=args.window, fps=args.fps, tokenizer_name=args.tokenizer_name, timestamp_key=args.val_timestamp_key)
train_data_loader = DataLoader(train_dataset, batch_size=args.train_batch_size, num_workers=args.train_num_workers, drop_last=False, shuffle=True, pin_memory=True, collate_fn=train_dataset.collater)
val_data_loader = DataLoader(val_dataset, batch_size=args.val_batch_size, num_workers=args.val_num_workers, drop_last=True, shuffle=True, pin_memory=True, collate_fn=train_dataset.collater)
print("===== Video features data loaded! =====")
model = matchvoice_model(llm_ckpt=args.tokenizer_name, tokenizer_ckpt=args.tokenizer_name ,window=args.window, num_query_tokens=args.num_query_tokens, num_video_query_token=args.num_video_query_token, num_features=args.num_features, device=args.device).to(args.device)
if args.continue_train:
model.load_state_dict(torch.load(args.load_ckpt))
optimizer = AdamW(model.parameters(), lr=args.lr)
os.makedirs(args.model_output_dir, exist_ok=True)
print("===== Model and Checkpoints loaded! =====")
max_val_CIDEr = max(float(0), args.pre_max_CIDEr)
for epoch in range(args.pre_epoch, args.num_epoch):
model.train()
train_loss_accum = 0.0
train_pbar = tqdm(train_data_loader, desc=f'Epoch {epoch+1}/{args.num_epoch} Training')
for samples in train_pbar:
optimizer.zero_grad()
try:
loss = model(samples)
loss.backward()
optimizer.step()
train_loss_accum += loss.item()
train_pbar.set_postfix({"Loss": f"{loss.item():.4f}"})
avg_train_loss = train_loss_accum / len(train_data_loader)
except:
pass
model.eval()
val_CIDEr = 0.0
val_pbar = tqdm(val_data_loader, desc=f'Epoch {epoch+1}/{args.num_epoch} Validation')
with torch.no_grad():
for samples in val_pbar:
temp_res_text, anonymized = model(samples, True)
cur_CIDEr_score = eval_cider(temp_res_text,anonymized)
val_CIDEr += sum(cur_CIDEr_score)/len(cur_CIDEr_score)
val_pbar.set_postfix({"Scores": f"|C:{sum(cur_CIDEr_score)/len(cur_CIDEr_score):.4f}"})
avg_val_CIDEr = val_CIDEr / len(val_data_loader)
print(f"Epoch {epoch+1} Summary: Average Training Loss: {avg_train_loss:.3f}, Average Validation scores: C:{avg_val_CIDEr*100:.3f}")
if epoch % 5 == 0:
file_path = f"{args.model_output_dir}/model_save_{epoch+1}.pth"
save_matchvoice_model(model, file_path)
if avg_val_CIDEr > max_val_CIDEr:
max_val_CIDEr = avg_val_CIDEr
file_path = f"{args.model_output_dir}/model_save_best_val_CIDEr.pth"
save_matchvoice_model(model, file_path)
def save_matchvoice_model(model, file_path):
state_dict = model.cpu().state_dict()
state_dict_without_llama = {}
# 遍历原始模型的 state_dict，并排除 llama_model 相关的权重
for key, value in state_dict.items():
if "llama_model.model.layers" not in key:
state_dict_without_llama[key] = value
torch.save(state_dict_without_llama, file_path)
model.to(model.device)
if __name__ == "__main__":
torch.manual_seed(42)
np.random.seed(42)
random.seed(42)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(42)
parser = argparse.ArgumentParser(description="Train a model with FRANZ dataset.")
parser.add_argument("--feature_root", type=str, default="./features/baidu_soccer_embeddings")
parser.add_argument("--window", type=float, default=15)
parser.add_argument("--tokenizer_name", type=str, default="meta-llama/Meta-Llama-3-8B")
parser.add_argument("--train_ann_root", type=str, default="./dataset/MatchTime/train")
parser.add_argument("--train_batch_size", type=int, default=32)
parser.add_argument("--train_num_workers", type=int, default=32)
parser.add_argument("--train_timestamp_key", type=str, default="gameTime")
parser.add_argument("--val_ann_root", type=str, default="./dataset/MatchTime/valid")
parser.add_argument("--val_batch_size", type=int, default=20)
parser.add_argument("--val_num_workers", type=int, default=32)
parser.add_argument("--val_timestamp_key", type=str, default="gameTime")
parser.add_argument("--lr", type=float, default=1e-4)
parser.add_argument("--num_epoch", type=int, default=80)
parser.add_argument("--num_query_tokens", type=int, default=32)
parser.add_argument("--num_video_query_token", type=int, default=32)
parser.add_argument("--num_features", type=int, default=512)
parser.add_argument("--fps", type=int, default=2)
parser.add_argument("--model_output_dir", type=str, default="./ckpt")
parser.add_argument("--device", type=str, default="cuda:0")
# If continue training from any epoch
parser.add_argument("--continue_train", type=bool, default=False)
parser.add_argument("--pre_max_CIDEr", type=float, default=0.0)
parser.add_argument("--pre_epoch", type=int, default=0)
parser.add_argument("--load_ckpt", type=str, default="./ckpt/model_save_best_val_CIDEr.pth")
args = parser.parse_args()
train(args)