-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
executable file
·185 lines (148 loc) · 8.07 KB
/
run.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
import os
import json
import os
import logging
import argparse
import pandas as pd
from collections import Counter
import torch
from torch.utils.data import DataLoader
from models import TextEncoder, AudioEncoder, EarlyFusionEncoderFinetune, LateFusionEncoderFinetune
from data_loader import TextDataset, AudioDataset, MultimodalDataset
from train_validate import train_model, validate_model
from utils import set_seed, NUM_FOLDS, set_logger, get_weights
def run_model(opt):
for key, value in vars(opt).items():
if value in ["True", "False"]:
value = value=="True"
setattr(opt, key, value)
dataset_df = pd.read_csv(opt.data_path, sep="\t")
with open(opt.dataset_fold_path, "r") as f:
data_folds = json.load(f)
dataset_df = dataset_df.dropna()
session_ids = list(set(dataset_df["session"].values.tolist()))
session_ids.sort()
output_dir = os.path.join(opt.out_path, opt.model, opt.output_filename)
os.makedirs(output_dir, exist_ok=True)
set_logger(os.path.join(output_dir, "logging.log"))
for key, value in vars(opt).items():
logging.info("{}, {}".format(key, value))
all_results_df = pd.DataFrame([])
for fold in range(NUM_FOLDS):
opt.fold = fold
log_str = "-"*15 + "split " + str(fold) + "-"*15
logging.info(log_str)
train_ids, val_ids, test_ids = data_folds[str(fold)]
if opt.model == "early_fusion_finetune":
model = EarlyFusionEncoderFinetune(opt)
Dataset = MultimodalDataset
elif opt.model == "late_fusion_finetune":
model = LateFusionEncoderFinetune(opt)
Dataset = MultimodalDataset
elif opt.model == "audio":
model = AudioEncoder(opt)
Dataset = AudioDataset
elif opt.model == "text":
model = TextEncoder(opt)
Dataset = TextDataset
model = model.cuda()
train_data = Dataset(train_ids, dataset_df, opt)
val_data = Dataset(val_ids, dataset_df, opt)
test_data = Dataset(test_ids, dataset_df, opt)
train_labels = train_data.get_labels()
val_labels = val_data.get_labels()
test_labels = test_data.get_labels()
print("train/val/test size", len(train_data), len(val_data), len(test_data))
print("train/val/test session size", len(train_ids), len(val_ids), len(test_ids))
print("Class Dist train:", Counter(train_labels), "val:", Counter(val_labels), "test:", Counter(test_labels))
train_dl = DataLoader(train_data, num_workers=opt.num_worker, batch_size=opt.batch_size, collate_fn=train_data.collate_fn, shuffle=True, pin_memory=True, drop_last=True)
val_dl = DataLoader(val_data, num_workers=opt.num_worker, batch_size=opt.batch_size, collate_fn=val_data.collate_fn, shuffle=False, pin_memory=True)
test_dl = DataLoader(test_data, num_workers=opt.num_worker, batch_size=opt.batch_size, collate_fn=test_data.collate_fn, shuffle=False, pin_memory=True)
weights = get_weights(train_labels)
print("class weights {:.2f}, {:.2f}".format(weights[1], weights[0]))
if opt.model in ["late_fusion_finetune"]:
loss_func = torch.nn.NLLLoss(weight=torch.tensor(weights).cuda())
optimizer = torch.optim.AdamW([
{"params": model.modality_ratio, "lr": 1e-2},
{"params": model.audio_encoder.parameters()},
{"params": model.text_encoder.parameters()}],
lr=opt.learning_rate, weight_decay=opt.weight_decay)
else:
loss_func = torch.nn.CrossEntropyLoss(weight=torch.tensor(weights).cuda())
optimizer = torch.optim.AdamW(model.parameters(), lr=opt.learning_rate, weight_decay=opt.weight_decay)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode="min", patience=opt.patience, verbose=True)
last_best_epoch = -1
best_score = -1
for epoch in range(1, opt.epochs_num+1):
log_str = "Epoch: {}/{}".format(epoch, opt.epochs_num)
logging.info(log_str)
loss, train_score = train_model(model, train_dl, optimizer, loss_func, opt)
val_loss, val_score, _ = validate_model(model, val_dl, loss_func, opt)
scheduler.step(val_loss)
if val_score > best_score:
best_score = val_score
last_best_epoch = epoch
best_state = {"epoch":epoch, "model":model.state_dict(), "opt":opt}
torch.save(best_state, os.path.join(output_dir, "{}_{}.pt".format(opt.model_name, fold)))
log_str = "Train Loss: {:.4f} Train F1: {:.3f} Val F1: {:.3f}".format(loss, train_score, val_score)
logging.info(log_str)
if epoch - last_best_epoch >= opt.patience:
# Early stopping
break
model.load_state_dict(torch.load(os.path.join(output_dir, "{}_{}.pt".format(opt.model_name, fold)))["model"])
_, test_score, result_df = validate_model(model, test_dl, loss_func, opt)
result_df["fold"] = [fold] * result_df.shape[0]
all_results_df = pd.concat([all_results_df, result_df], axis=0)
final_log_str = "Test F1 for fold {}: {:.3f}\n".format(fold, test_score) + "*"*15 + "\n"
logging.info(final_log_str)
if opt.model in ["late_fusion_finetune"]:
log_str = f"Modality ratio: {model.sigmoid(model.modality_ratio)}\n"
logging.info(log_str)
all_results_df.set_index("ID", inplace=True)
all_results_df.sort_index(inplace=True)
all_results_df.to_csv(os.path.join(output_dir, "predictions.csv"), header=True, index=True)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# data
parser.add_argument("--data_root", type=str, default="./data")
parser.add_argument("--data_path", type=str, default="./data/combined_dataset_feats.tsv")
parser.add_argument("--dataset_fold_path", type=str, default="./data/independent_dataset_folds.json")
parser.add_argument("--out_path", type=str, default="./exps_independent")
parser.add_argument("--dataset", type=str, default="combined")
parser.add_argument("--output_filename", type=str)
parser.add_argument("--model_name", type=str)
parser.add_argument("--by_speaker", type=str, default="both")
parser.add_argument("--quartile", type=int, default=-1)
# model
parser.add_argument("--model", type=str, default="text")
parser.add_argument("--context_window", type=int, default=20)
parser.add_argument("--hop", type=int, default=10)
parser.add_argument("--dropout", type=float, default=0.1)
parser.add_argument("--class_num", type=int, default=2)
# text model
parser.add_argument("--text_model", type=str, default="distil-roberta-emotion")
parser.add_argument("--finetune", type=int, default="2")
parser.add_argument("--hidden_dim_1", type=int, default=128)
parser.add_argument("--hidden_dim_2", type=int, default=128)
parser.add_argument("--nlayer", type=int, default=2)
parser.add_argument("--bidirectional", type=str, default="True")
parser.add_argument("--encoder_pooling", type=str, default="cls")
parser.add_argument("--self_attn", type=str, default="True")
parser.add_argument("--attn_pooling", type=str, default="mean_max")
parser.add_argument("--attn_heads", type=int, default=1)
parser.add_argument("--speaker_encoding", type=str, default="True")
# audio base model
parser.add_argument("--audio_feat", type=str, default="hubert")
parser.add_argument("--feat_dim", type=int, default=768)
parser.add_argument("--hidden_dim", type=int, default=128)
# training
parser.add_argument("--num_worker", type=int, default=4)
parser.add_argument("--batch_size", type=int, default=8)
parser.add_argument("--epochs_num", type=int, default=15)
parser.add_argument("--learning_rate", type=float, default=1e-5)
parser.add_argument("--weight_decay", type=float, default=1e-5)
parser.add_argument("--gradient_clip", type=float, default=1.0)
parser.add_argument("--patience", type=int, default=5)
opt = parser.parse_args()
set_seed()
run_model(opt)