FacialMMT/main.py at main · NUSTM/FacialMMT · GitHub

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
# -*- encoding: utf-8 -*-

import os
import torch
import argparse
from utils.util import *
from torch.utils.data import DataLoader,RandomSampler, SequentialSampler
from train import Lite

load_project_path = os.path.abspath(os.path.dirname(__file__))

parser = argparse.ArgumentParser(description='A Facial Expression-Aware Multimodal Multi-task Learning Framework for Emotion Recognition in Multi-party Conversations')

#---------------------------------------------------------------------------------------------------------------------------------------------#
'''MELD dataset loading'''
parser.add_argument('--load_anno_csv_path', type=str, default='/media/devin/data/meld/MELD.Raw')
parser.add_argument('--meld_text_path', type=str, default='/media/devin/data/meld/preprocess_data/')
parser.add_argument('--num_labels', type=int, default=7, help='classes number of meld')
parser.add_argument('--data_load_path', type=str, default=os.path.join(load_project_path,'preprocess_data/'),
                    help='path for storing the data')
parser.add_argument('--save_Model_path', default=os.path.join(load_project_path,'saved_model'))
parser.add_argument('--plm_name', type=str, default='roberta-large', choices='[roberta-large, bert-large]')
parser.add_argument('--choice_modality', type=str, default='T+A+V', choices='[T+A+V, V]')

#---------------------------------------------------------------------------------------------------------------------------------------------#
'''Aff-Wild2 dataset loading'''
parser.add_argument('--data_folder', type=str,
                    default='/media/devin/data/aff-wild2/cropped_aligned')
parser.add_argument('--anno_folder', type=str,
                    default='/media/devin/data/aff-wild2/Third_ABAW_Annotations/EXPR_Classification_Challenge_used_FacialMMT/Train_Set/')
parser.add_argument('--data_list_train', type=str,
                    default='/media/devin/data/aff-wild2/preprocess_data/FacialMMT/affwild2_train_img_path_list.txt')

'''Swin Transformer backbone loading'''
parser.add_argument("--backbone_type", type=str, default='SwinTransformer')
parser.add_argument("--backbone_conf_file", type = str, default= os.path.join(load_project_path,'modules/SwinTransformer/swin_conf.yaml'),
                    help = "The path of backbone_conf.yaml.")
parser.add_argument('--pretrained_backbone_path', type=str,
                    default=os.path.join(load_project_path,'pretrained_model','Swin_tiny_Ms-Celeb-1M.pt'))

parser.add_argument('--tau', type=float, default=1, help='temperature parameter, default:1')
parser.add_argument('--FacialEmoImpor_threshold', type=float, default=0.2,
                    help='filter out the emotion-blurred frames')
#---------------------------------------------------------------------------------------------------------------------------------------------#
#tuning
parser.add_argument('--num_epochs', type=int, default=1,
                    help='number of epochs')
parser.add_argument('--aux_lr', type=float, default=5e-5,
                    help='initial learning rate of aux task')
parser.add_argument('--trg_lr', type=float, default=7e-6,
                    help='initial learning rate of trg task')
parser.add_argument('--weight_decay', type=float, default=0.01, help='0.01 for FaialMMT-BERT')
parser.add_argument('--warm_up', type=float, default=0.1, help='dynamic adjust learning rate')

parser.add_argument('--aux_batch_size', type=int, default=150, help='num of images in Aff-Wild2')
parser.add_argument('--trg_batch_size', type=int, default=1, help='num of dialogues in MELD')

parser.add_argument('--aux_accumulation_steps',type=int, default=1,
                    help='gradient accumulation for src task')
parser.add_argument('--trg_accumulation_steps',type=int, default=4,
                    help='gradient accumulation for trg task')
#-------------------------------------------
#multi-modal fusion
parser.add_argument('--crossmodal_layers_TA', type=int, default=2, help='crossmodal layers of text and audio')
parser.add_argument('--crossmodal_num_heads_TA', type=int, default=12)
parser.add_argument('--crossmodal_attn_dropout_TA', type=float, default=0.1, help='dropout applied on the attention weights')

parser.add_argument('--crossmodal_layers_TA_V', type=int, default=2, help='crossmodal layers of text_audio and vision')
parser.add_argument('--crossmodal_num_heads_TA_V', type=int, default=12)
parser.add_argument('--crossmodal_attn_dropout_TA_V', type=float, default=0.1, help='dropout applied on the attention weights')

#---------------------------------------------------------------------------------------------------------------------------------------------#
#self-attention transformer for audio and vision
parser.add_argument('--audio_utt_Transformernum',type=int, default= 5, help='num of self-attention for audio')
parser.add_argument('--vision_utt_Transformernum',type=int, default= 2, help='num of self-attention for vision')

parser.add_argument('--hidden_size', type=int, default=768, help='embedding size in the transformer, 768')
parser.add_argument('--num_attention_heads', type=int, default=12, help='number of heads for the transformer network, 12')
parser.add_argument('--intermediate_size', type=int, default=3072, help='embedding intermediate layer size, 4*hidden_size, 3072')
parser.add_argument('--hidden_act', type=str, default='gelu', help='non-linear activation function')
parser.add_argument('--hidden_dropout_prob',type=float, default=0.1, help='multimodal dropout')
parser.add_argument('--attention_probs_dropout_prob',type=float, default=0.1,help='attention dropout')
parser.add_argument('--layer_norm_eps', type=float, default=1e-12, help='1e-12')
parser.add_argument('--initializer_range',type=int, default=0.02)
#---------------------------------------------------------------------------------------------------------------------------------------------#

parser.add_argument('--clip', type=float, default=0.8,
                    help='gradient clip value (default: 0.8)')
parser.add_argument('--aux_log_interval', type=int, default=1000,
                    help='frequency of result logging')
parser.add_argument('--trg_log_interval', type=int, default=1600,
                    help='frequency of result logging')
parser.add_argument('--seed', type=int, default=1111, help='random seed')

#---------------------------------------------------------------------------------------------------------------------------------------------#
#Evaluate the model on the test set directly
parser.add_argument('--doEval', type=bool, default=True, help='whether to evaluate the model on the test set directly')
parser.add_argument('--load_unimodal_path', type=str, default='unimodal_model_V.pt',
                    help='path to load the best unimodal to evaluate on the test set')
parser.add_argument('--load_multimodal_path', type=str, default= 'multimodal_model_T+A+V_RoBERTa.pt',
                    help='path to load the best multimodal to evaluate on the test set')
parser.add_argument('--load_swin_path', type=str, default= 'best_swin_RoBERTa.pt',
                    help='path to load the best auxiliary model to evaluate on the test set')

args = parser.parse_args()
torch.manual_seed(args.seed)
if torch.cuda.is_available():
    torch.cuda.manual_seed(args.seed)
    torch.set_default_tensor_type('torch.FloatTensor')

if args.choice_modality == 'V':
    trg_train_data = get_meld_vision(args, 'train')
    trg_valid_data = get_meld_vision(args, 'val')
    trg_test_data = get_meld_vision(args, 'test')


elif args.choice_modality == 'T+A+V':
    args.pretrainedtextmodel_path = os.path.join(load_project_path,'pretrained_model',args.plm_name)
    trg_train_data = get_multimodal_data(args, 'train')
    trg_valid_data = get_multimodal_data(args, 'val')
    trg_test_data = get_multimodal_data(args, 'test')
    if not args.doEval:
        #auxiliary dataset
        aux_train_data = get_affwild2_dataset(args)
        aux_train_loader = DataLoader(aux_train_data, sampler=RandomSampler(aux_train_data), batch_size=args.aux_batch_size)
        args.aux_n_train = len(aux_train_data)

trg_train_loader = DataLoader(trg_train_data, sampler=RandomSampler(trg_train_data), batch_size=args.trg_batch_size)
trg_valid_loader = DataLoader(trg_valid_data, sampler=SequentialSampler(trg_valid_data), batch_size=args.trg_batch_size)
trg_test_loader = DataLoader(trg_test_data, sampler=SequentialSampler(trg_test_data), batch_size=args.trg_batch_size)
args.trg_n_train, args.trg_n_valid, args.trg_n_test = len(trg_train_data), len(trg_valid_data), len(trg_test_data)


if args.choice_modality == 'T+A+V':
    args.audio_featExtr_dim = trg_train_data.get_audio_featExtr_dim()

if args.choice_modality in ('V', 'T+A+V'):
    args.vision_featExtr_dim = trg_train_data.get_vision_featExtr_dim()

if args.choice_modality == 'T+A+V':
    args.get_text_utt_max_lens = trg_train_data.get_text_max_utt_len()
    args.get_audio_utt_max_lens = max(trg_train_data.get_audio_max_utt_len(),trg_valid_data.get_audio_max_utt_len(),trg_test_data.get_audio_max_utt_len())

if args.choice_modality in ('V', 'T+A+V'):
    args.get_vision_utt_max_lens = max(trg_train_data.get_vision_max_utt_len(),trg_valid_data.get_vision_max_utt_len(),trg_test_data.get_vision_max_utt_len())

if __name__ == '__main__':
    print('&'*50)
    if args.doEval:
        print('Evaluating on the test set directly...')
        if args.choice_modality == 'V':
            Lite(strategy='dp', devices=1, accelerator="gpu", precision=32).run(args, None, None, None, trg_test_loader)
        elif args.choice_modality == 'T+A+V':
            Lite(strategy='dp', devices=1, accelerator="gpu", precision=16).run(args, None, None, None, trg_test_loader)
    else:
        print('Training from scratch...')
        if args.choice_modality == 'V':
            Lite(strategy='dp', devices=1, accelerator="gpu", precision=32).run(args, None, trg_train_loader, trg_valid_loader, trg_test_loader)
        elif args.choice_modality == 'T+A+V':
            Lite(strategy='dp', devices=1, accelerator="gpu", precision=16).run(args, aux_train_loader, trg_train_loader, trg_valid_loader, trg_test_loader)