-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_read.py
178 lines (150 loc) · 8.64 KB
/
run_read.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
import argparse
import os
os.environ['MPLCONFIGDIR'] = './plt/'
import sys
import torch
basepath = os.path.dirname(os.path.dirname(sys.path[0]))
sys.path.append(basepath)
import dataloader as dataloader
import models
import numpy as np
import warnings
from tqdm import tqdm
from utilities import accuracy, seed_everything
from TTA import READ
# TTA for the cav-mae-finetuned model
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--dataset', type=str, default='vggsound', choices=['vggsound', 'ks50'], help='dataset name')
parser.add_argument("--json-root", type=str, default='/xlearning/mouxing/workspace/TTA/READ/_code_clean/json_csv_files/vgg', help="validation data json")
parser.add_argument("--label-csv", type=str, default='/xlearning/mouxing/workspace/TTA/READ/_code_clean/json_csv_files/class_labels_indices_vgg.csv', help="csv with class labels")
parser.add_argument("--n_class", type=int, default=50, help="number of classes")
parser.add_argument("--model", type=str, default='cav-mae-ft', help="the model used")
parser.add_argument("--dataset_mean", type=float, default=-5.081, help="the dataset mean, used for input normalization")
parser.add_argument("--dataset_std", type=float, default=4.4849, help="the dataset std, used for input normalization")
parser.add_argument("--target_length", type=int, default=1024, help="the input length in frames")
parser.add_argument('--lr', '--learning-rate', default=1e-4, type=float, metavar='LR', help='initial learning rate')
parser.add_argument("--optim", type=str, default="adam", help="training optimizer", choices=["sgd", "adam"])
parser.add_argument('-b', '--batch-size', default=64, type=int, metavar='N', help='mini-batch size')
parser.add_argument('-w', '--num-workers', default=32, type=int, metavar='NW', help='# of workers for dataloading (default: 32)')
# parser.add_argument("--pretrain_path", type=str, default='/xlearning/mouxing/workspace/MM-TTA/egs/kinetics/exp/testmae02-k50-cav-mae-ft-1e-4-2-0.5-1-bs32-ldaFalse-multimodal-fzFalse-h10-a5/models/audio_model_wa.pth', help="pretrained model path")
parser.add_argument("--pretrain_path", type=str, default='/xlearning/mouxing/workspace/TTA/READ/_code_clean/pretrained_model/vgg_65.5.pth', help="pretrained model path")
parser.add_argument("--gpu", type=str, default='7', help="gpu device number")
parser.add_argument("--testmode", type=str, default='multimodal', help="how to test the model")
parser.add_argument('--tta-method', type=str, default='READ', choices=['READ', 'Tent', 'SAR', 'None'], help='which TTA method to be used')
parser.add_argument('--corruption-modality', type=str, default='video', choices=['video', 'audio', 'none'], help='which modality to be corrupted')
# parser.add_argument('--data-val', type=str, default='/xlearning/mouxing/workspace/MM-TTA/audioset-processing/data/ks50_test_json_files/gaussian_noise/severity_5.json', help='path to the validation data json')
parser.add_argument('--severity-start', type=int, default=5, help='the start severity of the corruption')
parser.add_argument('--severity-end', type=int, default=5, help='the end severity of the corruption')
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
print(args)
if args.dataset == 'vggsound':
args.n_class = 309
elif args.dataset == 'ks50':
args.n_class = 50
if args.corruption_modality == 'video':
corruption_list = [
'gaussian_noise',
'shot_noise',
'impulse_noise',
'defocus_blur',
'glass_blur',
'motion_blur',
'zoom_blur',
'snow',
'frost',
'fog',
'brightness',
'contrast',
'elastic_transform',
'pixelate',
'jpeg_compression'
]
elif args.corruption_modality == 'audio':
corruption_list = [
'gaussian_noise',
'traffic',
'crowd',
'rain',
'thunder',
'wind'
]
elif args.corruption_modality == 'none':
corruption_list = ['clean']
args.severity_start = args.severity_end = 0
for corruption in corruption_list:
for severity in range(args.severity_start, args.severity_end+1):
epoch_accs = []
if args.corruption_modality == 'none':
data_val = os.path.join(args.json_root, corruption, 'severity_{}.json').format(severity)
else:
data_val = os.path.join(args.json_root, args.corruption_modality, '{}', 'severity_{}.json').format(corruption, severity)
print('===> Now handling: ', data_val)
for itr in range(1, 6):
seed = int(str(itr)*3)
seed_everything(seed=seed)
print("### Seed= {}, Round {} ###".format(seed, itr))
# all exp in this work is based on 224 * 224 image
im_res = 224
val_audio_conf = {'num_mel_bins': 128, 'target_length': args.target_length, 'freqm': 0, 'timem': 0, 'mixup': 0, 'dataset': args.dataset,
'mode': 'eval', 'mean': args.dataset_mean, 'std': args.dataset_std, 'noise': False, 'im_res': im_res}
tta_loader = torch.utils.data.DataLoader(
dataloader.AudiosetDataset(data_val, label_csv=args.label_csv, audio_conf=val_audio_conf),
batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True, drop_last=False)
if args.model == 'cav-mae-ft':
print('test a cav-mae model with 11 modality-specific layers and 1 modality-sharing layers')
va_model = models.CAVMAEFT(label_dim=args.n_class, modality_specific_depth=11)
else:
raise ValueError('model not supported')
if args.pretrain_path == 'None':
warnings.warn("Note no pre-trained models are specified.")
else:
# TTA based on a CAV-MAE finetuned model
mdl_weight = torch.load(args.pretrain_path)
if not isinstance(va_model, torch.nn.DataParallel):
va_model = torch.nn.DataParallel(va_model)
miss, unexpected = va_model.load_state_dict(mdl_weight, strict=False)
print('now load cav-mae finetuned weights from ', args.pretrain_path)
print(miss, unexpected)
# exit()
# evaluate with multiple frames
if not isinstance(va_model, torch.nn.DataParallel):
va_model = torch.nn.DataParallel(va_model)
va_model.cuda()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#######
print(data_val)
print(f'use TTA or no?# {args.tta_method}')
if args.tta_method == 'None':
adapt_flag = False
else:
adapt_flag = True
if args.tta_method == 'READ' or args.tta_method == 'None':
va_model = READ.configure_model(va_model)
trainables = [p for p in va_model.parameters() if p.requires_grad]
print('Total parameter number is : {:.3f} million'.format(sum(p.numel() for p in va_model.parameters()) / 1e6))
print('Total trainable parameter number is : {:.3f} million'.format(sum(p.numel() for p in trainables) / 1e6))
params, param_names = READ.collect_params(va_model)
optimizer = torch.optim.Adam([{'params': params, 'lr': 1e-4}],
weight_decay=0.,
betas=(0.9, 0.999))
read_model = READ.READ(va_model, optimizer, device, args)
read_model.eval()
overlap = []
with torch.no_grad():
for epoch in range(1):
data_bar = tqdm(tta_loader)
batch_accs = []
for i, (a_input, v_input, labels) in enumerate(data_bar):
a_input = a_input.to(device)
v_input = v_input.to(device)
outputs, loss= read_model((a_input, v_input), adapt_flag=adapt_flag) # now it infers and adapts!
batch_acc = accuracy(outputs[1], labels, topk=(1,))
batch_acc = round(batch_acc[0].item(), 2)
batch_accs.append(batch_acc)
data_bar.set_description(f'Batch#{i}: L0#{loss[0]:.4f}, L1#{loss[1]:.6f}, ACC#{batch_acc:.2f}')
epoch_acc = round(sum(batch_accs) / len(batch_accs), 2)
epoch_accs.append(epoch_acc)
print('Epoch{}: all acc is {}'.format(epoch, epoch_acc))
continue
print('===>{}-{}, mean: {}, std: {}'.format(corruption,severity,np.round(np.mean(epoch_accs), 2),np.round(np.std(epoch_accs), 2)))