-
Notifications
You must be signed in to change notification settings - Fork 0
/
infer_debugging.py
172 lines (120 loc) · 5.35 KB
/
infer_debugging.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
import sys
# sys.path.append('../../')
from utils.evaluate import eval
from utils.dataset import get_dataset_ssl_clip
from train.fas import flip_mcl
import numpy as np
from train.config import configC, configM, configI, configO
import time
from timeit import default_timer as timer
import os
import torch
import argparse
from train.params import parse_args
import json
from PIL import Image
import matplotlib.pyplot as plt
import torchvision.transforms as transforms
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = 'cuda'
def overlay(image, attention_map):
attention_map = attention_map[1:, 1:]
attention_map = attention_map.mean(axis=0).reshape(14, 14)
attention_map = attention_map / attention_map.max()
cmap = plt.get_cmap('jet')
attention_map_colored = cmap(attention_map)[:, :, :3]
image = transforms.ToTensor()(image).permute(1, 2, 0).numpy()
attention_map_resized = np.array(Image.fromarray((attention_map_colored * 255).astype(np.uint8)).resize(image.shape[:2], Image.BILINEAR)) / 255
overlay = (0.5 * image + 0.5 * attention_map_resized)
overlay = np.clip(overlay, 0, 1)
return overlay
def save_image(args, path, attention_maps, type='incorrect'):
try:
img = Image.open(path).convert("RGB")
image_filename = os.path.basename(path)
image_base, _ = os.path.splitext(image_filename)
output_dir = os.path.join(args.report_logger_path, args.name, type + "_debugging_images")
os.makedirs(output_dir, exist_ok=True)
image_output_dir = os.path.join(output_dir, image_base)
os.makedirs(image_output_dir, exist_ok=True)
img.save(os.path.join(image_output_dir, f"{image_base}_original.jpg"))
for idx, attention_map in enumerate(attention_maps):
overlay_img = overlay(img, attention_map)
plt.imshow(overlay_img)
overlay_dir = os.path.join(image_output_dir, f"layer_{idx}.jpg")
plt.savefig(overlay_dir, bbox_inches='tight', pad_inches=0)
plt.close()
except Exception as e:
print(f"failed to save attention maps for {path}: {e}")
def infer(args, config):
_, _, _, _, _, _, _, _, _, _, test_dataloader = get_dataset_ssl_clip(args,
config.src1_data, config.src1_train_num_frames,
config.src2_data, config.src2_train_num_frames,
config.src3_data, config.src3_train_num_frames,
config.src4_data, config.src4_train_num_frames,
config.src5_data, config.src5_train_num_frames,
config.tgt_data, config.tgt_test_num_frames)
best_model_ACC = 0.0
best_model_HTER = 1.0
best_model_ACER = 1.0
best_model_AUC = 0.0
best_TPR_FPR = 0.0
valid_args = [np.inf, 0, 0, 0, 0, 0, 0, 0]
net1 = flip_mcl(args, device, in_dim=512, ssl_mlp_dim=4096, ssl_emb_dim=256).to(device)
if config.checkpoint:
ckpt = torch.load(config.checkpoint)
net1.load_state_dict(ckpt['state_dict'])
epoch = ckpt['epoch']
iter_num_start = epoch*100
print(f'Loaded checkpoint from epoch {epoch} at iteration : {iter_num_start}' )
######### eval #########
valid_args = eval(test_dataloader, net1, norm_flag=True, vis=args.vis)
# judge model according to HTER
is_best = valid_args[3] <= best_model_HTER
best_model_HTER = min(valid_args[3], best_model_HTER)
threshold = valid_args[5]
best_model_ACC = valid_args[6]
best_model_AUC = valid_args[4]
best_TPR_FPR = valid_args[7]
return best_model_HTER*100.0, best_model_AUC*100.0, best_TPR_FPR*100.0, valid_args[8], valid_args[9], valid_args[10], valid_args[11]
def main(args):
args = parse_args(args)
with open(os.path.join(os.getcwd(), 'train/model_config/'+args.t_model+'.json'), 'r') as f:
args.t_embed_dim = json.load(f)['embed_dim']
with open(os.path.join(os.getcwd(), 'train/model_config/'+args.model+'.json'), 'r') as f:
args.s_embed_dim = json.load(f)['embed_dim']
# get the name of the experiments
if args.name is None:
args.name = '-'.join([
args.current_time.strftime("%Y_%m_%d-%H_%M_%S"),
f"t_model_{args.t_model}",
f"s_model_{args.model}",
f"lr_{args.lr}",
f"b_{args.batch_size}"
])
log_base_path = os.path.join(args.report_logger_path, args.name)
os.makedirs(log_base_path, exist_ok = True)
# 0-shot / 5-shot
if args.config == 'I':
config = configI
if args.config == 'C':
config = configC
if args.config == 'M':
config = configM
if args.config == 'O':
config = configO
for attr in dir(config):
if attr.find('__') == -1:
print('%s = %r' % (attr, getattr(config, attr)))
with open(os.path.join(args.report_logger_path, args.name, 'out.log'), "w") as f:
f.write('HTER, AUC, TPR@FPR=1%\n')
config.checkpoint = args.ckpt
hter, auc, tpr_fpr, incorrect_list, incorrect_attmaps, correct_list, correct_attmaps = infer(args, config)
f.write(f'{hter},{auc},{tpr_fpr}\n')
for i, image_path in enumerate(incorrect_list):
save_image(args, image_path, incorrect_attmaps[i], type='incorrect')
for i, image_path in enumerate(correct_list):
save_image(args, image_path, correct_attmaps[i], type='correct')
if __name__ == "__main__":
main(sys.argv[1:])