-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate_metric_cls_cpu_ViT.py
231 lines (202 loc) · 11.4 KB
/
evaluate_metric_cls_cpu_ViT.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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
#!/usr/bin/env python
# coding: utf-8
import os
import argparse
from pprint import pprint
import torch
import models.group1 as models
import numpy as np
from scipy.stats import weightedtau
import json
import time
from metrics import LEEP, NLEEP, LogME_Score, SFDA_Score, PARC_Score, LogME_optimal, EMMS
from scipy.stats import kendalltau
from scipy.stats import weightedtau
import pprint
import json
from scipy.stats import pearsonr
from w_pearson import wpearson
def save_score(score_dict, fpath):
with open(fpath, "w") as f:
# write dict
json.dump(score_dict, f)
def exist_score(model_name, fpath):
with open(fpath, "r") as f:
result = json.load(f)
if model_name in result.keys():
return True
else:
return False
def recall_k(score, dset, k):
#succed = 0
sorted_score = sorted(score.items(), key=lambda i: i[1], reverse=True)
sorted_score = {a[0]:a[1] for a in sorted_score}
gt = finetune_acc[dset]
sorted_gt = sorted(gt.items(), key=lambda i: i[1], reverse=True)
sorted_gt = {a[0]:a[1] for a in sorted_gt}
top_k_gt = sorted_gt.keys()[:k]
succed = 1 if sorted_score.keys()[0] in top_k_gt else 0
return succed
def rel_k(score, dset, k):
sorted_score = sorted(score.items(), key=lambda i: i[1], reverse=True)
gt = finetune_acc[dset]
sorted_gt = sorted(gt.items(), key=lambda i: i[1], reverse=True)
best_model = sorted_gt[0][0]
sorted_gt = {a[0]:a[1] for a in sorted_gt}
max_gt = sorted_gt[best_model]
topk_score_model = [a[0] for i, a in enumerate(sorted_score) if i < k]
topk_score_ft = [sorted_gt[a] for a in topk_score_model]
return max(topk_score_ft) / max_gt
def pearson_coef(score,dset):
global finetune_acc
score = score.items()
metric_score = [a[1] for a in score]
gt = finetune_acc[dset]
gt_ = []
for a in score:
gt_.append(gt[a[0]])
tw_metric,_ = pearsonr(metric_score,gt_)
return tw_metric
def wpearson_coef(score,dset):
global finetune_acc
score = score.items()
metric_score = [a[1] for a in score]
gt = finetune_acc[dset]
gt_ = []
for a in score:
gt_.append(gt[a[0]])
tw_metric = wpearson(metric_score,gt_)
return tw_metric
def w_kendall_metric(score, dset):
global finetune_acc
score = score.items()
metric_score = [a[1] for a in score]
gt = finetune_acc[dset]
gt_ = []
for a in score:
gt_.append(gt[a[0]])
tw_metric,_ = weightedtau(metric_score,gt_)
return tw_metric
def kendall_metric(score, dset):
global finetune_acc_ssl
score = score.items()
metric_score = [a[1] for a in score]
gt = finetune_acc[dset]
gt_ = []
for a in score:
gt_.append(gt[a[0]])
t_metric,_ = kendalltau(metric_score,gt_)
return t_metric
finetune_acc = {'aircraft': {'deit_tiny': 71.26, 'deit_small': 73.12, 'deit_base': 78.39, 'dino_small': 72.18, 'dino_base': 67.13, 'mocov3_small': 76.04, 'pvtv2_b2': 84.14, 'pvtv2_b3': 84.7, 'pvt_tiny': 69.76, 'pvt_small': 75.2, 'pvt_medium': 76.7, 'swin_t': 81.9, 'swin_s': 83.24},
'caltech101': {'deit_tiny': 89.39, 'deit_small': 92.7, 'deit_base': 93.47, 'dino_small': 86.76, 'dino_base': 92.34, 'mocov3_small': 89.84, 'pvtv2_b2': 93.13, 'pvtv2_b3': 94.4, 'pvt_tiny': 90.04, 'pvt_small': 93.02, 'pvt_medium': 93.75, 'swin_t': 91.9, 'swin_s': 94.0},
'cars': {'deit_tiny': 82.09, 'deit_small': 86.72, 'deit_base': 89.26, 'dino_small': 79.81, 'dino_base': 80.74, 'mocov3_small': 82.18, 'pvtv2_b2': 90.6, 'pvtv2_b3': 91.22, 'pvt_tiny': 84.1, 'pvt_small': 87.61, 'pvt_medium': 87.66, 'swin_t': 88.93, 'swin_s': 89.81},
'cifar10': {'deit_tiny': 96.52, 'deit_small': 97.69, 'deit_base': 98.56, 'dino_small': 97.96, 'dino_base': 98.31, 'mocov3_small': 97.92, 'pvtv2_b2': 97.96, 'pvtv2_b3': 98.44, 'pvt_tiny': 94.87, 'pvt_small': 97.34, 'pvt_medium': 97.93, 'swin_t': 97.34, 'swin_s': 98.06},
'cifar100': {'deit_tiny': 81.58, 'deit_small': 86.62, 'deit_base': 89.96, 'dino_small': 85.66, 'dino_base': 89.38, 'mocov3_small': 85.84, 'pvtv2_b2': 88.24, 'pvtv2_b3': 89.3, 'pvt_tiny': 75.26, 'pvt_small': 86.2, 'pvt_medium': 87.36, 'swin_t': 85.97, 'swin_s': 88.42},
'dtd': {'deit_tiny': 71.86, 'deit_small': 75.08, 'deit_base': 77.66, 'dino_small': 75.96, 'dino_base': 76.01, 'mocov3_small': 71.88, 'pvtv2_b2': 77.16, 'pvtv2_b3': 77.37, 'pvt_tiny': 72.92, 'pvt_small': 75.77, 'pvt_medium': 77.1, 'swin_t': 77.04, 'swin_s': 77.34},
'flowers': {'deit_tiny': 95.5, 'deit_small': 96.79, 'deit_base': 97.98, 'dino_small': 95.96, 'dino_base': 96.28, 'mocov3_small': 93.89, 'pvtv2_b2': 97.89, 'pvtv2_b3': 98.06, 'pvt_tiny': 95.8, 'pvt_small': 97.32, 'pvt_medium': 97.36, 'swin_t': 97.4, 'swin_s': 96.87},
'food': {'deit_tiny': 81.96, 'deit_small': 86.26, 'deit_base': 88.96, 'dino_small': 85.69, 'dino_base': 87.1, 'mocov3_small': 82.84, 'pvtv2_b2': 88.67, 'pvtv2_b3': 89.08, 'pvt_tiny': 83.78, 'pvt_small': 86.98, 'pvt_medium': 85.56, 'swin_t': 86.67, 'swin_s': 87.7},
'pets': {'deit_tiny': 91.44, 'deit_small': 94.02, 'deit_base': 94.61, 'dino_small': 92.59, 'dino_base': 93.41, 'mocov3_small': 90.44, 'pvtv2_b2': 93.86, 'pvtv2_b3': 95.14, 'pvt_tiny': 91.48, 'pvt_small': 94.13, 'pvt_medium': 94.48, 'swin_t': 94.5, 'swin_s': 94.8},
'sun397': {'deit_tiny': 58.4, 'deit_small': 64.76, 'deit_base': 68.62, 'dino_small': 64.14, 'dino_base': 64.78, 'mocov3_small': 60.6, 'pvtv2_b2': 66.44, 'pvtv2_b3': 67.54, 'pvt_tiny': 61.86, 'pvt_small': 65.78, 'pvt_medium': 67.22, 'swin_t': 65.51, 'swin_s': 67.03},
'voc2007': {'deit_tiny': 83.1, 'deit_small': 86.62, 'deit_base': 87.88, 'dino_small': 84.8, 'dino_base': 86.72, 'mocov3_small': 81.84, 'pvtv2_b2': 86.44, 'pvtv2_b3': 88.08, 'pvt_tiny': 84.6, 'pvt_small': 86.62, 'pvt_medium': 87.36, 'swin_t': 87.54, 'swin_s': 88.26}
}
# Main code
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Calculate transferability score.')
parser.add_argument('-m', '--model', type=str, default='deepcluster-v2',
help='name of the pretrained model to load and evaluate (deepcluster-v2 | supervised)')
parser.add_argument('-d', '--dataset', type=str, default='dtd',
help='name of the dataset to evaluate on')
parser.add_argument('-me', '--metric', type=str, default='logme',
help='name of the method for measuring transferability')
parser.add_argument('--nleep-ratio', type=float, default=5,
help='the ratio of the Gaussian components and target data classess')
parser.add_argument('--parc-ratio', type=float, default=2,
help='PCA reduction dimension')
parser.add_argument('--output_dir', type=str, default='./results_metrics/group1',
help='dir of output score')
args = parser.parse_args()
print(args)
score_dict = {}
fpath = os.path.join(args.output_dir, args.metric)
if not os.path.exists(fpath):
os.makedirs(fpath)
fpath = os.path.join(fpath, f'{args.dataset}_metrics_clip.json')
finetune = []
score = []
models_hub = ['deit_base', 'deit_tiny', 'deit_small',
'dino_small', 'mocov3_small',
'pvtv2_b2',
'pvt_tiny', 'pvt_small', 'pvt_medium',
'swin_t'
]
datasets_hub = ['aircraft','caltech101','cars','cifar10','cifar100','dtd','flowers','food','pets','sun397','voc2007']
for dataset in datasets_hub:
start_time = time.time()
args.dataset = dataset
finetune = []
score = []
score_dict = {}
for model in models_hub:
args.model = model
model_npy_feature = os.path.join('/data/results_f/group4_bnorm', f'{args.model}_{args.dataset}_feature.npy')
model_npy_label = os.path.join('/data/results_f/group4_bnorm', f'{args.model}_{args.dataset}_label.npy')
X_features, y_labels = np.load(model_npy_feature), np.load(model_npy_label)
print(y_labels.max())
embedding_npy_label = f'/{args.dataset}_nonorm_bert.npy'
embedding_npy_label2 = f'/{args.dataset}_clip_1024_nonorm.npy'
embedding_npy_label3 = f'/{args.dataset}_gpt2_1024_nonorm.npy'
embedding_label = np.load(embedding_npy_label) #47,512
embedding_label2 = np.load(embedding_npy_label2) #47,512
embedding_label3 = np.load(embedding_npy_label3) #47,512
y_labels_0 = np.zeros([y_labels.shape[0],embedding_label.shape[1]])
y_labels_1 = np.zeros([y_labels.shape[0],embedding_label2.shape[1]])
y_labels_2 = np.zeros([y_labels.shape[0],embedding_label3.shape[1]])
for i in range(y_labels.shape[0]):
y_labels_0[i] = embedding_label[y_labels[i]]
y_labels_1[i] = embedding_label2[y_labels[i]]
y_labels_2[i] = embedding_label3[y_labels[i]]
# y_labels = y_labels2
y_labels1 = np.stack((y_labels_0, y_labels_1, y_labels_2), axis=2)
print(X_features.shape,y_labels1.shape,dataset)
args.metric = 'EMMS'
if args.metric == 'EMMS':
score_dict[args.model] = EMMS(X_features, y_labels1)
elif args.metric == 'logme':
score_dict[args.model] = LogME_Score(X_features, y_labels)
elif args.metric == 'transrate':
score_dict[args.model] = Transrate(X_features, y_labels)
elif args.metric == 'leep':
score_dict[args.model] = LEEP(X_features, y_labels, model_name=args.model)
elif args.metric == 'nleep':
ratio = 1 if args.dataset in ('food', 'pets') else args.nleep_ratio
score_dict[args.model] = NLEEP(X_features, y_labels, component_ratio=ratio)
else:
raise NotImplementedError
finetune.append(finetune_acc[args.dataset][args.model])
score.append(score_dict[args.model])
print(f'{args.metric} of {args.model}: {score_dict[args.model]}\n')
end_time = time.time()
elapsed_time = end_time - start_time
print("Elapsed time: {:.2f} s".format(elapsed_time))
tw_metric, _ = weightedtau(score, finetune)
print(tw_metric,args.dataset)
results = sorted(score_dict.items(), key=lambda i: i[1], reverse=True)
print(f'Models ranking on {args.dataset} based on {args.metric}: ')
print(results)
tw = w_kendall_metric(score_dict, args.dataset)
t = kendall_metric(score_dict, args.dataset)
pear = pearson_coef(score_dict, args.dataset)
wpear = wpearson_coef(score_dict, args.dataset)
rel_3 = rel_k(score_dict, args.dataset, k=3)
rel_1 = rel_k(score_dict, args.dataset, k=1)
# results = {a[0]: a[1] for a in results}
# save_score(results, fpath)
print("Rel@1 dataset:{:12s} our:{:2.3f}".format(args.dataset,rel_1))
print("Rel@3 dataset:{:12s} our:{:2.3f}".format(args.dataset,rel_3))
print("Pearson dataset:{:12s} our:{:2.3f}".format(args.dataset,pear))
print("WPearson dataset:{:12s} our:{:2.3f}".format(args.dataset,wpear))
print("Kendall dataset:{:12s} our:{:2.3f}".format(args.dataset,t))
print("WKendall dataset:{:12s} our:{:2.3f}".format(args.dataset,tw))
print('*'*80)
# results = {a[0]: a[1] for a in results}
# save_score(results, fpath)