-
Notifications
You must be signed in to change notification settings - Fork 0
/
metric.py
350 lines (273 loc) · 13.3 KB
/
metric.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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
import numpy as np
from numba import njit
from sklearn.covariance import LedoitWolf
from sklearn.random_projection import GaussianRandomProjection, SparseRandomProjection
import torch
import pandas as pd
import itertools
def random_projection( method = 'gaussian', n_components = 'auto'):
if method == 'gaussian':
rp = GaussianRandomProjection(n_components)
elif method =='sparse':
rp = SparseRandomProjection(n_components)
return rp
def log_maximum_evidence(features: np.ndarray, targets: np.ndarray, regression=False, return_weights=False):
r"""
Log Maximum Evidence in `LogME: Practical Assessment of Pre-trained Models
for Transfer Learning (ICML 2021) <https://arxiv.org/pdf/2102.11005.pdf>`_.
Args:
features (np.ndarray): feature matrix from pre-trained model.
targets (np.ndarray): targets labels/values.
regression (bool, optional): whether to apply in regression setting. (Default: False)
return_weights (bool, optional): whether to return bayesian weight. (Default: False)
Shape:
- features: (N, F) with element in [0, :math:`C_t`) and feature dimension F, where :math:`C_t` denotes the number of target class
- targets: (N, ) or (N, C), with C regression-labels.
- weights: (F, :math:`C_t`).
- score: scalar.
"""
f = features.astype(np.float64)
y = targets
if regression:
y = targets.astype(np.float64)
fh = f
f = f.transpose()
D, N = f.shape
v, s, vh = np.linalg.svd(f @ fh, full_matrices=True)
evidences = []
weights = []
if regression:
C = y.shape[1]
for i in range(C):
y_ = y[:, i]
evidence, weight = each_evidence(y_, f, fh, v, s, vh, N, D)
evidences.append(evidence)
weights.append(weight)
else:
C = int(y.max() + 1)
for i in range(C):
y_ = (y == i).astype(np.float64)
evidence, weight = each_evidence(y_, f, fh, v, s, vh, N, D)
evidences.append(evidence)
weights.append(weight)
score = np.mean(evidences)
weights = np.vstack(weights)
if return_weights:
return score, weights
else:
return score
@njit
def each_evidence(y_, f, fh, v, s, vh, N, D):
"""
compute the maximum evidence for each class
"""
alpha = 1.0
beta = 1.0
lam = alpha / beta
tmp = (vh @ (f @ y_))
for _ in range(11):
# should converge after at most 10 steps
# typically converge after two or three steps
gamma = (s / (s + lam)).sum()
m = v @ (tmp * beta / (alpha + beta * s))
alpha_de = (m * m).sum()
alpha = gamma / alpha_de
beta_de = ((y_ - fh @ m) ** 2).sum()
beta = (N - gamma) / beta_de
new_lam = alpha / beta
if np.abs(new_lam - lam) / lam < 0.01:
break
lam = new_lam
evidence = D / 2.0 * np.log(alpha) \
+ N / 2.0 * np.log(beta) \
- 0.5 * np.sum(np.log(alpha + beta * s)) \
- beta / 2.0 * beta_de \
- alpha / 2.0 * alpha_de \
- N / 2.0 * np.log(2 * np.pi)
return evidence / N, m
def h_score(features: np.ndarray, labels: np.ndarray):
r"""
H-score in `An Information-theoretic Approach to Transferability in Task Transfer Learning (ICIP 2019)
<http://yangli-feasibility.com/home/media/icip-19.pdf>`_.
The H-Score :math:`\mathcal{H}` can be described as:
.. math::
\mathcal{H}=\operatorname{tr}\left(\operatorname{cov}(f)^{-1} \operatorname{cov}\left(\mathbb{E}[f \mid y]\right)\right)
where :math:`f` is the features extracted by the model to be ranked, :math:`y` is the groud-truth label vector
Args:
features (np.ndarray):features extracted by pre-trained model.
labels (np.ndarray): groud-truth labels.
Shape:
- features: (N, F), with number of samples N and feature dimension F.
- labels: (N, ) elements in [0, :math:`C_t`), with target class number :math:`C_t`.
- score: scalar.
"""
f = features
y = labels
covf = np.cov(f, rowvar = False)
C = int(y.max() + 1)
g = np.zeros_like(f)
for i in range(C):
Ef_i = np.mean(f[y == i, :], axis=0)
g[y == i] = Ef_i
covg = np.cov(g, rowvar = False)
score = np.trace(np.dot(np.linalg.pinv(covf, rcond=1e-15), covg))
return score
def regularized_h_score(features: np.ndarray, labels: np.ndarray):
f = features.astype('float64')
f = f - np.mean(f, axis= 0, keepdims= True) #Center the features for correct Ledoit-Wolf Estimation
y = labels
C = int(y.max() + 1)
g = np.zeros_like(f)
cov = LedoitWolf(assume_centered= False).fit(f)
alpha = cov.shrinkage_
covf_alpha = cov.covariance_
for i in range(C):
Ef_i = np.mean(f[y == i, :], axis=0)
g[y == i] = Ef_i
covg = np.cov(g, rowvar = False)
score = np.trace(np.dot(np.linalg.pinv(covf_alpha, rcond=1e-15), (1-alpha) * covg))
return score
def coding_rate(features : np.ndarray, eps = 1e-4):
f = features
n, d = f.shape
(_, rate) = np.linalg.slogdet((np.eye(d) + 1 / (n * eps) * f.transpose() @ f))
return 0.5 * rate
def transrate(features : np.ndarray, labels : np.ndarray, eps = 1e-4):
f = features
y = labels
f = f - np.mean(f, axis = 0, keepdims = True)
Rf = coding_rate(f, eps)
Rfy = 0.0
C = int(y.max() + 1)
for i in range(C):
Rfy += coding_rate(f[(y==i).flatten()], eps)
return Rf - Rfy / C
def compute_metric_per_layer(dataset = 'BCCD', metric = 'label_LogME', rp = None):
feats_metric, bbox_metric = [], []
annots= torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset}/labels/all_labels.pt")
labels = annots[:,-1].cpu().int().numpy() - 1 # Extract only labels, convert to int32 and remove background class (0 class)
if dataset == 'Open_Images':
labels[labels==12] = 10
labels[labels==11] = 2 #Label 10 and 2 are not there so we replace label 10 by label 12 #dirty patch up to modify !
for i in range(1,6):
feats = torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset}/layer_{i}/all_features.pt").cpu().numpy()
feats_bbox = torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset}/layer_{i}/all_features_bbox.pt").cpu().numpy()
#Random projection
if rp is not None :
feats = rp.fit_transform(feats)
feats_bbox = rp.fit_transform(feats)
if metric == 'label_LogME':
logme = log_maximum_evidence(feats, labels)
feats_metric.append(logme)
logme = log_maximum_evidence(feats_bbox, labels)
bbox_metric.append(logme)
if metric == 'xy_LogME':
logme_xy = log_maximum_evidence(feats,annots[:,0:4].int().numpy(), regression = True)
feats_metric.append(logme_xy)
logme = log_maximum_evidence(feats_bbox,annots[:,0:4].int().numpy(), regression = True)
bbox_metric.append(logme_xy)
elif metric == 'hscore':
hscore = h_score(feats, labels)
feats_metric.append(hscore)
hscore = h_score(feats_bbox, labels)
bbox_metric.append(hscore)
elif metric == 'regularized_hscore':
hscore = regularized_h_score(feats, labels)
feats_metric.append(hscore)
hscore = regularized_h_score(feats_bbox, labels)
bbox_metric.append(hscore)
elif metric == 'transrate':
tr = transrate(feats, labels)
feats_metric.append(tr)
tr = transrate(feats_bbox, labels)
bbox_metric.append(tr)
return feats_metric, bbox_metric
def compute_metric_per_dataset(metric = 'LogME', rp = None):
metric_bccd, bbox_metric_bccd = compute_metric_per_layer(dataset = 'BCCD', metric = metric, rp = rp)
metric_chess, bbox_metric_chess = compute_metric_per_layer(dataset = 'CHESS', metric = metric, rp = rp)
metric_gw, bbox_metric_gw = compute_metric_per_layer(dataset = 'Global_Wheat', metric = metric, rp = rp)
metric_voc, bbox_metric_voc = compute_metric_per_layer(dataset = 'VOC', metric = metric, rp = rp)
metric_oi, bbox_metric_oi = compute_metric_per_layer(dataset = 'Open_Images', metric = metric, rp = rp)
metric_df = np.vstack((metric_bccd, metric_chess, metric_gw, metric_voc, metric_oi))
metric_df = pd.DataFrame(metric_df, index = ['BCCD', 'CHESS', 'Global_Wheat', 'VOC', 'Open_Images'])
bbox_metric_df = np.vstack((bbox_metric_bccd, bbox_metric_chess, bbox_metric_gw, bbox_metric_voc, bbox_metric_oi))
bbox_metric_df = pd.DataFrame(bbox_metric_df, index = ['BCCD', 'CHESS', 'Global_Wheat', 'VOC', 'Open_Images'])
return metric_df, bbox_metric_df
def compute_metric_synth(metric = 'label_LogME', rp = None, layer = 'layer_5', device = 'cpu'):
datasets = ['MNIST', 'KMNIST', 'EMNIST', 'FASHION_MNIST', 'USPS']
df_metric = pd.DataFrame(np.zeros((5,5)), index = datasets, columns = datasets)
bbox_metric_df = pd.DataFrame(np.zeros((5,5)), index = datasets, columns = datasets)
for dataset_source, dataset_target in itertools.permutations(datasets, 2):
annots= torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset_target}/from_{dataset_source}/labels/all_labels.pt").to(device)
labels = annots[:,-1].to(device).int().numpy() - 1
feats = torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset_target}/from_{dataset_source}/{layer}/all_features.pt").to(device).numpy()
feats_bbox = torch.load(f"/data.nfs/AUTO_TL_OD/extracted_feats/{dataset_target}/from_{dataset_source}/{layer}/all_features_bbox.pt").to(device).numpy()
if rp is not None:
feats = rp.fit_transform(feats)
feats_bbox = rp.fit_transform(feats)
elif metric == 'label_LogME':
logme = log_maximum_evidence(feats, labels)
df_metric.loc[dataset_source, dataset_target] = logme
logme = log_maximum_evidence(feats_bbox, labels)
bbox_metric_df.loc[dataset_source, dataset_target] = logme
elif metric == 'xy_LogME':
logme = log_maximum_evidence(feats,annots[:,0:4].int().numpy(), regression = True)
df_metric.loc[dataset_source, dataset_target] = logme
logme = log_maximum_evidence(feats_bbox,annots[:,0:4].int().numpy(), regression = True)
bbox_metric_df.loc[dataset_source, dataset_target] = logme
elif metric == 'hscore':
hscore = h_score(feats, labels)
df_metric.loc[dataset_source, dataset_target] = hscore
hscore = h_score(feats_bbox, labels)
bbox_metric_df.loc[dataset_source, dataset_target] = hscore
elif metric == 'regularized_hscore':
hscore = regularized_h_score(feats, labels)
df_metric.loc[dataset_source, dataset_target] = hscore
hscore = regularized_h_score(feats_bbox, labels)
bbox_metric_df.loc[dataset_source, dataset_target] = hscore
elif metric == 'transrate':
tr = transrate(feats, labels)
df_metric.loc[dataset_source, dataset_target] = tr
tr = transrate(feats_bbox, labels)
bbox_metric_df.loc[dataset_source, dataset_target] = tr
return df_metric, bbox_metric_df
def compute_metric_real(metric = 'label_LogME', rp = None, dataset_paths = None, data_dir = "/data.nfs/AUTO_TL_OD/extracted_feats/", layer = 'layer_5'):
metrics = []
bbox_metrics = []
for dataset_path in dataset_paths:
annots= torch.load(data_dir + f"{dataset_path}/labels/all_labels.pt").cpu()
labels = annots[:,-1].cpu().int().numpy() - 1
feats = torch.load(data_dir + f"{dataset_path}/{layer}/all_features.pt").cpu().numpy()
feats_bbox = torch.load(data_dir + f"{dataset_path}/{layer}/all_features_bbox.pt").cpu().numpy()
if dataset_path == 'Open_Images':
labels[labels==12] = 10
labels[labels==11] = 2 #Label 10 and 2 are not there so we replace label 10 by label 12 #dirty patch up to modify !
if rp is not None:
feats = rp.fit_transform(feats)
feats_bbox = rp.fit_transform(feats)
if metric == 'label_LogME':
logme = log_maximum_evidence(feats, labels)
metrics.append(logme)
logme = log_maximum_evidence(feats_bbox, labels)
bbox_metrics.append(logme)
elif metric == 'xy_LogME':
logme = log_maximum_evidence(feats,annots[:,0:4].int().numpy(), regression = True)
metrics.append(logme)
logme = log_maximum_evidence(feats_bbox,annots[:,0:4].int().numpy(), regression = True)
bbox_metrics.append(logme)
elif metric == 'hscore':
hscore = h_score(feats, labels)
metrics.append(hscore)
hscore = h_score(feats_bbox, labels)
bbox_metrics.append(hscore)
elif metric == 'regularized_hscore':
hscore = regularized_h_score(feats, labels)
metrics.append(hscore)
hscore = regularized_h_score(feats_bbox, labels)
bbox_metrics.append(hscore)
elif metric == 'transrate':
tr = transrate(feats, labels)
metrics.append(tr)
tr = transrate(feats_bbox, labels)
bbox_metrics.append(tr)
return metrics, bbox_metrics