-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_utils.py
207 lines (164 loc) · 6.1 KB
/
test_utils.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
import numpy as np
import random
from utils import get_name
def get_random_splits(rec_IDs, N_folds=10, fold_size=1000):
N = len(rec_IDs)
st = random.getstate()
random.seed(1234)
assert N_folds >= 1
if fold_size <= 0:
fold_size = len(rec_IDs) // N_folds
splits = []
for i in range(N_folds):
idxs = random.sample(range(N), fold_size)
splits.append(idxs)
random.setstate(st)
return splits
def rank_only(img_embs, rec_embs, mode="i2t"):
assert mode in ["i2t", "t2i"], "unsupported cross modal ranking"
assert img_embs.shape == rec_embs.shape
N = img_embs.shape[0]
ranks = []
recall = {1: 0.0, 5: 0.0, 10: 0.0}
if N <= 30000:
if mode == "i2t":
sims = np.dot(img_embs, rec_embs.T)
else:
sims = np.dot(rec_embs, img_embs.T)
for i in range(N):
# sort in descending order
sorting = np.argsort(sims[i,:])[::-1].tolist()
# where this index 'i' is in the sorted list
pos = sorting.index(i)
if pos == 0:
recall[1] += 1
if pos < 5:
recall[5] += 1
if pos < 10:
recall[10] += 1
ranks.append(pos+1)
else:
for i in range(N):
if mode == "i2t":
sims = np.dot(img_embs[i,:], rec_embs.T)
else:
sims = np.dot(rec_embs[i,:], img_embs.T)
sorting = np.argsort(sims)[::-1].tolist()
# where this index 'i' is in the sorted list
pos = sorting.index(i)
if pos == 0:
recall[1] += 1
if pos < 5:
recall[5] += 1
if pos < 10:
recall[10] += 1
ranks.append(pos+1)
medRank = np.median(ranks)
for k in recall:
recall[k] = recall[k] / N
dcg = np.array([1/np.log2(r+1) for r in ranks]).mean()
return medRank, recall, dcg, ranks
def rank(img_embs, rec_embs, rec_IDs, mode="i2t", N_folds=10, fold_size=1000):
assert img_embs.shape == rec_embs.shape
assert img_embs.shape[0] == len(rec_IDs)
assert N_folds >= 1
if fold_size <= 0:
fold_size = len(rec_IDs) // N_folds
st = random.getstate()
random.seed(1234)
N = len(rec_IDs)
global_recall = {1: 0.0, 5: 0.0, 10: 0.0}
global_rank = []
all_ranks = []
global_dcg = []
for i in range(N_folds):
# sampling fold_size samples
idxs = random.sample(range(N), fold_size)
img_emb_sub = img_embs[idxs,:]
rec_emb_sub = rec_embs[idxs,:]
medRank, recall, dcg, ranks = rank_only(img_emb_sub, rec_emb_sub, mode)
global_rank.append(medRank)
all_ranks.append(np.mean(ranks))
global_dcg.append(dcg)
for k in global_recall:
global_recall[k] += recall[k]
for k in global_recall:
global_recall[k] = global_recall[k] / N_folds
random.setstate(st)
return np.average(global_rank), global_recall, np.mean(all_ranks), np.mean(global_dcg)
def rank_2(img_embs, rec_embs, rec_IDs, mode="i2t", splits=None):
assert img_embs.shape == rec_embs.shape
assert img_embs.shape[0] == len(rec_IDs)
assert splits is not None
N_folds = len(splits)
#fold_size = len(splits[0])
N = len(rec_IDs)
global_recall = {1: 0.0, 5: 0.0, 10: 0.0}
global_rank = []
all_ranks = []
global_dcg = []
for i, idxs in enumerate(splits):
# sampling fold_size samples
img_emb_sub = img_embs[idxs,:]
rec_emb_sub = rec_embs[idxs,:]
medRank, recall, dcg, ranks = rank_only(img_emb_sub, rec_emb_sub, mode)
global_rank.append(medRank)
all_ranks.append(np.mean(ranks))
global_dcg.append(dcg)
for k in global_recall:
global_recall[k] += recall[k]
for k in global_recall:
global_recall[k] = global_recall[k] / N_folds
return np.average(global_rank), global_recall, np.mean(all_ranks), np.mean(global_dcg)
def rank_3(img_embs, rec_embs, img_splits, rec_splits, mode="i2t"):
assert img_splits is not None and rec_splits is not None
assert len(img_splits) == len(rec_splits)
N_folds = len(img_splits)
global_recall = {1: 0.0, 5: 0.0, 10: 0.0}
global_rank = []
all_ranks = []
global_dcg = []
for i in range(N_folds):
# sampling fold_size samples
img_emb_sub = img_embs[img_splits[i],:]
rec_emb_sub = rec_embs[rec_splits[i],:]
medRank, recall, dcg, ranks = rank_only(img_emb_sub, rec_emb_sub, mode)
global_rank.append(medRank)
all_ranks.append(np.mean(ranks))
global_dcg.append(dcg)
for k in global_recall:
global_recall[k] += recall[k]
for k in global_recall:
global_recall[k] = global_recall[k] / N_folds
return np.average(global_rank), global_recall, np.mean(all_ranks), np.mean(global_dcg)
def find_top_k(query_emb, embs, ids, k=1, emb_pos=None):
d = np.dot(query_emb, embs.T)
sort_ind = np.argsort(d)[::-1].tolist()
nns = sort_ind[:k]
IDs = [ids[j] for j in nns]
if emb_pos is None:
return IDs
else:
rank = sort_ind.index(emb_pos) + 1
return rank, IDs
def find_top_k_both(query_emb, emb_data, k=1):
img_embs = emb_data["img_embeds"]
txt_embs = emb_data["rec_embeds"]
ids = emb_data["rec_ids"]
img_IDs = find_top_k(query_emb, img_embs, ids, k)
txt_IDs = find_top_k(query_emb, txt_embs, ids, k)
return img_IDs, txt_IDs
def find_top_k_image(query_emb, emb_data, k=1):
img_embs = emb_data["img_embeds"]
ids = emb_data["rec_ids"]
img_IDs = find_top_k(query_emb, img_embs, ids, k)
return img_IDs
def find_top_k_text(query_emb, emb_data, k=1):
txt_embs = emb_data["rec_embeds"]
ids = emb_data["rec_ids"]
txt_IDs = find_top_k(query_emb, txt_embs, ids, k)
return txt_IDs
def get_model_signature(opts):
name = get_name(opts.test_model_path)
name = "{:s}_{:s}_{:s}_{:s}_{:s}_{:s}".format(name, opts.ingrInLayer, opts.instInLayer, opts.docInLayer, "ingrAtt" if opts.ingrAtt else "noIngrAtt", "instAtt" if opts.instAtt else "noIngrAtt")
return name