-
Notifications
You must be signed in to change notification settings - Fork 93
/
Copy pathcdp.py
201 lines (175 loc) · 8.21 KB
/
cdp.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
import numpy as np
import os
import pdb
import sys
import time
import json
from . import graph
import pickle
from . import eval_cluster
from .create_pair_set import create
from .mediator import Mediator
from .utils import log
def sample(base, committee, accept=7, th=0.7):
pairs = []
scores = []
if len(committee) > 0:
knn = base[0]
k = knn.shape[1]
tile_knn = np.tile(knn.reshape(len(knn), -1, 1), (1, 1, k))
simi = 1.0 - base[1]
anchor = np.tile(np.arange(len(knn)).reshape(len(knn), 1), (1, knn.shape[1]))
vote_num = np.zeros(knn.shape, dtype=np.int)
for cmt in committee:
tile_cmt = np.tile(cmt[0].reshape(len(cmt[0]), 1, -1), (1, k, 1))
vote_num += (tile_knn == tile_cmt).sum(axis=2)
selidx = np.where((simi > th) & (vote_num >= accept) & (knn != -1) & (knn != anchor))
else:
knn = base[0]
simi = 1.0 - base[1]
anchor = np.tile(np.arange(len(knn)).reshape(len(knn), 1), (1, knn.shape[1]))
selidx = np.where((simi > th) & (knn != -1) & (knn != anchor))
pairs = np.hstack((anchor[selidx].reshape(-1, 1), knn[selidx].reshape(-1, 1)))
scores = simi[selidx]
pairs = np.sort(pairs, axis=1)
pairs, unique_idx = np.unique(pairs, return_index=True, axis=0)
scores = scores[unique_idx]
return pairs, scores
def cdp(args):
exp_root = os.path.dirname(args.config)
setattr(args, 'exp_root', exp_root)
with open("data/{}/list.txt".format(args.data_name), 'r') as f:
fns = f.readlines()
args.total_num = len(fns)
if args.strategy == "vote":
output_cdp = '{}/output/k{}_{}_accept{}_th{}'.format(exp_root, args.k, args.strategy, args.vote['accept_num'], args.vote['threshold'])
elif args.strategy == "mediator":
output_cdp = '{}/output/k{}_{}_{}{}{}_th{}'.format(
exp_root,
args.k,
args.strategy,
int('relationship' in args.mediator['input']),
int('affinity' in args.mediator['input']),
int('structure' in args.mediator['input']),
args.mediator['threshold'])
elif args.strategy == 'groundtruth':
output_cdp = '{}/output/gt'.format(exp_root)
else:
raise Exception('No such strategy: {}'.format(args.strategy))
output_sub = '{}/sz{}_step{}'.format(output_cdp, args.propagation['max_sz'], args.propagation['step'])
log('Output folder: {}'.format(output_sub))
outmeta = '{}/meta.txt'.format(output_sub)
if not os.path.isdir(output_sub):
os.makedirs(output_sub)
# pair selection
if args.strategy == 'vote':
pairs, scores = vote(output_cdp, args)
elif args.strategy == 'mediator':
pairs, scores = mediator(args)
elif args.strategy == 'groundtruth': # only for debug
pairs, scores = groundtruth(args)
else:
raise Exception("No such strategy: {}".format(args.strategy))
log("\tpair num: {}".format(len(pairs)))
if len(pairs) == 0:
raise Exception('No positive pair is discovered, please decrease the threshold.')
# propagation
log("Propagation ...")
components = graph.graph_propagation(pairs, scores, args.propagation['max_sz'], args.propagation['step'], args.propagation['max_iter'])
# collect results
cdp_res = []
for c in components:
cdp_res.append(sorted([n.name for n in c]))
pred = -1 * np.ones(args.total_num, dtype=np.int)
for i,c in enumerate(cdp_res):
pred[np.array(c)] = i
valid = np.where(pred != -1)
_, unique_idx = np.unique(pred[valid], return_index=True)
pred_unique = pred[valid][np.sort(unique_idx)]
pred_mapping = dict(zip(list(pred_unique), range(pred_unique.shape[0])))
pred_mapping[-1] = -1
pred = np.array([pred_mapping[p] for p in pred])
# analyse results
num_valid = len(valid[0])
num_class = len(pred_unique)
log("\n------------- Analysis --------------")
log('num_images: {}\tnum_class: {}\tnum_per_class: {:.2g}'.format(num_valid, num_class, num_valid/float(num_class)))
log("Discard ratio: {:.4g}".format(1 - num_valid / float(len(pred))))
# evaluate
if args.evaluation:
log("\n------------- Evaluation --------------")
if not os.path.isfile("data/{}/meta.txt".format(args.data_name)):
raise Exception("Meta file not exist: {}, please create meta.txt or set evaluation to False".format("data/{}/meta.txt".format(args.data_name)))
with open("data/{}/meta.txt".format(args.data_name), 'r') as f:
label = f.readlines()
label = np.array([int(l.strip()) for l in label])
assert len(label) == len(pred), "numbers of labels and predictions are different: {} vs {}".format(len(label), len(pred))
# pair evaluation
log("Pair accuracy: {:.4g}".format((label[pairs[:,0]] == label[pairs[:,1]]).sum() / float(len(pairs))))
# clustering evaluation
pred_with_singular = pred.copy()
pred_with_singular[np.where(pred == -1)] = np.arange(num_class, num_class + (pred == -1).sum()) # to assign -1 with new labels
log('(singular removed) prec / recall / fscore: {:.4g}, {:.4g}, {:.4g}'.format(*eval_cluster.fscore(label[valid], pred[valid])))
log('(singular kept) prec / recall / fscore: {:.4g}, {:.4g}, {:.4g}'.format(*eval_cluster.fscore(label, pred_with_singular)))
# write to list
new_label = ['{}\n'.format(p) for p in pred]
if not os.path.isdir(os.path.dirname(outmeta)):
os.makedirs(os.path.dirname(outmeta))
log('Writing to: {}'.format(outmeta))
with open(outmeta, 'w') as f:
f.writelines(new_label)
log("\n--------------- End -----------------")
def vote(output, args):
assert args.vote['accept_num'] <= len(args.committee)
base_knn_fn = 'data/{}/knn/{}_k{}.npz'.format(args.data_name, args.base, args.k)
committee_knn_fn = ['data/{}/knn/{}_k{}.npz'.format(args.data_name, cmt, args.k) for cmt in args.committee]
if not os.path.isfile(output + '/vote_pairs.npy'):
log('Extracting pairs by voting ...')
knn_file = np.load(base_knn_fn)
knn_base = (knn_file['idx'], knn_file['dist'])
knn_committee = []
for i,cfn in enumerate(committee_knn_fn):
knn_file = np.load(cfn)
knn_committee.append((knn_file['idx'], knn_file['dist']))
pairs, scores = sample(knn_base, knn_committee, accept=args.vote['accept_num'], th=args.vote['threshold'])
np.save(output + '/vote_pairs.npy', pairs)
np.save(output + '/vote_scores.npy', scores)
else:
log('Loading pairs by voting ...')
pairs = np.load(output + '/vote_pairs.npy')
scores = np.load(output + '/vote_scores.npy')
return pairs, scores
def mediator(args):
args.mediator['model_name'] = "data/{}/models/k{}_{}{}{}.pth.tar".format(
args.mediator['train_data_name'],
args.k,
int('relationship' in args.mediator['input']),
int('affinity' in args.mediator['input']),
int('structure' in args.mediator['input']),
)
med = Mediator(args)
if not os.path.isfile(args.mediator['model_name']) or args.mediator['force_retrain']:
log("Creating pair set for: labeled")
create(args.mediator['train_data_name'], args, phase="train")
log("Training")
med.train()
else:
log("Mediator model exists: {}".format(args.mediator['model_name']))
log("Creating pair set for: unlabeled")
create(args.data_name, args)
log("Testing")
med.test()
raw_pairs = np.load("{}/output/pairset/k{}/pairs.npy".format(args.exp_root, args.k))
pair_pred = np.load("{}/output/pairset/k{}/pairs_pred.npy".format(args.exp_root, args.k))
sel = np.where(pair_pred > args.mediator['threshold'])[0]
pairs = raw_pairs[sel, :]
scores = pair_pred[sel]
return pairs, scores
def groundtruth(args):
raw_pairs = np.load("{}/output/{}/k{}/pairs.npy".format(args.exp_root, args.data_name, args.k))
pair_gt = np.load("{}/output/{}/k{}/pair_label.npy".format(args.exp_root, args.data_name, args.k))
pairs = raw_pairs[np.where(pair_gt == 1)[0], :]
pairs = pairs[np.where(pairs[:,0] != pairs[:,1])]
pairs = np.unique(np.sort(pairs, axis=1), axis=0)
scores = np.ones((pairs.shape[0]), dtype=np.float32)
return pairs, scores