-
Notifications
You must be signed in to change notification settings - Fork 1
/
generate_dataset.py
236 lines (212 loc) · 8.09 KB
/
generate_dataset.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
import os.path
import sys
import json
import random
"""
@author Arjun Albert
@email aalbert@mit.edu
@date 11/2/2020
@notes Generate training, test, and validation datasets from
Visual Genome scene ontology dataset for knowledge
graph embedding for scene semantic understanding.
"""
"""
Convert a tuple representation of a head, relationship, tail triple into a string.
Returns the string representation of the triple.
"""
def get_triple_as_string(h, r, t):
return h + "\t" + r + "\t" + t + "\n"
"""
Checks if a relationship json object has the proper subject, object, predicate
(relationship) fields required to form a head, relation, tail triple.
Returns true if the input json had proper fields, false otherwise.
"""
def check_valid_relationship(rel):
if "subject" not in rel:
return False
if "predicate" not in rel:
return False
if "object" not in rel:
return False
return True
"""
Gets an objects name from each databases json object representation.
Returns the object name string regardless of database versions object.
"""
def get_object_name(obj):
if "names" in obj:
return obj["names"][0]
if "name" in obj:
return obj["name"]
return None
"""
Gets a subjects name from each databases json subject representation.
Returns the subject name string regardless of database versions subject.
"""
def get_subject_name(sbj):
if "names" in sbj:
return sbj["names"][0]
if "name" in sbj:
return sbj["name"]
return None
"""
Imports the json data from each of the visual genomes datasets.
Adds the relationships (h, r, t) for each dataset by scene to
the it's respective dataset.
"""
def import_datasets(set1, set2, set3):
if set1:
print("Importing dataset 1.")
with open(os.path.join(sys.path[0], "scenes", "relationships_1.json"), 'r') as f:
scenes = json.loads(f.read())
for scene in scenes:
scene_id = scene["image_id"]
if scene_id not in dataset_1:
dataset_1[scene_id] = []
relationships = scene["relationships"]
for relationship in relationships:
if check_valid_relationship(relationship):
h = get_subject_name(relationship["subject"])
r = relationship["predicate"]
t = get_object_name(relationship["object"])
dataset_1[scene_id].append((h, r, t))
if set2:
print("Importing dataset 2.")
with open(os.path.join(sys.path[0], "scenes", "relationships_1.json"), 'r') as f:
scenes = json.loads(f.read())
for scene in scenes:
scene_id = scene["image_id"]
if scene_id not in dataset_2:
dataset_2[scene_id] = []
relationships = scene["relationships"]
for relationship in relationships:
if check_valid_relationship(relationship):
h = get_subject_name(relationship["subject"])
r = relationship["predicate"]
t = get_object_name(relationship["object"])
dataset_2[scene_id].append((h, r, t))
if set3:
print("Importing dataset 3.")
with open(os.path.join(sys.path[0], "scenes", "relationships_1.json"), 'r') as f:
scenes = json.loads(f.read())
for scene in scenes:
scene_id = scene["image_id"]
if scene_id not in dataset_3:
dataset_3[scene_id] = []
relationships = scene["relationships"]
for relationship in relationships:
if check_valid_relationship(relationship):
h = get_subject_name(relationship["subject"])
r = relationship["predicate"]
t = get_object_name(relationship["object"])
dataset_3[scene_id].append((h, r, t))
"""
Split the a relationship in a dataset randomly between training, test, and validation.
Returns the modified training, test, or validation datasets.
"""
def split_dataset(tr, te, va, rel):
h, r, t = rel
s = get_triple_as_string(h, r, t)
i = random.randint(0, 2)
if i == 0:
tr += s
elif i == 1:
te += s
else:
va += s
return tr, te, va
"""
Input training, test, validation data, and a dataset number to write to.
Write the training, test, and validation data to the proper directories.
"""
def write_datasets(tr, te, va, i):
with open(os.path.join(sys.path[0], "pykg2vec", "scene_data_" + str(i), "ad_scene_relationships-train.txt"), 'w') as f:
f.write(tr)
with open(os.path.join(sys.path[0], "pykg2vec", "scene_data_" + str(i), "ad_scene_relationships-test.txt"), 'w') as f:
f.write(te)
with open(os.path.join(sys.path[0], "pykg2vec", "scene_data_" + str(i), "ad_scene_relationships-valid.txt"), 'w') as f:
f.write(va)
"""
Label a relationships head and tail with an arbitrary prefix.
Return the new head and tail strings in the triple.
"""
def label_relationship(rel, pfx):
h, r, t = rel
return pfx + h, r, pfx + t
"""
Decide proper ratio to split training, test, and validation data into.
Also downsample the training data with ratio 1 : downsampling.
Gather triples from memory and then writeout each dataset.
Will overwrite previously filled out datasets with new datasets
even if they are not included or are empty.
"""
def export_data_for_training(downsampling, testing_limit):
print("Exporting datasets for training.")
max_test = testing_limit
keys = dataset_1.keys()
train_1, test_1, valid_1 = "", "", ""
relationship_count = 0
for key in keys:
for rel in dataset_1[key]:
if relationship_count % downsampling == 0:
c = random.randint(0, 10)
h, r, t = rel
s = get_triple_as_string(h, r, t)
if c < 8:
train_1 += s
elif c < 9:
if len(valid_1.split("\n")) < max_test:
valid_1 += s
else:
if len(test_1.split("\n")) < max_test:
test_1 += s
relationship_count += 1
write_datasets(train_1, test_1, valid_1, 1)
keys = dataset_2.keys()
train_2, test_2, valid_2 = "", "", ""
relationship_count = 0
for key in keys:
for rel in dataset_2[key]:
if relationship_count % downsampling == 0:
c = random.randint(0, 10)
h, r, t = rel
s = get_triple_as_string(h, r, t)
if c < 8:
train_2 += s
elif c < 9:
if len(valid_2.split("\n")) < max_test:
valid_2 += s
else:
if len(test_2.split("\n")) < max_test:
test_2 += s
relationship_count += 1
write_datasets(train_2, test_2, valid_2, 2)
keys = dataset_3.keys()
train_3, test_3, valid_3 = "", "", ""
relationship_count = 0
for key in keys:
for rel in dataset_3[key]:
if relationship_count % downsampling == 0:
c = random.randint(0, 10)
h, r, t = rel
s = get_triple_as_string(h, r, t)
if c < 8:
train_3 += s
elif c < 9:
if len(valid_3.split("\n")) < max_test:
valid_3 += s
else:
if len(test_3.split("\n")) < max_test:
test_3 += s
relationship_count += 1
write_datasets(train_3, test_3, valid_3, 3)
print("Done.")
# Store scene ontology relationships for each dataset
dataset_1 = {}
dataset_2 = {}
dataset_3 = {}
# Load dataset 1, dataset 2, dataset 3 into memory and format them
import_datasets(True, True, True)
# Export data downsampled by 100x for training, test, and validation
# Set a max length of 4000 relations each for testing and validation
export_data_for_training(50, 4000)