-
Notifications
You must be signed in to change notification settings - Fork 0
/
process_datasets.py
225 lines (190 loc) · 7.25 KB
/
process_datasets.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
import dgl
import _pickle as pickle
import torch
import params
import networkx as nx
import matplotlib.pyplot as plt
import random
from tqdm import tqdm
from argparse import ArgumentParser
def show_statistic(g):
print("nodes: %d" % (g.number_of_nodes()))
print("edges: %d" % (g.number_of_edges()))
print(g.ndata['norm'])
return
def create_type2id(dependency2id, pos2id):
'''combine dependecy2id and pos2id to type2id
'''
type2id = dict()
count = 0
for dep in dependency2id:
type2id[dep] = count
count += 1
for pos in pos2id:
type2id[pos] = count
count += 1
return type2id
def extract_bin(atom,
sth2id,
graph,
name,
heterogeneous=False,
visualize=False,
save_gml=False):
''' parse data in each atom of train/valid/test.bin and create dgl_graph
each sample contains topk + 1 atom which are[summary,pos,neg_1,...,neg_topk-1]
each atom is a dictionary : {'sentence':str,'edges':list of tuple(src_word,src_id, src_pos,dep,tgt_word,tgt_id, tgt_pos)}
'''
if heterogeneous:
return None
else:
g = dgl.DGLGraph()
# we treat dep as a node but different dependency relations with same dep type are treated as one node
word_count = len(atom['edges'])
dep_set = set([x[3] for x in atom['edges']])
dep_count = len(dep_set)
g.add_nodes(word_count + dep_count)
onehot = [-1 for _ in range(word_count + dep_count)]
# transform word and dep into nodes id
# add dep node onehot
dep2node = dict()
count = word_count
for dep in dep_set:
dep2node[dep] = count
onehot[count] = sth2id[dep]
count += 1
# add edges and word node onehot
edge_list = []
if count == word_count + dep_count:
for idx, edge in enumerate(atom['edges']):
# add src -> dep
edge_list.append(tuple([edge[1], dep2node[edge[3]]]))
# add dep -> tgt
edge_list.append(tuple([dep2node[edge[3]], edge[5]]))
# add attribute (word pos)
onehot[idx] = sth2id[edge[2]]
else:
print("wrong")
print(count, word_count, dep_count)
print(atom['sentence'])
# save gml graph for better visualization
if save_gml:
print(atom['index'])
print(atom['sentence'])
if atom['index'] == 9:
G_gml = nx.Graph()
for idx, edge in enumerate(atom['edges']):
# add src -> dep
G_gml.add_edge(edge[2] + '_' + str(edge[1]), edge[3])
# add dep -> tgt
G_gml.add_edge(edge[3], edge[6] + '_' + str(edge[5]))
nx.write_gml(G_gml,
'./G_' + name + "_" + str(atom['index']) + ".gml")
# add edges into DGL graph
# double direction?
src, dst = tuple(zip(*edge_list))
g.add_edges(src, dst)
if graph == "undirected":
g.add_edges(dst, src)
# add norm for all nodes
unnormed = g.in_degrees(g.nodes()) + g.out_degrees(g.nodes())
g.ndata['norm'] = torch.sqrt(unnormed.float())
# add symmetric norm value on edge
for i in range(g.number_of_edges()):
src, tgt = g.find_edges(i)
g.edges[i].data['sym_norm'] = 1.0 / \
(g.nodes[src].data['norm'] * g.nodes[tgt].data['norm'])
if visualize:
# visualize graph
id2sth = {v: k for k, v in sth2id.items()}
labels = dict()
for idx, i in enumerate(onehot):
labels[idx] = id2sth[i]
nx_G = g.to_networkx()
# pos = nx.kamada_kawai_layout(nx_G)
pos = nx.nx_agraph.graphviz_layout(nx_G, prog='dot')
nx.draw(nx_G,
pos,
with_labels=True,
labels=labels,
node_size=800,
node_color=[[.7, .7, .7]],
arrowsize=5)
plt.show()
return g, onehot
def build_homogeneous(train_bin, graph):
''' only use pos and dependency as feature of nodes, all nodes share the same type
'''
# get id
dependency2id = pickle.load(open("./data/dependency2id", "rb"))
pos2id = pickle.load(open("./data/pos2id", "rb"))
type2id = create_type2id(dependency2id, pos2id)
pickle.dump(type2id, open("./data/type2id", "wb"))
# prepare processed sample lists
result_list = []
sentence_pair_list = []
id_list = []
count = 0
bin_num = 0
if graph == "directed":
prefix = "./data/large_directed"
elif graph == "undirected":
prefix = "./data/large_undirected"
for sample in tqdm(train_bin):
try:
summary_graph, summary_onehot = extract_bin(sample[0],
type2id,
graph,
"gold",
save_gml=True)
except IndexError:
print(sample)
exit
s = input()
pos_graph, pos_onehot = extract_bin(sample[1],
type2id,
graph,
"pos",
save_gml=True)
rand_choose = random.randint(2, params.topk)
neg_graph, neg_onehot = extract_bin(sample[rand_choose],
type2id,
graph,
"neg",
save_gml=True)
temp = tuple([
summary_graph, summary_onehot, pos_graph, pos_onehot, neg_graph,
neg_onehot
])
result_list.append(temp)
sentence_pair_list.append(
tuple([sample[0]['sentence'], sample[1]['sentence']]))
id_list.append(sample[0]['index'])
count += 1
if count % params.bin_size == 0:
pickle.dump(result_list, open(prefix + ".bin" + str(bin_num),
"wb"))
pickle.dump(
sentence_pair_list,
open(prefix + "sentence_pair" + ".bin" + str(bin_num), "wb"))
pickle.dump(id_list,
open(prefix + "id_list" + ".bin" + str(bin_num), "wb"))
del (result_list)
del (sentence_pair_list)
del (id_list)
result_list = []
sentence_pair_list = []
id_list = []
bin_num += 1
if __name__ == "__main__":
# parse argument
parser = ArgumentParser()
parser.add_argument("-g",
"--graph",
help="graph type, undirected|directed",
default="undirected")
args = parser.parse_args()
graph = args.graph
train_bin = pickle.load(open("./data/train.bin", "rb"))
print("train_bin loaded")
build_homogeneous(train_bin, graph)