-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_label_embedding.py
324 lines (275 loc) · 12.3 KB
/
train_label_embedding.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
"""
Runs our hyperbolic entailment cones on the synthetic data representing a uniform tree of some
fixed branching factor and some fixed depth. These trees are in data/toy/. This code will produce
an animation when embedding dimension is 2, and the animation will be opened in your browser at the
end of the training.
"""
# +
import csv
from collections import OrderedDict
import os
import pickle
import random
import re
from random import *
from smart_open import smart_open
import argparse
from utils.poincare_model import *
from utils.relations import *
from utils.eucl_cones_model import *
from utils.hyp_cones_model import *
from utils.poincare_viz import *
from utils.eval import *
import logging
import numpy as np
import sys
import plotly
from plotly.offline import plot
sys.path.insert(1, os.path.join(sys.path[0], '..'))
# TODO: fill here your plotly details before running.
plotly.tools.set_credentials_file(username='', api_key='')
from config import get_dicts
parser = argparse.ArgumentParser()
parser.add_argument("--tree", type=str, help='go_emotion | ED | ED_easy_4 | ED_hard_a | ED_hard_b | ED_hard_c | ED_hard_d ', default='ED')
parser.add_argument("--model", type=str, help='poincare | hyp_cones ', default='poincare')
parser.add_argument("--dim", type=int, default=100)
args = parser.parse_args()
default_params = {
'model' : args.model, # poincare or hyp_cones (initialized with Poincare)
'tree_depth' : 3, # a number between 1 and 7
'level_branch' : 4, # a number between 3 and 4
'remove_root': True,
'print_every': 5,
'size': args.dim, # embedding dimension. If 2, this code will produce an animation.
'lr': 0.025, # Learning rate
'opt': 'rsgd', # rsgd or exp_map or sgd
'burn_in': 20, # Number of epochs to use for burn-in initialization
'epsilon': 1e-5,
'seed': 7, # random seed
'num_negative': 5, # Number of negative samples to use (5 for go_emotion, 10 for empathatic dialogs)
'neg_sampling_power': 0.75, # 0 for uniform, 1 for unigram, 0.75 for word2vec
'neg_sampl_strategy': 'true_neg', # 'all', 'true_neg' , 'all_non_leaves' or 'true_neg_non_leaves'
'where_not_to_sample': 'ancestors', # both or ancestors or children. Has no effect if neg_sampl_strategy = 'all'.
'neg_edges_attach': 'child', # How to form negative edges: 'parent' (u,v') or 'child' (u', v) or 'both'
'always_v_in_neg': True, # always include the true edge (u,v) as negative.
'loss_type': 'nll', # 'nll', 'neg', 'maxmargin'
'maxmargin_margin': 2,
'neg_r': 1,
'neg_t': 2,
'neg_mu': 1.0, # Balancing factor between the positive and negative terms
'epochs': 2000, # Number of epochs to use
'batch_size': 16, # Size of batch to use for training
}
params = default_params.copy()
param_str_list = ['%s:%s' % (key, params[key]) for key in sorted(params.keys())]
figure_name = '; '.join(param_str_list)
animation = create_animation(figure_name)
if args.tree == 'go_emotion':
data_directory = os.path.join((os.path.abspath('')), 'data', 'go_emotion')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED_easy_4':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED_easy_4')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED_hard_a':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED_hard_a')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED_hard_b':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED_hard_b')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED_hard_c':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED_hard_c')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
elif args.tree == 'ED_hard_d':
data_directory = os.path.join((os.path.abspath('')), 'data', 'ED_hard_d')
data_file_path = os.path.join(data_directory, 'label_tree.tsv')
root_label = 'root'
# Recovers the tree from the transitive closure of a DAG
def recover_tree_from_transitive_closure(relations):
all_nodes_set = set()
for rel in relations:
all_nodes_set.add(rel[0])
all_nodes_set.add(rel[1])
ancestors = {}
for node in all_nodes_set:
ancestors[node] = []
for rel in relations:
if rel[0] != rel[1]:
ancestors[rel[1]].append(rel[0])
new_relations = []
for node in all_nodes_set:
num_ancestors = len(ancestors[node])
for ancestor in ancestors[node]:
if len(ancestors[ancestor]) == num_ancestors - 1:
new_relations.append((ancestor, node))
return new_relations
def read_tree_data(data=args.tree):
# Load the tree data:
transitive_relations = Relations(file_path=data_file_path, reverse=True)
tree_relations = recover_tree_from_transitive_closure(transitive_relations)
# All direct children of root
transitive_relations_without_root = []
tree_relations_without_root = []
for rel in tree_relations:
if rel[0] != root_label:
tree_relations_without_root.append(rel)
for rel in transitive_relations:
if rel[0] != root_label:
transitive_relations_without_root.append(rel)
return transitive_relations_without_root, tree_relations_without_root
transitive_relations, tree_relations = read_tree_data()
label2idx, idx2label = get_dicts(args.tree, return_emb_dicts=False)
class_names = [v for k, v in sorted(idx2label.items(), key=lambda item: item[0])]
# Create the Poincare model
model = PoincareModel(train_data=transitive_relations,
dim=params['size'],
lr=params['lr'],
opt=params['opt'],
burn_in=params['burn_in'],
epsilon=params['epsilon'],
seed=params['seed'],
# logger=logger,
num_negative=params['num_negative'],
### How to sample negatives for an edge (u,v)
neg_sampl_strategy=params['neg_sampl_strategy'],
# 'all' (all nodes used for negative sampling) or 'true_neg' (only not connected nodes)
where_not_to_sample=params['where_not_to_sample'],
# both or ancestors or children. Has no effect if neg_sampl_strategy = 'all'.
always_v_in_neg=params['always_v_in_neg'], # always include the true edge (u,v) as negative.
neg_sampling_power=params['neg_sampling_power'], # 0 for uniform, 1 for unigram, 0.75 for word2vec
### How to use the negatives in the loss function
neg_edges_attach=params['neg_edges_attach'],
# How to form negative edges: 'parent' (u,v') or 'child' (u', v) or 'both'
loss_type=params['loss_type'],
maxmargin_margin=params['maxmargin_margin'],
neg_r=params['neg_r'],
neg_t=params['neg_t'],
neg_mu=params['neg_mu'],
)
for i in range(int(params['epochs'] / params['print_every'])):
print('Starting epoch ' + str(params['print_every'] * i))
model.train(epochs=params['print_every'], batch_size=params['batch_size'], print_every=params['print_every'])
# Animation
if params['size'] == 2:
figure = poincare_2d_visualization(
model,
animation=animation,
epoch=(params['print_every'] * i),
eval_result='',
avg_loss=0,
avg_pos_loss=0,
avg_neg_loss=0,
tree=list(tree_relations),
show_node_labels=class_names,
figure_title=figure_name,
num_nodes=None)
if default_params['model'] == 'poincare':
if args.dim == 2: plot(animation)
word2vec = {}
for k in label2idx.keys():
if k in model.kv:
word2vec[k] = model.kv[k].astype(float)
# word2vec['sad'].astype(float)
bin_directory = os.path.join((os.path.abspath('')), 'label_tree')
bin_file_path = os.path.join(bin_directory, args.tree + '.bin')
pickle.dump(word2vec, open(bin_file_path, 'wb'))
# model.save(f'saved_poincare_model_{args.tree}.pt')
os._exit(0)
def poincare_ball_dist(u, v):
euclidean_dists = np.linalg.norm(u - v)
u_norm = np.linalg.norm(u)
v_norm = np.linalg.norm(v)
poincare_dists = np.arccosh(
1 + 2 * (
(euclidean_dists ** 2) / ((1 - u_norm ** 2) * (1 - v_norm ** 2))
)
)
return poincare_dists
model = HypConesModel(transitive_relations,
dim=params['size'],
init_range=(-0.1, 0.1),
lr=0.0001,
seed=params['seed'],
logger=logger,
num_negative=5,
### How to sample negatives for an edge (u,v)
neg_sampl_strategy='true_neg',
# 'all', 'true_neg' , 'all_non_leaves' or 'true_neg_non_leaves'
where_not_to_sample='ancestors',
# both or ancestors or children. Has no effect if neg_sampl_strategy = 'all'.
neg_edges_attach='child',
# How to form negative edges: 'parent' (u,v') or 'child' (u', v) or 'both'
neg_sampling_power=0,
# 0 for uniform, 1 for unigram, 0.75 for word2vec
margin=0.01, # Margin for the loss.
opt=params['opt'],
K=0.1,
epsilon=1e-4,
cvpr_loss='sim'
)
vecs = model.kv.syn0
model.kv.syn0 = model._clip_vectors(vecs * 0.8)
print('Finished initialization. Now training the hyperbolic cones..')
# Train the model
for i in range(int(params['epochs'] / params['print_every'])):
print('Starting epoch ' + str(params['print_every'] * i))
# Train
if i < 1:
# No training here, just to plot the initial state.
avg_loss, avg_pos_loss, avg_neg_loss = \
model.train(epochs=0, batch_size=params['batch_size'], print_every=params['print_every'])
else:
avg_loss, avg_pos_loss, avg_neg_loss = \
model.train(epochs=params['print_every'], batch_size=params['batch_size'], print_every=params['print_every'])
# Animation
if params['size'] == 2:
figure = poincare_2d_visualization(
model,
animation=animation,
epoch=(params['epochs'] + params['print_every'] * (i+1)),
eval_result='',
avg_loss=avg_loss,
avg_pos_loss=avg_pos_loss,
avg_neg_loss=avg_neg_loss,
tree=list(tree_relations),
show_node_labels=show_node_labels,
figure_title=figure_name,
num_nodes=None)
if params['size'] == 2:
plot(animation)
word2vec = {}
for k in label2idx.keys():
if k in model.kv:
word2vec[k] = model.kv[k].astype(float)
if 'sad' in word2vec:
# little buggy in ED dataset
word2vec['sad'].astype(float)
pickle.dump(word2vec, open(os.path.join((os.path.abspath('..'), "label_tree", args.tree + ".bin") , 'wb')))
# ### If one wants to save and reload a model later:
# pickle.dump(model.kv, open(args.root_path + "model_" + str(dim) + "D.bin", 'wb'))
#
# model = PoincareModel(train_data=transitive_relations, dim=dim, logger=logger)
# model.kv = pickle.load(open(args.root_path + "model_" + str(dim) + "D.bin", 'rb'))
#
# if dim == 2:
# figure = poincare_2d_visualization(
# model,
# animation=animation,
# epoch=(0),
# eval_result='',
# avg_loss=0,
# avg_pos_loss=0,
# avg_neg_loss=0,
# tree=list(),
# show_node_labels=show_node_labels,
# figure_title=figure_name,
# num_nodes=None)
# plot(animation)