-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathrun.py
368 lines (331 loc) · 20.6 KB
/
run.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
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
import argparse
import time
from pprint import pprint
import numpy as np
import random
from pathlib import Path
import torch
from torch.utils.data import DataLoader
import dgl
from dgl.contrib.data import load_data
import os
import sys
import logging
from model import RGAT_DistMult, RGAT_ConvE, RGCN_ConvE, ConvE
from utils import process, TrainDataset, TestDataset, load_link
os.environ["TOKENIZERS_PARALLELISM"] = "false"
class Runner(object):
def __init__(self, params):
self.p = params
self.prj_path = Path(__file__).parent.resolve()
self.data = load_link(self.p.dataset)
self.num_ent, self.train_data, self.valid_data, self.test_data, self.num_rels = self.data.num_nodes, self.data.train, self.data.valid, self.data.test, self.data.num_rels
self.ent2textvector, self.rel2textvector, self.attr2vector = self.data.ent2textvector, self.data.rel2textvector, self.data.attr2vector
self.attrname, self.ent2value, self.ent2attrlabel = self.data.attrname, self.data.ent2value, self.data.ent2attrlabel
self.ent2imgvector = self.data.ent2imgvector
self.triplets = process({'train': self.train_data, 'valid': self.valid_data, 'test': self.test_data},
self.num_rels)
if self.p.gpu != -1 and torch.cuda.is_available():
self.device = torch.device(f'cuda:{self.p.gpu}')
else:
self.device = torch.device('cpu')
self.p.embed_dim = self.p.k_w * self.p.k_h if self.p.embed_dim is None else self.p.embed_dim # output dim of gnn
self.data_iter = self.get_data_iter()
self.g = self.build_graph()
self.edge_type, self.edge_norm = self.get_edge_dir_and_norm()
self.model = self.get_model()
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.p.lr, weight_decay=self.p.l2)
self.best_val_mrr, self.best_epoch, self.best_val_results = 0., 0., {}
self.logger = logging.getLogger()
self.log_name = self.p.name + '.log'
self.log_path = '/home/liangshuang/NewWork/logs'
pprint(vars(self.p))
def fit(self):
epoch_hits1 = []
epoch_hits3 = []
epoch_hits10 = []
epoch_mrr = []
epoch_mr = []
save_root = self.prj_path / 'checkpoints'
if not save_root.exists():
save_root.mkdir()
save_path = save_root / (self.p.name + '.pt')
if self.p.restore:
self.load_model(save_path)
print('Successfully Loaded previous model')
for epoch in range(self.p.max_epochs):
start_time = time.time()
train_loss = self.train()
val_results = self.evaluate('test')
epoch_mr.append(val_results['mr'])
epoch_mrr.append(val_results['mrr'])
epoch_hits1.append(val_results['hits@1'])
epoch_hits3.append(val_results['hits@3'])
epoch_hits10.append(val_results['hits@10'])
if val_results['mrr'] > self.best_val_mrr:
self.best_val_results = val_results
self.best_val_mrr = val_results['mrr']
self.best_epoch = epoch
self.save_model(save_path)
print(
f"[Epoch {epoch}]: Training Loss: {train_loss:.5}, Valid MRR: {val_results['mrr']:.5}, Best Valid MRR: {self.best_val_mrr:.5}, Cost: {time.time() - start_time:.2f}s")
pprint(vars(self.p))
self.load_model(save_path)
print(f'Loading best model in {self.best_epoch} epoch, Evaluating on Test data')
self.model.eval()
# entity_embedding, rel_embedding = self.model.get_embedding(self.g)
# np.save(os.path.join('./embedding', self.p.name + '_entity_embedding'), entity_embedding.cpu().numpy())
# np.save(os.path.join('./embedding', self.p.name + '_rel_embedding'), rel_embedding.cpu().numpy())
test_results = self.evaluate('test')
print(
f"MRR: Tail {test_results['left_mrr']:.5}, Head {test_results['right_mrr']:.5}, Avg {test_results['mrr']:.5}")
print(f"MR: Tail {test_results['left_mr']:.5}, Head {test_results['right_mr']:.5}, Avg {test_results['mr']:.5}")
print(f"hits@1 = {test_results['hits@1']:.5}")
print(f"hits@3 = {test_results['hits@3']:.5}")
print(f"hits@10 = {test_results['hits@10']:.5}")
print("rank: ", self.model.rank)
np.save(os.path.join('./result', self.p.name + '_epoch_mr'), epoch_mr)
np.save(os.path.join('./result', self.p.name + '_epoch_mrr'), epoch_mrr)
np.save(os.path.join('./result', self.p.name + '_epoch_hits@1'), epoch_hits1)
np.save(os.path.join('./result', self.p.name + '_epoch_hits@3'), epoch_hits1)
np.save(os.path.join('./result', self.p.name + '_epoch_hits@10'), epoch_hits1)
def train(self):
self.model.train()
losses = []
train_iter = self.data_iter['train']
for step, (triplets, labels) in enumerate(train_iter):
triplets, labels = triplets.to(self.device), labels.to(self.device)
subj, rel = triplets[:, 0], triplets[:, 1]
pred = self.model(self.g, subj, rel) # [batch_size, num_ent]
loss = self.model.calc_loss(pred, labels)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
losses.append(loss.item())
loss = np.mean(losses)
return loss
def evaluate(self, split):
"""
Function to evaluate the model on validation or test set
:param split: valid or test, set which data-set to evaluate on
:return: results['mr']: Average of ranks_left and ranks_right
results['mrr']: Mean Reciprocal Rank
results['hits@k']: Probability of getting the correct prediction in top-k ranks based on predicted score
results['left_mrr'], results['left_mr'], results['right_mrr'], results['right_mr']
results['left_hits@k'], results['right_hits@k']
"""
def get_combined_results(left, right):
results = dict()
assert left['count'] == right['count']
count = float(left['count'])
results['left_mr'] = round(left['mr'] / count, 5)
results['left_mrr'] = round(left['mrr'] / count, 5)
results['right_mr'] = round(right['mr'] / count, 5)
results['right_mrr'] = round(right['mrr'] / count, 5)
results['mr'] = round((left['mr'] + right['mr']) / (2 * count), 5)
results['mrr'] = round((left['mrr'] + right['mrr']) / (2 * count), 5)
for k in [1, 3, 10]:
results[f'left_hits@{k}'] = round(left[f'hits@{k}'] / count, 5)
results[f'right_hits@{k}'] = round(right[f'hits@{k}'] / count, 5)
results[f'hits@{k}'] = round((results[f'left_hits@{k}'] + results[f'right_hits@{k}']) / 2, 5)
return results
self.model.eval()
left_result = self.predict(split, 'tail')
right_result = self.predict(split, 'head')
res = get_combined_results(left_result, right_result)
return res
def predict(self, split='valid', mode='tail'):
"""
Function to run model evaluation for a given mode
:param split: valid or test, set which data-set to evaluate on
:param mode: head or tail
:return: results['mr']: Sum of ranks
results['mrr']: Sum of Reciprocal Rank
results['hits@k']: counts of getting the correct prediction in top-k ranks based on predicted score
results['count']: number of total predictions
"""
with torch.no_grad():
results = dict()
test_iter = self.data_iter[f'{split}_{mode}']
for step, (triplets, labels) in enumerate(test_iter):
triplets, labels = triplets.to(self.device), labels.to(self.device)
subj, rel, obj = triplets[:, 0], triplets[:, 1], triplets[:, 2]
pred = self.model(self.g, subj, rel)
b_range = torch.arange(pred.shape[0], device=self.device)
target_pred = pred[b_range, obj] # [batch_size, 1], get the predictive score of obj
# label=>-1000000, not label=>pred, filter out other objects with same sub&rel pair
pred = torch.where(labels.byte(), -torch.ones_like(pred) * 10000000, pred)
pred[b_range, obj] = target_pred # copy predictive score of obj to new pred
ranks = 1 + torch.argsort(torch.argsort(pred, dim=1, descending=True), dim=1, descending=False)[
b_range, obj] # get the rank of each (sub, rel, obj)
ranks = ranks.float()
results['count'] = torch.numel(ranks) + results.get('count', 0) # number of predictions
results['mr'] = torch.sum(ranks).item() + results.get('mr', 0)
results['mrr'] = torch.sum(1.0 / ranks).item() + results.get('mrr', 0)
for k in [1, 3, 10]:
results[f'hits@{k}'] = torch.numel(ranks[ranks <= k]) + results.get(f'hits@{k}', 0)
return results
def save_model(self, path):
"""
Function to save a model. It saves the model parameters, best validation scores,
best epoch corresponding to best validation, state of the optimizer and all arguments for the run.
:param path: path where the model is saved
:return:
"""
state = {
'model': self.model.state_dict(),
'best_val': self.best_val_results,
'best_epoch': self.best_epoch,
'optimizer': self.optimizer.state_dict(),
'args': vars(self.p)
}
torch.save(state, path)
def load_model(self, path):
"""
Function to load a saved model
:param path: path where model is loaded
:return:
"""
state = torch.load(path)
self.best_val_results = state['best_val']
self.best_val_mrr = self.best_val_results['mrr']
self.best_epoch = state['best_epoch']
self.model.load_state_dict(state['model'])
self.optimizer.load_state_dict(state['optimizer'])
def build_graph(self):
g = dgl.DGLGraph()
g.add_nodes(self.num_ent)
g.add_edges(self.train_data[:, 0], self.train_data[:, 2])
g.add_edges(self.train_data[:, 2], self.train_data[:, 0])
return g
def get_data_iter(self):
"""
get data loader for train, valid and test section
:return: dict
"""
def get_data_loader(dataset_class, split):
return DataLoader(
dataset_class(self.triplets[split], self.num_ent, self.p),
batch_size=self.p.batch_size,
shuffle=True,
num_workers=self.p.num_workers
)
return {
'train': get_data_loader(TrainDataset, 'train'),
'valid_head': get_data_loader(TestDataset, 'valid_head'),
'valid_tail': get_data_loader(TestDataset, 'valid_tail'),
'test_head': get_data_loader(TestDataset, 'test_head'),
'test_tail': get_data_loader(TestDataset, 'test_tail')
}
def get_edge_dir_and_norm(self):
"""
:return: edge_type: indicates type of each edge: [E]
"""
in_deg = self.g.in_degrees(range(self.g.number_of_nodes())).float().numpy()
norm = in_deg ** -0.5
norm[np.isinf(norm)] = 0
self.g.ndata['xxx'] = norm
self.g.apply_edges(lambda edges: {'xxx': edges.dst['xxx'] * edges.src['xxx']})
norm = self.g.edata.pop('xxx').squeeze().to(self.device)
edge_type = torch.tensor(np.concatenate([self.train_data[:, 1], self.train_data[:, 1] + self.num_rels])).to(
self.device)
return edge_type, norm
def get_model(self):
if self.p.model_func.lower() == 'distmult':
model = RGAT_DistMult(num_ent=self.num_ent, num_rel=self.num_rels, num_base=self.p.num_bases,
init_dim=self.p.init_dim, gcn_dim=self.p.gcn_dim, embed_dim=self.p.embed_dim,
n_layer=self.p.n_layer, edge_type=self.edge_type, edge_norm=self.edge_norm,
ent2textvector = self.ent2textvector, rel2textvector = self.rel2textvector,
ent2attr = self.ent2value, ent2attrlabel = self.ent2attrlabel, attr2vector = self.attr2vector, ent2imgvector = self.ent2imgvector,
use_text = self.p.text, use_img = self.p.img, use_attr = self.p.attr,
device = self.device, bias=self.p.bias, gcn_drop=self.p.gcn_drop, opn=self.p.opn,
hid_drop=self.p.hid_drop)
elif self.p.model_func.lower() == 'conve':
model = RGAT_ConvE(num_ent=self.num_ent, num_rel=self.num_rels, num_base=self.p.num_bases,
init_dim=self.p.init_dim, gcn_dim=self.p.gcn_dim, embed_dim=self.p.embed_dim,
n_layer=self.p.n_layer, edge_type=self.edge_type, edge_norm=self.edge_norm,
ent2textvector=self.ent2textvector, rel2textvector=self.rel2textvector,
ent2attr=self.ent2value, ent2attrlabel=self.ent2attrlabel, attr2vector=self.attr2vector, ent2imgvector = self.ent2imgvector,
use_text=self.p.text, use_img=self.p.img, use_attr=self.p.attr,
device=self.device,bias=self.p.bias, gcn_drop=self.p.gcn_drop, opn=self.p.opn,
hid_drop=self.p.hid_drop, input_drop=self.p.input_drop,
conve_hid_drop=self.p.conve_hid_drop, feat_drop=self.p.feat_drop,
num_filt=self.p.num_filt, ker_sz=self.p.ker_sz, k_h=self.p.k_h, k_w=self.p.k_w)
elif self.p.model_func.lower() == 'rgcn':
model = RGCN_ConvE(num_ent=self.num_ent, num_rel=self.num_rels, num_base=self.p.num_bases,
init_dim=self.p.init_dim, gcn_dim=self.p.gcn_dim, embed_dim=self.p.embed_dim,
n_layer=self.p.n_layer, edge_type=self.edge_type, edge_norm=self.edge_norm,
ent2textvector=self.ent2textvector, rel2textvector=self.rel2textvector,
ent2attr=self.ent2value, ent2attrlabel=self.ent2attrlabel, attr2vector=self.attr2vector, ent2imgvector = self.ent2imgvector,
use_text=self.p.text, use_img=self.p.img, use_attr=self.p.attr,
device=self.device,bias=self.p.bias, gcn_drop=self.p.gcn_drop, opn=self.p.opn,
hid_drop=self.p.hid_drop, input_drop=self.p.input_drop,
conve_hid_drop=self.p.conve_hid_drop, feat_drop=self.p.feat_drop,
num_filt=self.p.num_filt, ker_sz=self.p.ker_sz, k_h=self.p.k_h, k_w=self.p.k_w)
elif self.p.model_func.lower() == 'simpleconve':
model = ConvE(num_ent=self.num_ent, num_rel=self.num_rels, num_base=self.p.num_bases,
init_dim=self.p.init_dim, gcn_dim=self.p.gcn_dim, embed_dim=self.p.embed_dim,
n_layer=self.p.n_layer, edge_type=self.edge_type, edge_norm=self.edge_norm,
ent2textvector=self.ent2textvector, rel2textvector=self.rel2textvector,
ent2attr=self.ent2value, ent2attrlabel=self.ent2attrlabel, attr2vector=self.attr2vector, ent2imgvector = self.ent2imgvector,
use_text=self.p.text, use_img=self.p.img, use_attr=self.p.attr,
device=self.device,bias=self.p.bias, gcn_drop=self.p.gcn_drop, opn=self.p.opn,
hid_drop=self.p.hid_drop, input_drop=self.p.input_drop,
conve_hid_drop=self.p.conve_hid_drop, feat_drop=self.p.feat_drop,
num_filt=self.p.num_filt, ker_sz=self.p.ker_sz, k_h=self.p.k_h, k_w=self.p.k_w)
else:
raise KeyError(f'score function {self.p.model_func} not recognized.')
model.to(self.device)
return model
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Parser For Arguments',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--name', default='test_run', help='Set run name for saving/restoring models')
parser.add_argument('--data', dest='dataset', default='FB15k-237', help='Dataset to use, default: FB15k-237')
parser.add_argument('--model_func', dest='model_func', default='conve',
help='Score Function for Link prediction')
parser.add_argument('--opn', dest='opn', default='corr', help='Composition Operation to be used in CompGCN')
parser.add_argument('--batch', dest='batch_size', default=256, type=int, help='Batch size')
parser.add_argument('--gpu', type=int, default=0, help='Set GPU Ids : Eg: For CPU = -1, For Single GPU = 0')
parser.add_argument('--epoch', dest='max_epochs', type=int, default=500, help='Number of epochs')
parser.add_argument('--l2', type=float, default=0.0, help='L2 Regularization for Optimizer')
parser.add_argument('--lr', type=float, default=0.001, help='Starting Learning Rate')
parser.add_argument('--lbl_smooth', dest='lbl_smooth', type=float, default=0.1, help='Label Smoothing')
parser.add_argument('--num_workers', type=int, default=8, help='Number of processes to construct batches')
parser.add_argument('--seed', dest='seed', default=12345, type=int, help='Seed for randomization')
parser.add_argument('--restore', dest='restore', action='store_true',
help='Restore from the previously saved model')
parser.add_argument('--bias', dest='bias', action='store_true', help='Whether to use bias in the model')
parser.add_argument('--text', dest='text', action='store_true', help='Whether to use text in the model')
parser.add_argument('--img', dest='img', action='store_true', help='Whether to use img in the model')
parser.add_argument('--attr', dest='attr', action='store_true', help='Whether to use attr in the model')
parser.add_argument('--num_bases', dest='num_bases', default=-1, type=int,
help='Number of basis relation vectors to use')
parser.add_argument('--init_dim', dest='init_dim', default=100, type=int,
help='Initial dimension size for entities and relations')
parser.add_argument('--gcn_dim', dest='gcn_dim', default=200, type=int, help='Number of hidden units in GCN')
parser.add_argument('--embed_dim', dest='embed_dim', default=None, type=int,
help='Embedding dimension to give as input to score function')
parser.add_argument('--n_layer', dest='n_layer', default=1, type=int, help='Number of GCN Layers to use')
parser.add_argument('--gcn_drop', dest='gcn_drop', default=0.1, type=float, help='Dropout to use in GCN Layer')
parser.add_argument('--hid_drop', dest='hid_drop', default=0.3, type=float, help='Dropout after GCN')
# ConvE specific hyperparameters
parser.add_argument('--conve_hid_drop', dest='conve_hid_drop', default=0.3, type=float,
help='ConvE: Hidden dropout')
parser.add_argument('--feat_drop', dest='feat_drop', default=0.2, type=float, help='ConvE: Feature Dropout')
parser.add_argument('--input_drop', dest='input_drop', default=0.2, type=float, help='ConvE: Stacked Input Dropout')
parser.add_argument('--k_w', dest='k_w', default=20, type=int, help='ConvE: k_w')
parser.add_argument('--k_h', dest='k_h', default=10, type=int, help='ConvE: k_h')
parser.add_argument('--num_filt', dest='num_filt', default=200, type=int,
help='ConvE: Number of filters in convolution')
parser.add_argument('--ker_sz', dest='ker_sz', default=7, type=int, help='ConvE: Kernel size to use')
args = parser.parse_args()
if not args.restore:
args.name = time.strftime('%Y_%m_%d') + '_' + time.strftime(
'%H:%M:%S') + '-' + args.model_func.lower() + '-' + args.opn
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
runner = Runner(args)
runner.fit()