-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
159 lines (124 loc) · 8.51 KB
/
eval.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
#!/home/pahuja.9/miniconda3/envs/colake/bin/python
import sys, os
import argparse
import numpy as np
from tqdm import tqdm
import json
import random
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
from utils.gen_utils import *
from dataset import SubGraphDataset
from dataset_eval import LinkPredSubGraphDataset
from model_self_attn import GraphTransformer
from model_cross_attn import GraphTransformer as GraphTransformerCrossAttn
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--dataset-path', type=str, required=True)
parser.add_argument('--split', type=str, choices=['train', 'valid', 'test'], default='valid')
parser.add_argument('--model-type', type=str, choices=['self-attn', 'cross-attn'], default='cross-attn')
parser.add_argument('--ckpt-path', type=str, required=True)
parser.add_argument('--sample-size', type=int, default=20, help='sample size in terms of no. of edges')
parser.add_argument('--embed-dim', type=int, default=320, help='embedding dim.')
parser.add_argument('--n-attn-heads', type=int, default=8)
parser.add_argument('--n-bert-layers', type=int, default=3)
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--beta', type=float, default=0.999, help='beta_2 of adam')
parser.add_argument('--cuda', action='store_true')
parser.add_argument('--n-epochs', type=int, default=10)
parser.add_argument('--neigh-size', type=int, default=10)
parser.add_argument('--batch-size', type=int, default=1)
parser.add_argument('--graph-connection', type=str, choices=['type_1', 'type_2'], default='type_3')
parser.add_argument('--sampling-type', type=str, choices=['bfs', 'rwr', 'khop', 'bfs-complete', 'minerva'], default='rwr')
parser.add_argument('--seed', type=int, default=231327)
parser.add_argument('--beam-size', type=int, default=100, help='beam size for Minerva model decoding')
parser.add_argument('--debug', action='store_true')
parser.add_argument("--add-segment-embed", action="store_true", default=False)
parser.add_argument('--add-inverse-rels', action='store_true', help='add inverse relations for query tower')
parser.add_argument('--label-smoothing', type=float, default=0.0)
parser.add_argument('--hidden-dropout-prob', type=float, default=0.1)
parser.add_argument('--attention-probs-dropout-prob', type=float, default=0.1)
# Bert model args
args = parser.parse_args()
device = 'cuda' if args.cuda and torch.cuda.is_available() else 'cpu'
# set seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
dataset = SubGraphDataset(args.dataset_path, 'train', args.neigh_size, args.sampling_type, args.graph_connection, device, False, args)
my_collator_eval = MyCollatorCustom(dataset.triples_dataset.entity2id['<pad>'], dataset.triples_dataset.relation2id['<pad>'], include_edge=True, add_segment_embed=args.add_segment_embed)
val_dataset_head = LinkPredSubGraphDataset(args.dataset_path, args.split, 'head', args.sample_size, args.neigh_size, args.sampling_type, args.graph_connection, device, args)
val_dataloader_head = DataLoader(val_dataset_head, batch_size=16, collate_fn=my_collator_eval)
val_dataset_tail = LinkPredSubGraphDataset(args.dataset_path, args.split, 'tail', args.sample_size, args.neigh_size, args.sampling_type, args.graph_connection, device, args)
val_dataloader_tail = DataLoader(val_dataset_tail, batch_size=16, collate_fn=my_collator_eval)
if args.model_type == 'self-attn':
model = GraphTransformer(dataset.num_entities, dataset.num_relations, args).to(device)
else:
model = GraphTransformerCrossAttn(dataset.num_entities, dataset.num_relations, args).to(device)
model.load_state_dict(torch.load(args.ckpt_path), strict=False)
model.eval()
with torch.no_grad():
right_results = {} # head is corrupted
for batch_id, batch in enumerate(tqdm(val_dataloader_head)):
# read a batch of triples
# sub, rel, obj = batch['edge']
sub, rel, obj = batch['edge'][:, 0], batch['edge'][:, 1], batch['edge'][:, 2]
label = batch['label'].to(device)
# print('sub = {}'.format(sub))
# print('rel = {}'.format(rel))
# print('obj = {}'.format(obj))
# do the forward pass
if 'segment_ids' not in batch:
outputs = model(input_ids=batch['input_ids'].to(device), attention_mask=batch['attention_mask'].to(device), token_type_ids=batch['token_type_ids'].to(device), ent_masked_lm_labels=batch['ent_masked_lm_labels'].to(device), ent_len_tensor=batch['ent_len_tensor'].to(device), ent_ids=batch['ent_ids'].to(device), rel_ids=batch['rel_ids'].to(device))
else:
outputs = model(input_ids=batch['input_ids'].to(device), attention_mask=batch['attention_mask'].to(device), token_type_ids=batch['token_type_ids'].to(device), ent_masked_lm_labels=batch['ent_masked_lm_labels'].to(device), ent_len_tensor=batch['ent_len_tensor'].to(device), ent_ids=batch['ent_ids'].to(device), rel_ids=batch['rel_ids'].to(device), segment_ids=batch['segment_ids'].to(device))
pred = outputs['ent_logits'].squeeze(1)
b_range = torch.arange(pred.size()[0], device=device)
target_pred = pred[b_range, sub] # get score of tail ent in (sub, rel, obj) triple [batch]
pred = torch.where(label.byte(), -torch.ones_like(pred) * 10000000, pred) # [batch, num_ent] at indices which corr to 'sub' position, put a big negative no. AT other places it remains unchanged.
pred[b_range, sub] = target_pred # assign it back for the ground truth object
ranks = 1 + torch.argsort(torch.argsort(pred, dim=1, descending=True), dim=1, descending=False)[b_range, sub]
# print('ranks = {}'.format(ranks))
# print('target_pred = {}'.format(target_pred))
ranks = ranks.float()
right_results['count'] = torch.numel(ranks) + right_results.get('count', 0.0)
right_results['mr'] = torch.sum(ranks).item() + right_results.get('mr', 0.0)
right_results['mrr'] = torch.sum(1.0/ranks).item() + right_results.get('mrr', 0.0)
# print(right_results)
for k in range(10):
right_results['hits@{}'.format(k+1)] = torch.numel(ranks[ranks <= (k+1)]) + right_results.get('hits@{}'.format(k+1), 0.0)
if args.debug and batch_id>100:
break
with torch.no_grad():
left_results = {} # tail is corrupted
for batch_id, batch in enumerate(tqdm(val_dataloader_tail)):
# read a batch of triples
# sub, rel, obj = batch['edge']
sub, rel, obj = batch['edge'][:, 0], batch['edge'][:, 1], batch['edge'][:, 2]
label = batch['label'].to(device)
# do the forward pass
if 'segment_ids' not in batch:
outputs = model(input_ids=batch['input_ids'].to(device), attention_mask=batch['attention_mask'].to(device), token_type_ids=batch['token_type_ids'].to(device), ent_masked_lm_labels=batch['ent_masked_lm_labels'].to(device), ent_len_tensor=batch['ent_len_tensor'].to(device), ent_ids=batch['ent_ids'].to(device), rel_ids=batch['rel_ids'].to(device))
else:
outputs = model(input_ids=batch['input_ids'].to(device), attention_mask=batch['attention_mask'].to(device), token_type_ids=batch['token_type_ids'].to(device), ent_masked_lm_labels=batch['ent_masked_lm_labels'].to(device), ent_len_tensor=batch['ent_len_tensor'].to(device), ent_ids=batch['ent_ids'].to(device), rel_ids=batch['rel_ids'].to(device), segment_ids=batch['segment_ids'].to(device))
pred = outputs['ent_logits'].squeeze(1)
b_range = torch.arange(pred.size()[0], device=device)
target_pred = pred[b_range, obj] # get score of tail ent in (sub, rel, obj) triple [batch]
pred = torch.where(label.byte(), -torch.ones_like(pred) * 10000000, pred) # [batch, num_ent] at indices which corr to 'obj' position, put a big negative no. AT other places it remains unchanged.
pred[b_range, obj] = target_pred # assign it back for the ground truth object
ranks = 1 + torch.argsort(torch.argsort(pred, dim=1, descending=True), dim=1, descending=False)[b_range, obj]
# print('ranks = {}'.format(ranks))
ranks = ranks.float()
left_results['count'] = torch.numel(ranks) + left_results.get('count', 0.0)
left_results['mr'] = torch.sum(ranks).item() + left_results.get('mr', 0.0)
left_results['mrr'] = torch.sum(1.0/ranks).item() + left_results.get('mrr', 0.0)
for k in range(10):
left_results['hits@{}'.format(k+1)] = torch.numel(ranks[ranks <= (k+1)]) + left_results.get('hits@{}'.format(k+1), 0.0)
if args.debug and batch_id>100:
break
results = get_combined_results(left_results, right_results)
print('[{}]: MRR: Tail : {:.5}, Head : {:.5}, Avg : {:.5}'.format(args.split, results['left_mrr'], results['right_mrr'], results['mrr']))
print(results)
if __name__=='__main__':
main()