-
Notifications
You must be signed in to change notification settings - Fork 0
/
main2.py
151 lines (122 loc) · 5.62 KB
/
main2.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
# main file for CNN based retrieval model
import os
import numpy as np
import torch
from torch.optim import Adam
from torch.utils.data import DataLoader
from torch.utils.data.sampler import SubsetRandomSampler
from conv_knrm import *
from data_reader import DataReader
from embedding_loader import create_embedding_matrix
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def train(model, data, n_epochs = 10, mini_batch_size = 16):
# Use 80% data from training and 20% for validation
data_size = len(data)
train_size = int(0.8 * len(data))
val_size = data_size - train_size
train_set, val_set = np.arange(0, train_size), np.arange(train_size, data_size)
train_size = len(train_set) / mini_batch_size
val_size = len(val_set) / mini_batch_size
train_loader = DataLoader(data,
batch_size = mini_batch_size,
sampler = SubsetRandomSampler(train_set),
num_workers = 1)
val_loader = DataLoader(data,
batch_size = mini_batch_size,
sampler = SubsetRandomSampler(val_set),
num_workers = 1)
# Create loss function and optimizer
criterion = HingeLoss()
optimizer = Adam(model.parameters(), lr=0.001)
chk_path = 'convknrm_chk.pth'
best_loss = 9999.
for epoch in range(n_epochs):
# Training
train_loss = 0.0
for batch_idx, (query, pos, neg) in enumerate(train_loader):
query = torch.stack(query, dim = 1).to(device)
pos = torch.stack(pos, dim = 1).to(device)
neg = torch.stack(neg, dim = 1).to(device)
# Forward pass
pos_score = model([query, pos]).view(-1)
neg_score = model([query, neg]).view(-1)
# Compute loss and backpropagate
optimizer.zero_grad()
loss = criterion(pos_score, neg_score)
loss.backward()
optimizer.step()
train_loss += loss.item()
if batch_idx % 100 == 0:
msg = 'Epoch: {0} Training Batch: {1} Loss: {2:.2f} '.format(epoch, batch_idx, loss.item())
print(msg, end = '\r')
# Validation
val_loss = 0.0
with torch.no_grad():
for batch_idx, (query, pos, neg) in enumerate(val_loader):
query = torch.stack(query, dim = 1).to(device)
pos = torch.stack(pos, dim = 1).to(device)
neg = torch.stack(neg, dim = 1).to(device)
pos_score = model([query, pos]).view(-1)
neg_score = model([query, neg]).view(-1)
# Compute loss
loss = criterion(pos_score, neg_score)
val_loss += loss.item()
if batch_idx % 100 == 0:
msg = 'Epoch: {0} Validation Batch: {1} Loss: {2:.2f} '.format(epoch, batch_idx, loss.item())
print(msg, end = '\r')
# Display average loss values
train_loss /= train_size
val_loss /= val_size
msg = '==> Epoch: {0} Avg. Training Loss: {1:.2f} Avg. Validation Loss: {2:.2f}'.format(epoch, train_loss, val_loss)
print(msg, end = '\n')
# Save weights as checkpoints
if val_loss < best_loss:
best_loss = val_loss
print('Saving checkpoint at ', chk_path)
torch.save(model.state_dict(), chk_path)
print('Training complete!')
return model
def test(model, query, pos_doc, neg_doc, weights_path = 'convknrm_chk.pth'):
if weights_path is not None:
model.load_state_dict(torch.load(weights_path))
x1 = torch.tensor([query], dtype = torch.long, device = device)
x2 = torch.tensor([pos_doc], dtype = torch.long, device = device)
x3 = torch.tensor([neg_doc], dtype = torch.long, device = device)
y1 = model([x1, x2]).squeeze()
y2 = model([x1, x3]).squeeze()
return (y1, y2)
def main():
# Default settings
vocab_size = 20000
max_data_count = 10000
pretrained_emb_path = './emb_weights.npy'
data_path = './triples.train.small.tsv'
# Load data
data_reader = DataReader(data_path = data_path,
data_count = max_data_count,
vocab_size = vocab_size)
data = data_reader.load_data()
text_data = data_reader.unprocessed_data
print('Data loaded successfully!')
# Load pretrained embedding matrix (20000 X 300)
if not os.path.exists(pretrained_emb_path):
create_embedding_matrix(data_reader.word2idx_dict, data_reader.vocab, save_path = pretrained_emb_path)
with open(pretrained_emb_path, 'rb') as fp:
emb_weights = np.load(fp)
print('Embeddings loaded successfully!')
# Create a model
model = ConvKNRM(vocab_size, emb_weights, n_filters = 64)
model = model.to(device)
params = filter(lambda p: p.requires_grad, model.parameters())
print('Trainable parameters: {}'.format(sum([np.prod(p.size()) for p in params])))
print('Model created!')
# Train the model
train(model, data, n_epochs = 50)
# Test the model
idx = 5000
x1, x2, x3 = data[idx]
scores = test(model, x1, x2, x3)
print('Query: {0} \n\nPositive Document: {1} \n\nNegative Document: {2}\n\n'.format(*text_data[idx]))
print('First document score = {0:.2f} \nSecond document score = {1:.2f}'.format(*scores))
if __name__ == '__main__':
main()