-
Notifications
You must be signed in to change notification settings - Fork 26
/
Copy pathranknet.py
218 lines (199 loc) · 12.4 KB
/
ranknet.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
import os
import ipdb
import time
import models
import random
import pickle
import numpy as np
import argparse
import tensorflow as tf
import collections
from toy_ndcg import ndcg
from ranking_utils import calc_err
from convert_data_to_np_features import *
class RankNetTrainer:
def __init__(self, n_hidden, train_relevance_labels, train_query_ids, train_features, test_relevance_labels,
test_query_ids, test_features, vali_relevance_labels, vali_query_ids, vali_features):
self.train_query_ids = train_query_ids
self.train_relevance_labels = train_relevance_labels
self.train_features = train_features
self.train_unique_query_ids = np.unique(self.train_query_ids)
self.train_unique_query_ids_subset = [self.train_unique_query_ids[i] for i in range(0, 500)]
self.vali_query_ids = vali_query_ids
self.vali_relevance_labels = vali_relevance_labels
self.vali_features = vali_features
self.vali_unique_query_ids = np.unique(self.vali_query_ids)
self.vali_unique_query_ids_subset = [self.vali_unique_query_ids[i] for i in range(0, 500)]
self.test_query_ids = test_query_ids
self.test_relevance_labels = test_relevance_labels
self.test_features = test_features
self.unique_ids = np.unique(train_query_ids)
np.random.shuffle(self.unique_ids)
self.unique_ids_subset = [self.unique_ids[i] for i in range(0, 500)]
self.models_directory = os.path.join('..', 'models/ranknet/')
if not os.path.exists(self.models_directory):
os.makedirs(self.models_directory)
self.n_hidden = n_hidden
self.best_cost = float('inf')
self.best_ndcg = float('inf')
self.all_costs = list()
self.all_ndcg_scores = list()
self.all_full_ndcg_scores = list()
self.all_err_scores = list()
self.all_validation_costs = list()
self.all_validation_ndcg_scores = list()
self.all_validation_full_ndcg_scores = list()
self.all_validation_err_scores = list()
def train(self, learning_rate, n_layers, batch_size, lambdarank, unfactorized, factorized):
x = tf.placeholder("float", [None, models.N_FEATURES])
relevance_scores = tf.placeholder("float", [None, 1])
sorted_relevance_scores = tf.placeholder("float", [None, 1])
index_range = tf.placeholder("float", [None, 1])
lr = tf.placeholder("float", [])
query_indices = tf.placeholder("float", [None])
self.learning_rate = learning_rate
self.start_time = time.time()
if lambdarank:
self.filename = 'nn_lambdarank_%slayers_%shidden_lr%s' % (n_layers, self.n_hidden, ('%.0E' % self.learning_rate).replace('-', '_'))
cost, optimizer, score = models.lambdarank_deep(x, relevance_scores, sorted_relevance_scores, index_range,
self.learning_rate, self.n_hidden, n_layers)
elif unfactorized:
self.filename = 'nn_unfactorized_ranknet_%slayers_%shidden_lr%s' % (n_layers, self.n_hidden, ('%.0E' % self.learning_rate).replace('-', '_'))
cost, optimizer, score = models.default_ranknet(x, relevance_scores, self.learning_rate, self.n_hidden, n_layers)
elif factorized:
self.filename = 'nn_factorized_ranknet_%slayers_%shidden_lr%s' % (n_layers, self.n_hidden, ('%.0E' % self.learning_rate).replace('-', '_'))
cost, optimizer, score = models.deep_factorized_ranknet(x, relevance_scores, self.learning_rate, self.n_hidden, n_layers)
else:
raise('Need to specify if this model should be unfactorized, factorized, or use lambdarank!')
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
c_iter = 0
while True:
if c_iter % 10 == 0:
self.check_progress(sess, saver, cost, score, x, relevance_scores, c_iter, True)
c_iter += 1
indices = np.random.randint(1, len(self.train_features), batch_size)
# c_id = np.random.choice(self.unique_ids)
# indices = np.where(self.train_query_ids == c_id)[0]
if len(indices) > batch_size:
indices = indices[:batch_size]
if lambdarank:
optimizer(sess, {
x: np.array(self.train_features[indices], ndmin=2),
relevance_scores: np.array(self.train_relevance_labels[indices], ndmin=2).T,
lr: self.learning_rate,
query_indices: indices,
index_range: np.array([float(i) for i in range(0,len(indices))], ndmin=2).T,
sorted_relevance_scores: np.sort(np.array(self.train_relevance_labels[indices], ndmin=2)).T[::-1]
})
else:
optimizer(sess, {
x: np.array(self.train_features[indices], ndmin=2),
relevance_scores: np.array(self.train_relevance_labels[indices], ndmin=2).T,
lr: self.learning_rate,
query_indices: indices
})
def check_progress(self, sess, saver, cost, score, x, relevance_scores, c_iter, save_data=True):
train_avg_cost, train_avg_err, train_avg_ndcg, train_avg_full_ndcg = self.check_scores(cost,
self.train_features,
self.train_query_ids,
self.train_relevance_labels,
relevance_scores, score, sess,
self.train_unique_query_ids_subset, x)
vali_avg_cost, vali_avg_err, vali_avg_ndcg, vali_avg_full_ndcg = self.check_scores(cost,
self.vali_features,
self.vali_query_ids,
self.vali_relevance_labels,
relevance_scores, score, sess,
self.vali_unique_query_ids_subset, x)
print('{} -- Train Cost: {:10f} NDCG: {:9f} ({:9f}) ERR: {:9f} -- Validation Cost: {:10f} NDCG: {:9f} ({:9f}) ERR: {:9f} -- {:9f} s'.format(
c_iter, train_avg_cost, train_avg_ndcg, train_avg_full_ndcg, train_avg_err, vali_avg_cost, vali_avg_ndcg, vali_avg_full_ndcg, vali_avg_err, time.time() - self.start_time))
self.all_costs.append(train_avg_cost)
self.all_full_ndcg_scores.append(train_avg_full_ndcg)
self.all_ndcg_scores.append(train_avg_ndcg)
self.all_err_scores.append(train_avg_err)
self.all_validation_costs.append(vali_avg_cost)
self.all_validation_full_ndcg_scores.append(vali_avg_full_ndcg)
self.all_validation_ndcg_scores.append(vali_avg_ndcg)
self.all_validation_err_scores.append(vali_avg_err)
if self.all_validation_costs[-1] < self.best_cost:
self.best_cost = self.all_validation_costs[-1]
saver.save(sess, os.path.join(self.models_directory, self.filename + '_best_validation_cost'))
if self.all_ndcg_scores[-1] < self.best_ndcg:
self.best_ndcg = self.all_ndcg_scores[-1]
saver.save(sess, os.path.join(self.models_directory, self.filename + '_best_validation_ndcg'))
if save_data:
saver.save(sess, os.path.join(self.models_directory, self.filename + '_most_recent'))
pickle.dump(self.all_costs, open(os.path.join(self.models_directory, self.filename + '_costs.p'), 'wb'))
pickle.dump(self.all_ndcg_scores, open(os.path.join(self.models_directory, self.filename + '_ndcg_scores.p'), 'wb'))
pickle.dump(self.all_full_ndcg_scores, open(os.path.join(self.models_directory, self.filename + '_full_ndcg_scores.p'), 'wb'))
pickle.dump(self.all_err_scores, open(os.path.join(self.models_directory, self.filename + '_err_scores.p'), 'wb'))
pickle.dump(self.all_validation_costs, open(os.path.join(self.models_directory, self.filename + '_validation_costs.p'), 'wb'))
pickle.dump(self.all_validation_ndcg_scores, open(os.path.join(self.models_directory, self.filename + '_validation_ndcg_scores.p'), 'wb'))
pickle.dump(self.all_validation_full_ndcg_scores, open(os.path.join(self.models_directory, self.filename + '_validation_full_ndcg_scores.p'), 'wb'))
pickle.dump(self.all_validation_err_scores, open(os.path.join(self.models_directory, self.filename + '_validation_err_scores.p'), 'wb'))
return train_avg_cost, train_avg_ndcg, train_avg_err, vali_avg_cost, vali_avg_err, vali_avg_ndcg
def check_scores(self, cost, features, query_ids, relevance_labels, relevance_scores, score, sess,
unique_query_ids, x):
costs = list()
ndcg_scores = list()
full_ndcg_scores = list()
err_scores = list()
assert len(unique_query_ids) > 0
for c_id in unique_query_ids:
query_indices = np.where(query_ids == c_id)[0]
c_cost = sess.run(cost, feed_dict={
x: np.array(features[query_indices], ndmin=2),
relevance_scores: np.array(relevance_labels[query_indices], ndmin=2).T })
predicted_score = score(sess, {
x: np.array(features[query_indices], ndmin=2),
relevance_scores: np.array(relevance_labels[query_indices], ndmin=2).T })
pred_query_type = np.dtype(
[('predicted_scores', predicted_score.dtype),
('query_int', query_indices.dtype)])
pred_query = np.empty(len(predicted_score), dtype=pred_query_type)
pred_query['predicted_scores'] = np.reshape(predicted_score, [-1])
pred_query['query_int'] = query_indices
scored_pred_query = np.sort(pred_query, order='predicted_scores')[::-1]
costs.append(c_cost)
ndcg_scores.append(ndcg(relevance_labels[scored_pred_query['query_int']]))
full_ndcg_scores.append(ndcg(relevance_labels[scored_pred_query['query_int']], top_ten=False))
err_scores.append(calc_err(relevance_labels[scored_pred_query['query_int']]))
avg_cost = sum(costs) / len(costs)
avg_ndcg = np.mean(np.array(ndcg_scores))
avg_full_ndcg = np.mean(np.array(full_ndcg_scores))
avg_err = np.mean(np.array(err_scores))
return avg_cost, avg_err, avg_ndcg, avg_full_ndcg
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--lr', type=float, help='learning rate')
parser.add_argument('--batch', type=int, help='batch size')
parser.add_argument('--n_hidden', type=int, help='n hidden units')
parser.add_argument('--n_layers', type=int, help='n layers')
parser.add_argument('--lambdarank', action='store_true')
parser.add_argument('--unfactorized', action='store_true')
parser.add_argument('--factorized', action='store_true')
args = parser.parse_args()
np_train_file_directory = os.path.join('..', 'data/np_train_files')
train_relevance_labels = np.load(os.path.join(np_train_file_directory, LABEL_LIST + '.npy'))
train_query_ids = np.load(os.path.join(np_train_file_directory, QUERY_IDS + '.npy'))
train_features = np.load(os.path.join(np_train_file_directory, FEATURES + '.npy'))
np_test_file_directory = os.path.join('..', 'data/np_test_files')
test_relevance_labels = np.load(os.path.join(np_test_file_directory, LABEL_LIST + '.npy'))
test_query_ids = np.load(os.path.join(np_test_file_directory, QUERY_IDS + '.npy'))
test_features = np.load(os.path.join(np_test_file_directory, FEATURES + '.npy'))
np_vali_file_directory = os.path.join('..', 'data/np_vali_files')
vali_relevance_labels = np.load(os.path.join(np_vali_file_directory, LABEL_LIST + '.npy'))
vali_query_ids = np.load(os.path.join(np_vali_file_directory, QUERY_IDS + '.npy'))
vali_features = np.load(os.path.join(np_vali_file_directory, FEATURES + '.npy'))
learning_rate = 1e-5 if args.lr is None else args.lr
network_desc = 'unfactorized'
if args.factorized:
network_desc = 'factorized'
elif args.lambdarank:
network_desc = 'lambdarank'
print('Training a %s network, learning rate %f, n_hidden %s, n_layers %s' % (network_desc, learning_rate, args.n_hidden, args.n_layers))
trainer = RankNetTrainer(args.n_hidden, train_relevance_labels, train_query_ids, train_features, test_relevance_labels,
test_query_ids, test_features, vali_relevance_labels, vali_query_ids, vali_features)
trainer.train(learning_rate, args.n_layers, args.batch, args.lambdarank, args.unfactorized, args.factorized)