-
Notifications
You must be signed in to change notification settings - Fork 2
/
drl_rec.py
363 lines (340 loc) · 19.2 KB
/
drl_rec.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
#!encoding=utf-8
"""
Distilled reinforcement learning framework for recommendation (DRL-Rec)
@2021-01-04
modriczhang
"""
import os
import sys
import math
import time
import random
import hashlib
import datetime
import subprocess
import numpy as np
import tensorflow.compat.v1 as tf
from layer_util import *
from data_reader import DataReader
from hyper_param import param_dict as pd
from replay_buffer import RB
tf.disable_eager_execution()
tf.logging.set_verbosity(tf.logging.ERROR)
g_working_mode = 'local_train'
g_training = False
# replay buffer
g_rb = None
# data reader
g_dr = DataReader(pd['batch_size'])
g_batch_counter = 0
g_topk_pool = [[], [], [], []]
class DRL4Rec(object):
def __init__(self, global_step):
self.global_step = global_step
# cumulative loss
self.teacher_loss_val, self.student_loss_val, self.topk_loss_val = 0.0, 0.0, 0.0
# placeholder
self.sph_user = tf.sparse_placeholder(tf.int32, name='sph_user')
self.sph_doc = tf.sparse_placeholder(tf.int32, name='sph_doc')
self.sph_con = tf.sparse_placeholder(tf.int32, name='sph_con')
self.ph_reward = tf.placeholder(tf.float32, name='ph_reward')
self.ph_seqlen = tf.placeholder(tf.int32, name='seq_len')
self.seqlen = tf.identity(self.ph_seqlen)
print('\n======\nbuilding teacher main Q-network...')
self.TMH, self.TMQ = self.build_q_network('teacher', 'main', pd['teacher_feat_dim'])
print('\n======\nbuilding teacher target Q-network...')
self.TTH, self.TTQ = self.build_q_network('teacher', 'target', pd['teacher_feat_dim'])
print('\n======\nbuilding student main Q-network...')
self.SMH, self.SMQ = self.build_q_network('student', 'main', pd['student_feat_dim'])
print('\n======\nbuilding student target Q-network...')
self.STH, self.STQ = self.build_q_network('student', 'target', pd['student_feat_dim'])
teacher_var, student_var = 0, 0
vs = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='')
for vi in vs:
dims = vi.get_shape().as_list()
if 'teacher' in vi.name:
teacher_var += dims[0] if len(dims) == 1 else dims[0] * dims[1]
elif 'student' in vi.name:
student_var += dims[0] if len(dims) == 1 else dims[0] * dims[1]
else:
pass
print('\n\n=====Teacher Network: ' + str(teacher_var))
print('=====Student Network: ' + str(student_var))
print('=====Compress ratio: ' + str(1.0 - 1.0 * student_var / teacher_var))
# teacher rl loss
self.TNQ = tf.stop_gradient(self.TTQ)
self.TNQ = tf.concat([self.TNQ[:, 1:], tf.zeros_like(self.TNQ)[:, :1]], axis=1)
self.TYT = tf.reshape(self.ph_reward, [-1]) + tf.scalar_mul(tf.constant(pd['rl_gamma']),
tf.reshape(self.TNQ, [-1]))
self.TTDE = tf.square(self.TYT - tf.reshape(self.TMQ, [-1]))
self.teacher_loss = tf.reduce_mean(self.TTDE)
# teacher optimization
vs = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='teacher/main')
vs.extend(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='teacher/embedding'))
self.t_grads = tf.clip_by_global_norm(tf.gradients(self.teacher_loss, vs), pd['grad_clip'])[0]
with tf.variable_scope('opt_teacher'):
optimizer = tf.train.AdamOptimizer(pd['teacher_lr'])
self.teacher_opt = optimizer.apply_gradients(zip(self.t_grads, vs), global_step=global_step)
m_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="teacher/main")
t_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="teacher/target")
alpha = pd['double_networks_sync_step']
self.teacher_sync_op = [tf.assign(t, (1.0 - alpha) * t + alpha * m) for t, m in zip(t_params, m_params)]
# student rl loss
self.SNQ = tf.stop_gradient(self.STQ)
self.SNQ = tf.concat([self.SNQ[:, 1:], tf.zeros_like(self.SNQ)[:, :1]], axis=1)
self.SYT = tf.reshape(self.ph_reward, [-1]) + tf.scalar_mul(tf.constant(pd['rl_gamma']),
tf.reshape(self.SNQ, [-1]))
self.STDE = tf.square(self.SYT - tf.reshape(self.SMQ, [-1]))
self.student_rl_loss = tf.reduce_mean(self.STDE)
# confidence
maxe = tf.reduce_max(tf.reshape(self.TTDE, [-1])) + 1e-6
self.guide_weights = tf.stop_gradient(1.0 - self.TTDE / maxe)
# student guide loss
logp = tf.nn.log_softmax(logits=tf.stop_gradient(self.TMQ) / pd['distill_temperature'], dim=-1)
logq = tf.nn.log_softmax(logits=self.SMQ, dim=-1)
self.student_kd_loss = tf.reduce_mean(self.kl_divergence(logp, logq))
# student hint loss
if pd['enable_confidence']:
self.hint_loss = tf.reduce_mean(tf.multiply(tf.stop_gradient(self.guide_weights),
tf.reduce_mean(tf.square(
tf.reshape(self.SMH - tf.stop_gradient(self.TMH), [-1])))))
else:
self.hint_loss = tf.reduce_mean(tf.square(tf.reshape(self.SMH - tf.stop_gradient(self.TMH), [-1])))
loss_weights = [0.4, 0.3, 0.3]
if pd['enable_distill']:
self.student_loss = loss_weights[0] * self.student_rl_loss + \
loss_weights[1] * self.student_kd_loss + \
loss_weights[2] * self.hint_loss
else:
self.student_loss = loss_weights[0] * self.student_rl_loss
vs = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='student/main')
vs.extend(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='student/embedding'))
self.s_grads = tf.clip_by_global_norm(tf.gradients(self.student_loss, vs), pd['grad_clip'])[0]
with tf.variable_scope('opt_student'):
optimizer = tf.train.AdamOptimizer(pd['student_lr'])
self.student_opt = optimizer.apply_gradients(zip(self.s_grads, vs), global_step=global_step)
self.topk_grads = tf.clip_by_global_norm(tf.gradients(self.hint_loss, vs), pd['grad_clip'])[0]
with tf.variable_scope('opt_topk_distill'):
optimizer = tf.train.AdamOptimizer(pd['topk_distill_lr'])
self.topk_distill_opt = optimizer.apply_gradients(zip(self.topk_grads, vs), global_step=global_step)
m_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="student/main")
t_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="student/target")
self.student_sync_op = [tf.assign(t, (1.0 - alpha) * t + alpha * m) for t, m in zip(t_params, m_params)]
def kl_divergence(self, logp, logq):
p = tf.exp(logp)
return tf.reduce_sum(p * logp, axis=-1) - tf.reduce_sum(p * logq, axis=-1)
def field_interact(self, fields):
global g_training
qkv = tf.layers.dropout(fields, rate=pd['dropout'], training=g_training)
with tf.variable_scope('fi'):
return multihead_attention(queries=qkv,
keys=qkv,
values=qkv,
num_heads=pd['head_num'],
dropout_rate=pd['dropout'],
training=g_training,
causality=False,
scope='mha')
def build_embedding_layer(self, sub_net, scope, feat_dim):
with tf.variable_scope(sub_net, reuse=tf.AUTO_REUSE):
feat_dict = get_embeddings(g_dr.unique_feature_num(),
feat_dim,
scope='embedding',
zero_pad=False)
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
user_embed = tf.nn.embedding_lookup_sparse(feat_dict,
self.sph_user,
sp_weights=None,
partition_strategy='div',
combiner='mean')
user_fields = tf.reshape(user_embed, shape=[-1, pd['user_field_num'], feat_dim])
doc_embed = tf.nn.embedding_lookup_sparse(feat_dict,
self.sph_doc,
sp_weights=None,
partition_strategy='div',
combiner='mean')
doc_fields = tf.reshape(doc_embed, shape=[-1, pd['doc_field_num'], feat_dim])
con_embed = tf.nn.embedding_lookup_sparse(feat_dict,
self.sph_con,
sp_weights=None,
partition_strategy='div',
combiner='mean')
con_fields = tf.reshape(con_embed, shape=[-1, pd['con_field_num'], feat_dim])
return tf.concat([user_fields, doc_fields, con_fields], axis=1)
def build_q_network(self, sub_net, scope, feat_dim):
global g_training
fields = self.build_embedding_layer(sub_net, scope, feat_dim)
with tf.variable_scope(sub_net, reuse=tf.AUTO_REUSE):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
fields_num = pd['user_field_num'] + pd['doc_field_num'] + pd['con_field_num']
finp = tf.reshape(fields, shape=[-1, fields_num, feat_dim])
# rnn seq embedding
seq_embed = tf.reshape(self.field_interact(finp), (-1, self.seqlen, fields_num * feat_dim))
gru = tf.nn.rnn_cell.GRUCell(seq_embed.get_shape().as_list()[-1])
drop = tf.nn.rnn_cell.DropoutWrapper(gru, output_keep_prob=1.0 - pd['dropout'] if g_training else 1.)
cell = tf.nn.rnn_cell.MultiRNNCell([drop for _ in range(pd['rnn_layer'])])
state_embed, last_states = tf.nn.dynamic_rnn(cell=cell, dtype=tf.float32, inputs=seq_embed,
time_major=False)
print('state_embed.shape:', state_embed.shape)
state_dim = state_embed.get_shape().as_list()[-1]
mlp_dims = [state_dim / 2, state_dim / 4, pd['hint_dim']]
fc = state_embed
for i in range(len(mlp_dims)):
fc = tf.layers.dense(fc, mlp_dims[i], name='mlp_' + str(i + 1), activation=tf.nn.tanh,
reuse=tf.AUTO_REUSE)
fc = tf.layers.dropout(fc, rate=pd['dropout'], training=g_training)
hint_layer = fc
print('hint_layer.shape:', hint_layer.shape)
q = tf.reshape(tf.layers.dense(hint_layer, 1), (-1, self.seqlen))
print('q.shape:', q.shape)
return hint_layer, q
def teacher_learn(self, sess, ph_dict):
loss, _ = sess.run([self.teacher_loss, self.teacher_opt], feed_dict={self.ph_reward: ph_dict['reward'],
self.ph_seqlen: pd['rnn_seq_len'],
self.sph_user: ph_dict['user'],
self.sph_doc: ph_dict['doc'],
self.sph_con: ph_dict['con']})
self.teacher_loss_val += loss
def student_learn(self, sess, ph_dict):
loss, _ = sess.run([self.student_loss, self.student_opt], feed_dict={self.ph_reward: ph_dict['reward'],
self.ph_seqlen: pd['rnn_seq_len'],
self.sph_user: ph_dict['user'],
self.sph_doc: ph_dict['doc'],
self.sph_con: ph_dict['con']})
self.student_loss_val += loss
def inference_candi_pool(self, sess, ph_dict):
tq, sq = sess.run([self.TMQ, self.SMQ], feed_dict={self.sph_user: ph_dict['user'],
self.sph_doc: ph_dict['doc'],
self.sph_con: ph_dict['con'],
self.ph_seqlen: 1})
return tq, sq
def topk_distill(self, sess, ph_dict):
loss, _ = sess.run([self.hint_loss, self.topk_distill_opt],
feed_dict={self.ph_reward: ph_dict['reward'],
self.ph_seqlen: 1,
self.sph_user: ph_dict['user'],
self.sph_doc: ph_dict['doc'],
self.sph_con: ph_dict['con']})
self.topk_loss_val += loss
def eval(self, sess, ph_dict):
tq, sq = sess.run([self.TMQ, self.SMQ], feed_dict={self.sph_user: ph_dict['user'],
self.sph_doc: ph_dict['doc'],
self.sph_con: ph_dict['con'],
self.ph_seqlen: pd['rnn_seq_len']})
return tq, sq
def handle(sess, net, sess_data):
def gen_sparse_tensor(fs):
global g_dr
kk, vv = [], []
for i in range(len(fs)):
ff = fs[i]
assert (isinstance(ff, set))
ff = list(ff)
for k in range(len(ff)):
kk.append(np.array([i, k], dtype=np.int32))
vv.append(ff[k])
return tf.SparseTensorValue(kk, vv, [len(fs), g_dr.unique_feature_num()])
global g_rb, g_topk_pool
g_rb.save(sess_data)
while g_rb.has_batch():
user, doc, con, rwd, rtn = g_rb.next_batch()
cps = pd['candi_pool_size']
g_topk_pool[0].extend(user)
g_topk_pool[1].extend(doc)
g_topk_pool[2].extend(con)
g_topk_pool[3].extend(rwd)
for i in range(len(g_topk_pool)):
g_topk_pool[i] = g_topk_pool[i][-cps:]
phd = {}
rec_user = np.array(user).reshape(pd['batch_size'] * pd['rnn_seq_len'] * pd['user_field_num'])
phd['user'] = gen_sparse_tensor(rec_user)
rec_doc = np.array(doc).reshape(pd['batch_size'] * pd['rnn_seq_len'] * pd['doc_field_num'])
phd['doc'] = gen_sparse_tensor(rec_doc)
rec_con = np.array(con).reshape(pd['batch_size'] * pd['rnn_seq_len'] * pd['con_field_num'])
phd['con'] = gen_sparse_tensor(rec_con)
phd['reward'] = rwd
global g_batch_counter, g_training
g_batch_counter += 1
if g_training:
net.teacher_learn(sess, phd)
net.student_learn(sess, phd)
if len(g_topk_pool[0]) == cps:
# eval all items in candidate pool
topk_user = np.array(g_topk_pool[0])
phd['user'] = gen_sparse_tensor(topk_user.reshape(-1))
topk_doc = np.array(g_topk_pool[1])
phd['doc'] = gen_sparse_tensor(topk_doc.reshape(-1))
topk_con = np.array(g_topk_pool[2])
phd['con'] = gen_sparse_tensor(topk_con.reshape(-1))
topk_rwd = np.array(g_topk_pool[3])
tq, sq = net.inference_candi_pool(sess, phd)
# select top-k items of teacher and student
tq = tq.reshape((1, -1))
sq = sq.reshape((1, -1))
arg_tq = np.argsort(-tq, axis=1).take(np.arange(pd['explore_filter_topk']))
arg_sq = np.argsort(-sq, axis=1).take(np.arange(pd['explore_filter_topk']))
ef_user = np.concatenate([topk_user.take(arg_tq, axis=0), topk_user.take(arg_sq, axis=0)], axis=0)
ef_doc = np.concatenate([topk_doc.take(arg_tq, axis=0), topk_doc.take(arg_sq, axis=0)], axis=0)
ef_con = np.concatenate([topk_con.take(arg_tq, axis=0), topk_con.take(arg_sq, axis=0)], axis=0)
ef_rwd = np.concatenate([topk_rwd.take(arg_tq, axis=0), topk_rwd.take(arg_sq, axis=0)], axis=0)
phd['user'] = gen_sparse_tensor(ef_user.reshape(-1))
phd['doc'] = gen_sparse_tensor(ef_doc.reshape(-1))
phd['con'] = gen_sparse_tensor(ef_con.reshape(-1))
phd['reward'] = ef_rwd
# top-k distillation
if pd['enable_distill'] and pd['enable_topk_distill']:
net.topk_distill(sess, phd)
if g_batch_counter % pd['double_networks_sync_freq'] == 0:
print('Run soft replacement for main networks and target networks...')
sess.run(net.teacher_sync_op)
sess.run(net.student_sync_op)
else:
tq, sq = net.eval(sess, phd)
tq = tq.reshape([-1])
sq = sq.reshape([-1])
global g_working_mode
for i in range(len(rtn)):
if 'local_predict' == g_working_mode:
print('%s %s %s' % (rwd[i], tq[i], sq[i]))
def local_run():
global_step = tf.train.get_or_create_global_step()
sess = tf.Session()
net = DRL4Rec(global_step)
saver = tf.train.Saver(max_to_keep=1)
g_init_op = tf.global_variables_initializer()
if os.path.exists('./ckpt'):
model_file = tf.train.latest_checkpoint('ckpt/')
saver.restore(sess, model_file)
else:
sess.run(g_init_op)
os.system('mkdir ckpt')
global g_batch_counter
for k in range(pd['num_epochs'] if g_training else 1):
if k > 0:
g_dr.load('sample.data')
data = g_dr.next()
while data is not None:
handle(sess, net, data)
data = g_dr.next()
if g_training and g_batch_counter % 10 == 0:
print(
'>>> epoch %d --- batch %d --- teacher net loss = %f --- student net loss = %f --- top-k distill loss = %f' % (
k, g_batch_counter, net.teacher_loss_val / (g_batch_counter + 1e-6),
net.student_loss_val / (g_batch_counter + 1e-6),
net.topk_loss_val / (g_batch_counter + 1e-6)))
saver.save(sess, 'ckpt/drl_rec.ckpt')
if __name__ == '__main__':
g_working_mode = 'local_train'
commander = {
'local_train': local_run,
'local_predict': local_run
}
if g_working_mode not in commander:
print('your working mode(%s) not recognized!!!' % g_working_mode)
sys.exit(1)
g_training = True if g_working_mode == 'local_train' else False
print('>>> working_model:', g_working_mode)
print('>>> is_training:', g_training)
g_dr.load('sample.data')
g_rb = RB(pd['batch_size'], pd['rnn_seq_len'], pd['user_field_num'], pd['doc_field_num'], pd['con_field_num'])
commander[g_working_mode]()