-
Notifications
You must be signed in to change notification settings - Fork 5
/
generator_prune.py
178 lines (142 loc) · 8.98 KB
/
generator_prune.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
import tensorflow as tf
import ops
import numpy as np
import config
# import pandas as pd
# this class is used for true system, only predict but no evaluation
class NextItNet_Decoder:
def __init__(self, model_para):
self.model_para = model_para
self.embedding_width = model_para['dilated_channels']
self.taskID= model_para['taskID']
self.allitem_embeddings = tf.get_variable('allitem_embeddings',
[model_para['bigemb'], self.embedding_width ],
initializer=tf.truncated_normal_initializer(stddev=0.02))
self.allitem_embeddings_out = tf.get_variable("softmax_w_{}".format(self.taskID),
[model_para['target_item_size'], self.embedding_width ],
initializer=tf.truncated_normal_initializer(stddev=0.02),
regularizer=tf.contrib.layers.l2_regularizer(0.02))
def train_graph(self, ispre=True):
model_para = self.model_para
self.itemseq_input = tf.placeholder('int32',
[None, None], name='itemseq_input')
if ispre==True:
self.dilate_input = self.model_graph(self.itemseq_input, train=True)
else:
self.dilate_input = self.model_graph(self.itemseq_input, train=True,ispre=False)
def model_graph(self, itemseq_input, train=True, ispre=True):
model_para = self.model_para
context_seq = itemseq_input[:, 0:-1]
# label_seq = itemseq_input[:, 1:]
self.context_embedding = tf.nn.embedding_lookup(self.allitem_embeddings,
context_seq, name="context_embedding")
# positional embedding
if self.model_para['has_positionalembedding']:
pos_emb = self.embedding(
tf.tile(tf.expand_dims(tf.range(tf.shape(context_seq)[1]), 0),
[tf.shape(itemseq_input)[0], 1]),
max_position=model_para['max_position'],
num_units=self.embedding_width,
zero_pad=False,
scale=False,
l2_reg=0.0,
scope="dec_pos",
with_t=False
)
# dilate_input = tf.concat([self.context_embedding, pos_emb], -1)
dilate_input =self.context_embedding+pos_emb
else:
dilate_input = self.context_embedding
# dilate_input = self.context_embedding
residual_channels = dilate_input.get_shape().as_list()[-1]
for layer_id, dilation in enumerate(model_para['dilations']):
if ispre == True:
dilate_input = ops.nextitnet_residual_block_withmask_pre_beforeln(dilate_input, dilation,
layer_id, residual_channels,
model_para['kernel_size'], self.taskID,
causal=True, train=train)
else:
dilate_input = ops.nextitnet_residual_block_withmask_fine_beforeln(dilate_input, dilation,
layer_id, residual_channels,
model_para['kernel_size'], self.taskID,
causal=True, train=train)
return dilate_input
#saving important weights
def save_impwei(self,mask_var,weight, curtaskID, reuse=False):
if reuse:
tf.get_variable_scope().reuse_variables()
init_zeros = tf.zeros_initializer()
trainable_vars = tf.trainable_variables()
self.mask_val_list_task = []
kernel_num = 1 * self.model_para['kernel_size'] * self.embedding_width * self.embedding_width
# cutoff_rank_task_remain = config.maskp_task1 * config.maskp_task2 * config.maskp_task3* (1-config.maskp_task4)
if curtaskID == config.taskID_1st:
cutoff = 1 - config.maskp_task1
elif curtaskID == config.taskID_2nd:
cutoff = config.maskp_task1 * (1 - config.maskp_task2)
elif curtaskID == config.taskID_3rd:
cutoff = config.maskp_task1 * config.maskp_task2 * (1 - config.maskp_task3)
elif curtaskID == config.taskID_4th:
cutoff = config.maskp_task1 * config.maskp_task2 * config.maskp_task3 * (
1 - config.maskp_task4)
elif curtaskID == config.taskID_5th:
cutoff = config.maskp_task1 * config.maskp_task2 * config.maskp_task3 * config.maskp_task4 * (
1 - config.maskp_task5)
elif curtaskID == config.taskID_6th:
cutoff = config.maskp_task1 * config.maskp_task2 * config.maskp_task3 * config.maskp_task4 * config.maskp_task5 * (
1 - config.maskp_task6)
cutoff_rank_task_remain = cutoff
cutoff_rank = tf.cast(cutoff_rank_task_remain* kernel_num, tf.int32)
graph = tf.get_default_graph()
with tf.variable_scope("mask_filter", reuse=tf.AUTO_REUSE):
for layer_id, dilation in enumerate(self.model_para['dilations']):
mask_name = "mask_val_layer_{}_{}".format(layer_id, dilation)
mask_name_layid_dilation="_{}_{}".format(layer_id, dilation)
mask_lay_dilation=[v for v in mask_var if v.name.find(mask_name_layid_dilation) != -1]# mask according to layer_id and dilation
frozen_matrix_conv1 = tf.zeros_like(weight[0])
frozen_matrix_conv2 = tf.zeros_like(weight[0])
for index in xrange(curtaskID-config.taskID_1st):
taskid=config.taskID_1st+index
mask_task=[v for v in mask_lay_dilation if v.name.find(str(taskid)) != -1]
mask_task_conv1=[v for v in mask_task if v.name.find("conv1") != -1]
mask_task_conv2 = [v for v in mask_task if v.name.find("conv2") != -1]
frozen_matrix_conv1+=mask_task_conv1[0]
frozen_matrix_conv2 += mask_task_conv2[0]
dilated_conv1 = [v for v in weight if
v.name.find("conv1") != -1 and v.name.find(mask_name_layid_dilation) != -1][0]
dilated_conv2 = [v for v in weight if
v.name.find("conv2") != -1 and v.name.find(mask_name_layid_dilation) != -1][0]
mask_conv1_ = tf.abs(frozen_matrix_conv1 - 1)
dilated_conv1_norm = tf.abs(dilated_conv1*mask_conv1_)
dilated_conv1_onedim = tf.reshape(dilated_conv1_norm, [kernel_num])
top_k_dilated_conv1 = tf.nn.top_k(dilated_conv1_onedim, cutoff_rank + 1).values[cutoff_rank]
one = tf.ones_like(dilated_conv1_norm)
zero = tf.zeros_like(dilated_conv1_norm)
mask_dilated_conv1 = tf.where(dilated_conv1_norm < top_k_dilated_conv1, x=zero, y=one)
mask_conv2_ = tf.abs(frozen_matrix_conv2 - 1)
dilated_conv2_norm = tf.abs(dilated_conv2*mask_conv2_)
dilated_conv2_onedim = tf.reshape(dilated_conv2_norm, [kernel_num])
top_k_dilated_conv2 = tf.nn.top_k(dilated_conv2_onedim, cutoff_rank + 1).values[
cutoff_rank] # e.g., 2.3
mask_dilated_conv2 = tf.where(dilated_conv2_norm < top_k_dilated_conv2, x=zero, y=one)
# mask_dilated_conv2 = tf.where(dilated_conv2_norm < top_k_dilated_conv2, x=one, y=zero)
self.mask_val_list_task.append(mask_dilated_conv1)
self.mask_val_list_task.append(mask_dilated_conv2)
def embedding(self, inputs, max_position, num_units, zero_pad=True, scale=True, l2_reg=0.0, scope="embedding",
with_t=False):
with tf.variable_scope(scope):
lookup_table = tf.get_variable('lookup_table_position',
dtype=tf.float32,
shape=[max_position, num_units],
# initializer=tf.contrib.layers.xavier_initializer(),
regularizer=tf.contrib.layers.l2_regularizer(l2_reg))
if zero_pad:
lookup_table = tf.concat((tf.zeros(shape=[1, num_units]),
lookup_table[1:, :]), 0)
outputs = tf.nn.embedding_lookup(lookup_table, inputs)
if scale:
outputs = outputs * (num_units ** 0.5)
if with_t:
return outputs, lookup_table
else:
return outputs