-
Notifications
You must be signed in to change notification settings - Fork 0
/
hyper_param.py
36 lines (35 loc) · 1.4 KB
/
hyper_param.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
#!encoding:utf-8
"""
Model Hyper Parameter Dict
@2020-02-24
by Modric Zhang
"""
param_dict = {
'booster_feat_dim': 16, # feature embedding dimension
'lite_feat_dim': 8, # feature embedding dimension
'user_field_num': 5, # number of user feature fields
'doc_field_num': 5, # number of doc feature fields
'con_field_num': 5, # number of context feature fields
'future_field_num': 2, # number of future feature fields
'num_epochs': 1, # training epoch
'batch_size': 256, # batch size
'booster_lr': 0.0002, # learning rate
'lite_lr': 0.0002, # learning rate
'hint_dim': 32, # learning rate
'dropout': 0.3, # learning rate of q/value network
'rl_gamma': 0.6, # gamma in rl
'grad_clip': 5, # grad clip
'booster_rnn_dim': 160, # rnn dimension of booster network
'lite_rnn_dim': 120, # rnn dimension of light network
'clk_seq_len': 10, # click sequence length
'rnn_seq_len': 10, # rnn sequence length
'head_num': 4, # head number for all self-attention units
'rnn_layer': 1, # RNN layer
'encoder_layer': 1, # encoder layer
'decoder_layer': 1, # decoder layer
'double_networks_sync_freq': 10, # sync frequency for both policy and value network
'double_networks_sync_step': 0.005, # sync step is designed with reference to deepmind
'enable_gan': True,
'enable_distill': True,
'enable_future': True,
}