/
tree_tf_op.py
243 lines (178 loc) · 8.59 KB
/
tree_tf_op.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
import tensorflow as tf
import numpy as np
import sys
sys.path.append("..")
import global_vars as gv
import os
DEVICE = '/gpu:1'
hdir = os.getenv('HOME')
imgs_shape = [gv.BATCH_SZ, gv.n_rows, gv.n_cols, gv.n_input_channels]
map_prod = np.prod(gv.map_sz)
with tf.device(DEVICE):
sess = tf.InteractiveSession()
tf_op = tf.load_op_library('cuda_op_kernel.so')
##################### set / load vars
set_var_int32 = tf.placeholder(tf.int32, shape=[None])
set_var_int8 = tf.placeholder(tf.int8, shape=[None])
gm_var_nms = ['board', 'valid_mv_map_internal']
gm_var_placeholders = ['set_var_int8']*2
gm_vars = {}; set_gm_vars = {}
def return_vars():
v = {}
for var in gm_var_nms:
exec('v["%s"] = sess.run(gm_vars["%s"])' % (var, var))
return v
def set_vars(v):
for var, placeholder in zip(gm_var_nms, gm_var_placeholders):
exec('sess.run(set_gm_vars["%s"], feed_dict={%s: v["%s"].ravel()})' % (var, placeholder, var))
########################
def tf_pearsonr(val, val_target_nmean):
val_nmean = val - tf.reduce_mean(val)
val_target_nmean = val_target - tf.reduce_mean(val_target)
val_std = tf.sqrt(tf.reduce_sum(val_nmean**2))
val_target_std = tf.sqrt(tf.reduce_sum(val_target_nmean**2))
return -tf.reduce_sum(val_nmean * val_target_nmean) / (val_std * val_target_std)
def init_model(N_FILTERS, FILTER_SZS, STRIDES, N_FC1, EPS, MOMENTUM, \
LSQ_LAMBDA, LSQ_REG_LAMBDA, POL_CROSS_ENTROP_LAMBDA, VAL_LAMBDA, VALR_LAMBDA, L2_LAMBDA, WEIGHT_STD=1e-2):
global convs, weights, outputs, output_nms, pol, pol_pre, pol_mean_sq_err, train_step
global val, val_mean_sq_err, pol_loss, entrop, saver, update_ops
global val_pearsonr, pol_mean_sq_reg_err, loss, Q_map, P_map, visit_count_map
global move_random_ai, init_state, nn_move_unit, nn_prob_move_unit
global tree_to_coords, nn_max_to_coords, nn_prob_to_coords
global tree_prob_move_unit, backup_visit, backup_visit_terminal, tree_det_move_unit
global nn_prob_move_unit, nn_max_move_unit, tree_prob_visit_coord, tree_det_visit_coord
global sess, imgs, valid_mv_map, pol_target, val_target, moving_player
global gm_vars, set_gm_vars, oFC1, session_restore, session_backup
global winner, dir_pre, dir_a
global games_running, score, n_captures, pol_cross_entrop_err
global nn_prob_to_coords_valid_mvs, nn_max_prob_to_coords_valid_mvs
global nn_prob_move_unit_valid_mvs, nn_max_prob_move_unit_valid_mvs
assert len(N_FILTERS) == len(FILTER_SZS) == len(STRIDES)
with tf.device(DEVICE):
#### init state
init_state = tf_op.init_state()
dir_pre = tf.placeholder(tf.float32, shape=())
dir_a = tf.placeholder(tf.float32, shape=())
moving_player = tf.placeholder(tf.int32, shape=())
winner, score, n_captures = tf_op.return_winner(moving_player)
games_running = tf.ones(gv.BATCH_SZ, dtype=tf.int8)
session_restore = tf_op.session_restore()
session_backup = tf_op.session_backup()
##### vars
for var, placeholder in zip(gm_var_nms, gm_var_placeholders):
exec('gm_vars["%s"] = tf_op.%s()' % (var, var))
exec('set_gm_vars["%s"] = tf_op.set_%s(%s)' % (var, var, placeholder))
#### imgs
imgs, valid_mv_map = tf_op.create_batch(moving_player)
#print imgs.shape, imgs_shape
assert imgs.shape == tf.placeholder(tf.float32, shape=imgs_shape).shape, 'tf op shape not matching global_vars'
move_random_ai = tf_op.move_random_ai(moving_player)
global move_frm_inputs, to_coords_input
to_coords_input = tf.placeholder(tf.int32, shape=gv.BATCH_SZ)
move_frm_inputs = tf_op.move_unit(to_coords_input, moving_player)
####
pol_target = tf.placeholder(tf.float32, shape=[gv.BATCH_SZ, map_prod])
val_target = tf.placeholder(tf.float32, shape=[gv.BATCH_SZ])
convs = []; weights = []; outputs = []; output_nms = []
convs += [tf.nn.relu(tf.contrib.layers.batch_norm(tf.layers.conv2d(inputs=imgs, filters=N_FILTERS[0], kernel_size=[FILTER_SZS[0]]*2,
strides=[STRIDES[0]]*2, padding="same", activation=None, name='conv0',
kernel_regularizer=tf.contrib.layers.l2_regularizer(scale=L2_LAMBDA),
bias_regularizer=tf.contrib.layers.l2_regularizer(scale=L2_LAMBDA))))]
weights += [tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'conv0')[0]]
outputs += [convs[-1]]
output_nms += ['conv0']
for i in range(1, len(N_FILTERS)):
output_nms += ['conv' + str(i)]
conv_out = tf.contrib.layers.batch_norm(\
tf.layers.conv2d(inputs=convs[i-1], filters=N_FILTERS[i], kernel_size=[FILTER_SZS[i]]*2,
strides=[STRIDES[i]]*2, padding="same", activation=None, name=output_nms[-1],
kernel_regularizer=tf.contrib.layers.l2_regularizer(scale=L2_LAMBDA),
bias_regularizer=tf.contrib.layers.l2_regularizer(scale=L2_LAMBDA)))
# residual bypass
if (i % 2) == 0:
conv_out += convs[i-2]
convs += [tf.nn.relu(conv_out)]
weights += [tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, output_nms[-1])[0]]
outputs += [convs[-1]]
out_sz = np.int(np.prod(convs[-1].shape[1:]))
convr = tf.reshape(convs[-1], [gv.BATCH_SZ, out_sz])
################### pol
# FC layer
wFC1p = tf.Variable(tf.random_normal([out_sz, N_FC1], stddev=WEIGHT_STD), name='wFC1')
bFC1p = tf.Variable(tf.random_normal([N_FC1], mean=WEIGHT_STD*2, stddev=WEIGHT_STD), name='bFC1')
oFC1p = tf.nn.relu(tf.matmul(convr, wFC1p) + bFC1p)
weights += [wFC1p]
outputs += [oFC1p]
output_nms += ['oFC1p']
# FC layer
wFC2p = tf.Variable(tf.random_normal([N_FC1, map_prod], stddev=WEIGHT_STD), name='wFC2')
bFC2p = tf.Variable(tf.random_normal([map_prod], mean=WEIGHT_STD*2, stddev=WEIGHT_STD), name='bFC2')
pol_pre = tf.nn.relu(tf.matmul(oFC1p, wFC2p) + bFC2p)
weights += [wFC2p]
outputs += [pol_pre]
output_nms += ['pol_pre']
pol = tf.nn.softmax(pol_pre)
outputs += [pol]
output_nms += ['pol']
nn_max_to_coords = tf.argmax(pol_pre, 1, output_type=tf.int32)
nn_prob_to_coords = tf_op.prob_to_coord(pol, dir_pre, dir_a)
nn_prob_to_coords_valid_mvs = tf_op.prob_to_coord_valid_mvs(pol)
nn_max_prob_to_coords_valid_mvs = tf_op.max_prob_to_coord_valid_mvs(pol)
# move unit
nn_max_move_unit = tf_op.move_unit(nn_max_to_coords, moving_player)
nn_prob_move_unit = tf_op.move_unit(nn_prob_to_coords, moving_player)
nn_prob_move_unit_valid_mvs = tf_op.move_unit(nn_prob_to_coords_valid_mvs, moving_player)
nn_max_prob_move_unit_valid_mvs = tf_op.move_unit(nn_max_prob_to_coords_valid_mvs, moving_player)
# sq
sq_err = tf.reduce_sum((pol - pol_target)**2, axis=1)
pol_mean_sq_err = tf.reduce_mean(sq_err)
# sq reg
sq_err_reg = tf.reduce_sum(pol_pre**2, axis=1)
pol_mean_sq_reg_err = tf.reduce_mean(sq_err_reg)
# cross entrop
pol_ln = tf.log(pol)
pol_cross_entrop_err = -tf.reduce_mean(pol_target*pol_ln)
################# val
# FC layer
wFC1v = tf.Variable(tf.random_normal([out_sz, N_FC1], stddev=WEIGHT_STD), name='val_wFC1')
bFC1v = tf.Variable(tf.random_normal([N_FC1], mean=WEIGHT_STD*2, stddev=WEIGHT_STD), name='val_bFC1')
#oFC1v = tf.nn.relu(tf.matmul(convr, wFC1v) + bFC1v)
oFC1v = tf.matmul(convr, wFC1v) + bFC1v
weights += [wFC1v]
outputs += [oFC1v]
output_nms += ['oFC1v']
# FC layer
wFC2v = tf.Variable(tf.random_normal([N_FC1, 1], stddev=WEIGHT_STD), name='val')
bFC2v = tf.Variable(tf.random_normal([1], mean=WEIGHT_STD*2, stddev=WEIGHT_STD), name='val')
val = tf.tanh(tf.squeeze(tf.matmul(oFC1v, wFC2v) + bFC2v))
weights += [wFC2v]
outputs += [val]
output_nms += ['val']
# sq error
val_mean_sq_err = tf.reduce_mean((val - val_target)**2)
# pearson
val_pearsonr = tf_pearsonr(val, val_target)
########## FC l2 reg
FC_L2_reg = 0
for weights in [wFC1v, wFC2v, bFC1v, bFC2v, wFC1p, wFC2p, bFC1p, bFC2p]:
FC_L2_reg += tf.reduce_sum(weights**2)
FC_L2_reg *= (L2_LAMBDA/2.)
################### movement from tree statistics
visit_count_map = tf.placeholder(tf.float32, shape=(gv.BATCH_SZ, gv.map_szt))
tree_prob_visit_coord = tf_op.prob_to_coord(visit_count_map, dir_pre, dir_a)
tree_det_visit_coord = tf.argmax(visit_count_map, 1, output_type=tf.int32)
tree_det_move_unit = tf_op.move_unit(tree_det_visit_coord, moving_player)
tree_prob_move_unit = tf_op.move_unit(tree_prob_visit_coord, moving_player)
################### initialize
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
loss = LSQ_LAMBDA * pol_mean_sq_err + \
LSQ_REG_LAMBDA * pol_mean_sq_reg_err + \
POL_CROSS_ENTROP_LAMBDA * pol_cross_entrop_err + \
VAL_LAMBDA * val_mean_sq_err + \
VALR_LAMBDA * val_pearsonr + \
tf.losses.get_regularization_loss() + FC_L2_reg
with tf.control_dependencies(update_ops):
train_step = tf.train.MomentumOptimizer(EPS, MOMENTUM).minimize(loss)
sess.run(tf.global_variables_initializer())
# saving
saver = tf.train.Saver()