Repository navigation
Expand file tree
/
Copy pathk2net.py
More file actions
404 lines (324 loc) · 13.9 KB
/
Copy pathk2net.py
File metadata and controls
404 lines (324 loc) · 13.9 KB
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
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
""" modeled after dual_net w/ TF2.4 """
import logging
import os.path
from typing import List, Tuple, Any
import numpy as np
import tensorflow as tf
from tensorflow import keras
import coords
import features as features_lib
import go
import myconf
def get_features():
return features_lib.EXP3_FEATURES # REDUX_FEATURES # DLGO_FEATURES
def get_features_planes():
return features_lib.EXP3_FEATURES_PLANES
Conv2D = keras.layers.Conv2D
conv2d_kwargs = dict(padding='same', kernel_initializer='he_normal', data_format='channels_last')
def residual_module(layer_in, n_filters, kernel_size=(3, 3)):
"""
this aligns more w/ AGZ resnet setup
"""
merge_input = layer_in
# check if the number of filters needs to be increase
if layer_in.shape[-1] != n_filters:
x = Conv2D(n_filters, (1, 1), activation='relu', **conv2d_kwargs)(layer_in)
merge_input = keras.layers.BatchNormalization()(x)
# conv1
x = Conv2D(n_filters, kernel_size, activation=None, **conv2d_kwargs)(layer_in)
x = keras.layers.BatchNormalization()(x)
conv1 = keras.layers.Activation('relu')(x)
# conv2
x = Conv2D(n_filters, kernel_size, activation=None, **conv2d_kwargs)(conv1)
conv2 = keras.layers.BatchNormalization()(x)
# add
x = keras.layers.add([conv2, merge_input])
layer_out = keras.layers.Activation('relu')(x)
return layer_out
def build_model(input_shape):
"""
Trainable params: 72k
2nd round: 125k
6-block 9x9: 396k
"""
inputs = keras.Input(shape=input_shape, name='input')
# add "ones" feature plain
x = tf.pad(inputs, [[0, 0], [0, 0], [0, 0], [0, 1]], 'CONSTANT', constant_values=1)
# block 1
x = residual_module(x, 32, (5, 5))
for i in range(5):
x = residual_module(x, 64, (3, 3))
features_common = x
# value head
x = Conv2D(1, (1, 1), **conv2d_kwargs)(features_common)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Flatten()(x)
x = keras.layers.Dense(64, activation='relu', kernel_regularizer=keras.regularizers.l2(0.01))(x)
# predicting win/loss now
output_value = keras.layers.Dense(1, activation='tanh', kernel_regularizer=keras.regularizers.l2(0.01),
name='value')(x)
# policy head
# final conv, to get down to 1 filter (for score)
x = Conv2D(1, (1, 1), **conv2d_kwargs)(features_common)
move_prob = keras.layers.Flatten()(x)
# x = tf.pad(x, [(0, 0), (0, 1)], mode='constant', constant_values=-1e6)
pass_inputs = tf.stack([
tf.reduce_mean(move_prob, axis=1),
tf.reduce_max(move_prob, axis=1),
tf.math.reduce_std(move_prob, axis=1),
tf.squeeze(output_value, axis=1)
], axis=1)
pass_prob = keras.layers.Dense(1, activation=None)(pass_inputs)
x = tf.concat([move_prob, pass_prob], axis=1)
output_policy = keras.layers.Activation('softmax', name='policy')(x)
# somehow a dense layer makes it harder to train
# output_policy = keras.layers.Dense(82, activation='softmax', name='policy')(x)
model = keras.Model(inputs, [output_policy, output_value])
return model
def build_model_v1(input_shape):
"""
Trainable params: 72k
2nd round: 125k
6-block 9x9: 392k
"""
inputs = keras.Input(shape=input_shape)
# add "ones" feature plain
x = tf.pad(inputs, [[0, 0], [0, 0], [0, 0], [0, 1]], 'CONSTANT', constant_values=1)
# block 1
x = residual_module(x, 32, (5, 5))
# +2 more blocks: #params = 72k
x = residual_module(x, 64, (3, 3))
x = residual_module(x, 32, (3, 3))
features_common = x
# value head
x = Conv2D(1, (1, 1), **conv2d_kwargs)(features_common)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Flatten()(x)
x = keras.layers.Dense(64, activation='relu', kernel_regularizer=keras.regularizers.l2(0.01))(x)
# predicting win/loss now
output_value = keras.layers.Dense(1, activation='tanh', kernel_regularizer=keras.regularizers.l2(0.01),
name='value')(x)
# policy head
# final conv, to get down to 1 filter (for score)
x = Conv2D(1, (1, 1), **conv2d_kwargs)(features_common)
move_prob = keras.layers.Flatten()(x)
# x = tf.pad(x, [(0, 0), (0, 1)], mode='constant', constant_values=-1e6)
pass_inputs = tf.stack([
tf.reduce_mean(move_prob, axis=1),
tf.reduce_max(move_prob, axis=1),
tf.math.reduce_std(move_prob, axis=1),
tf.squeeze(output_value, axis=1)
], axis=1)
pass_prob = keras.layers.Dense(1, activation=None)(pass_inputs)
x = tf.concat([move_prob, pass_prob], axis=1)
output_policy = keras.layers.Activation('softmax', name='policy')(x)
# somehow a dense layer makes it harder to train
# output_policy = keras.layers.Dense(82, activation='softmax', name='policy')(x)
model = keras.Model(inputs, [output_policy, output_value])
return model
def build_model_for_eval():
""" load_model will try to resolve all custom objects used during training, a pain to use
Use this, and model.load_weights() instead
"""
# self.model = keras.models.load_model(save_file,
# custom_objects={'custom_BCE_loss': None})
input_shape = (go.N, go.N, get_features_planes())
model = build_model(input_shape)
return model
class DualNetwork:
""" interface that evaluates a board and returns a policy and value.
While it is most likely implemented by DNN, it could also be
- a remote server like Kata (which has MCTS built-in)
- an MCTS-enhanced policy
"""
def run(self, position: go.Position) -> Tuple[Any, float]:
pass
def run_many(self, positions: List[go.Position]) -> Tuple[np.ndarray, np.ndarray]:
pass
def model_id(self):
return 'unknown'
8000
class GCNetwork:
""" goal-conditioned DualNetwork
For now, goal is game-specific
"""
def run(self, position: go.Position, goal) -> Tuple[Any, float]:
pass
def run_many(self, positions: List[go.Position], goal) -> Tuple[np.ndarray, np.ndarray]:
pass
def model_id(self):
return 'unknown'
class TFDualNetwork(GCNetwork):
def __init__(self, save_file):
self.model_id = save_file or 'random-init'
model = build_model_for_eval()
if save_file:
model.load_weights(save_file)
self.model = model
def run(self, position: go.Position, goal):
probs, values = self.run_many([position], goal)
return probs[0], values[0]
# @tf.function(experimental_relax_shapes=True)
@tf.function(input_signature=(tf.TensorSpec(shape=[None, 9, 9, 12], dtype=tf.uint8),))
def tf_run(self, input):
""" https://www.tensorflow.org/guide/function
"""
print('tracing...')
return self.model(input, training=False)
def run_many(self, positions: List[go.Position], goal) -> Tuple[np.ndarray, np.ndarray]:
f = get_features()
processed = [features_lib.extract_features(p, f, goal) for p in positions]
# model.predict() doc suggests to use __call__ for small batch
# probs, values = self.model(tf.convert_to_tensor(processed), training=False)
probs, values = self.tf_run(tf.convert_to_tensor(processed, dtype=tf.uint8))
return probs.numpy(), values.numpy().squeeze(axis=-1)
class A0JaxNet(DualNetwork):
def __init__(self, saved_model_path: str):
self.model_id = saved_model_path
self.model = tf.saved_model.load(saved_model_path)
def run(self, position: go.Position):
f = features_lib.A0JAX_FEATURES
processed = features_lib.extract_features(position, f)
processed *= position.to_play
probs, value = self.model.f(processed)
return probs.numpy(), value.numpy() * position.to_play
def run_many(self, positions: List[go.Position]) -> Tuple[np.ndarray, np.ndarray]:
f = features_lib.A0JAX_FEATURES
processed = [features_lib.extract_features(position, f) * position.to_play for position in positions]
xs = np.stack(processed)
probs, values = self.model.f_batched(xs)
# values are from current player's perspective. convert to values for black
values_black = values.numpy() * np.array([position.to_play for position in positions])
return probs.numpy(), values_black
class MaskedNet(DualNetwork):
def __init__(self, dnn: DualNetwork, policy_mask: np.array):
self.dnn = dnn
self.mask = policy_mask
def run(self, position: go.Position):
probs, values = self.run_many([position])
return probs[0], values[0]
def run_many(self, positions: List[go.Position]) -> Tuple[np.ndarray, np.ndarray]:
probs, values = self.dnn.run_many(positions)
# np will broadcast mask along the batch dimension
probs = np.multiply(probs, self.mask)
return probs, values
class CoreMLNet(GCNetwork):
""" use coreml for prediction """
def __init__(self, save_file):
self.model_id = save_file
self.model = self.load_mlmodel(save_file)
def run(self, position: go.Position, goal):
probs, values = self.run_many([position], goal)
return probs[0], values[0]
def run_many(self, positions: List[go.Position], goal) -> Tuple[np.ndarray, np.ndarray]:
f = get_features()
processed = [features_lib.extract_features(p, f, goal) for p in positions]
nparray = np.stack(processed).astype(np.float16)
results = self.model.predict({'input': nparray})
probs, values = results['Identity'], results['Identity_1']
return probs, values
@staticmethod
def convert_tf2_to_coreml(save_file):
import coremltools as ct
model = build_model_for_eval()
if save_file:
model.load_weights(save_file)
mlmodel = ct.convert(model,
source='tensorflow',
convert_to="mlprogram",
compute_precision=ct.precision.FLOAT16,
compute_units=ct.ComputeUnit.ALL)
return mlmodel
@staticmethod
def load_mlmodel(mlmodel_fname):
import coremltools as ct
return ct.models.MLModel(mlmodel_fname)
class DummyNetwork(GCNetwork):
""" same interface as DualNetwork. Flat policy, Tromp score as value """
def __init__(self):
self.mod
98A7
el_id = 'dummy'
def run(self, position, goal):
probs, values = self.run_many([position], goal)
return probs[0], values[0]
@staticmethod
def zeroout_edges(probs: np.ndarray):
# not allow edge moves when position.n < 10
prototype = probs
for irow in (0, go.N - 1):
prototype[irow * go.N: (irow+1) * go.N] = 0
# disallow pass as well
prototype[0:: go.N] = 0
prototype[go.N - 1: go.N * go.N: go.N] = 0
return probs / np.sum(probs)
def run_many(self, positions: List[go.Position], goal) -> Tuple[np.ndarray, np.ndarray]:
probs = np.ones(myconf.TOTAL_MOVES) / myconf.TOTAL_MOVES
# if positions[0].n < 10:
# probs = self.zeroout_edges(probs)
probs = np.tile(probs, (len(positions), 1))
values = np.array([p.score() for p in positions])
return probs, np.sign(values)
def bootstrap():
N = go.N
input_shape = (N, N, get_features_planes())
model = build_model(input_shape)
fname = '/tmp/k2net.0.h5'
model.save(fname)
def load_net(model_fpath):
""" instantiate the right network, according to filename """
if model_fpath:
logging.info('loading %s', model_fpath)
if model_fpath.endswith('.mlpackage'):
network = CoreMLNet(model_fpath)
elif model_fpath.endswith('.h5'):
network = TFDualNetwork(model_fpath)
else: # saved_model
assert os.path.isfile(f'{model_fpath}/saved_model.pb')
network = A0JaxNet(model_fpath)
else:
logging.info('use DummyNetwork')
network = DummyNetwork()
return network
def test_filter_np():
probs = np.ones(myconf.TOTAL_MOVES) / myconf.TOTAL_MOVES
probs_filtered = DummyNetwork.zeroout_edges(probs)
print(np.reshape(probs_filtered[:go.N * go.N], (go.N, go.N)))
print(probs_filtered[-1-go.N:])
def test_load_model():
fname = f'{myconf.EXP_HOME}/checkpoints/model7_epoch1.h5'
model = build_model_for_eval()
model.load_weights(fname)
model.summary()
def test_mlmodel():
model = CoreMLNet('/tmp/model7_4.mlpackage')
pos0 = go.Position()
pos1 = pos0.play_move(coords.from_gtp('E5'))
probs, values = model.run_many([pos0, pos1])
print(probs.shape, values)
print(probs)
def test_batch_convert_tf2_to_coreml():
for model_spec in (x.split('_') for x in ['1_5', '2_2', '3_3', '4_4', '5_2', '6_2', '7_4', '8_4', '9_4', '10_4', '11_2']):
generation, epoch = model_spec[0], model_spec[1]
fname = f'{myconf.EXP_HOME}/checkpoints/model{generation}_epoch{epoch}.h5'
print(f'checking {fname}')
assert os.path.isfile(fname)
mlmodel = CoreMLNet.convert_tf2_to_coreml(fname)
mlmodel.save(f'{myconf.EXP_HOME}/checkpoints/model{generation}_{epoch}.mlpackage')
def test_convert_tf2_to_coreml():
MODEL_DIR = f'{myconf.EXP_HOME}/checkpoints'
generation = 0
for epoch in range(1):
fname = f'{MODEL_DIR}/model{generation}_{epoch}.h5'
mlmodel = CoreMLNet.convert_tf2_to_coreml(None) #fname)
mlmodel.save(f'{MODEL_DIR}/model{generation}_{epoch}.mlpackage')
def test_a0jax():
saved_model = "/Users/hyu/PycharmProjects/a0-jax/exp-go5C2/tfmodel/model5-25"
a0net = A0JaxNet(saved_model)
# a0net = load_net('/Users/hyu/PycharmProjects/dlgo/5x5/checkpoints/model11_epoch2.h5')
pos0 = go.Position()
probs, value = a0net.run(pos0)
print(probs, value)
assert np.argmax(probs) == 12
pos1 = pos0.play_move(coords.from_gtp('C2'))
probs, values = a0net.run_many([pos0, pos1])
print(probs.shape, values)