def residual_layer(self, initial_input):
# Residual layer as defined in literature
output_1 = Conv2D(filters=Param.FILTER_PER_LAYER,
kernel_size=Param.FILTER_KERNEL_SIZE,
data_format="channels_first",
padding='same',
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(
Param.L2_REG_CONST))(initial_input)
output_2 = BN(axis=1)(output_1)
output_3 = LR()(output_2)
output_4 = Conv2D(filters=Param.FILTER_PER_LAYER,
kernel_size=Param.FILTER_KERNEL_SIZE,
data_format="channels_first",
padding='same',
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(
Param.L2_REG_CONST))(output_3)
output_5 = BN(axis=1)(output_4)
output_6 = add([initial_input, output_5])
final_output = LR()(output_6)
return final_output
def value_head(self, initial_input):
# Value head as defined in literature
output_1 = Conv2D(
filters=1,
kernel_size=(1, 1), # 1x1 convolution
data_format="channels_first",
padding='same',
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(
Param.L2_REG_CONST))(initial_input)
output_2 = BN(axis=1)(output_1)
output_3 = LR()(output_2)
output_4 = Flatten()(output_3)
output_5 = Dense(
Param.
FILTER_PER_LAYER, # Use same convention of having the same number of nodes here as filter layers
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(Param.L2_REG_CONST))(output_4)
output_6 = LR()(output_5)
final_output = Dense(1,
use_bias=False,
activation='tanh',
kernel_regularizer=regularizers.l2(
Param.L2_REG_CONST),
name='value_head')(output_6)
return final_output
def residual_layer(self, initial_input):
output_1 = Dense(hive_command.RES_LAYER_NODE_COUNT,
use_bias=False,
activation='sigmoid',
kernel_regularizer=regularizers.l2(
hive_command.L2_REG_CONST))(initial_input)
output_2 = BN(axis=1)(output_1)
output_3 = LR()(output_2)
output_4 = Dense(hive_command.RES_LAYER_NODE_COUNT,
use_bias=False,
activation='sigmoid',
kernel_regularizer=regularizers.l2(
hive_command.L2_REG_CONST))(output_3)
output_5 = BN(axis=1)(output_4)
output_6 = add([initial_input, output_5])
final_output = LR()(output_6)
return final_output
def policy_head(self, initial_input):
# Policy head as defined in literature
output_1 = Conv2D(
filters=2,
kernel_size=(1, 1), # 1x1 convolution
data_format="channels_first",
padding='same',
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(Param.L2_REG_CONST),
kernel_initializer='random_uniform')(initial_input)
output_2 = BN(axis=1)(output_1)
output_3 = LR()(output_2)
output_4 = Flatten()(output_3)
final_output = Dense(
361, # 19 * 19,
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(Param.L2_REG_CONST),
name='policy_head')(output_4)
return final_output
def fertilize(self):
initial_input = Input(hive_command.INPUT_DIM)
output_1 = Dense(
128,
use_bias=False,
activation='sigmoid',
kernel_regularizer=regularizers.l2(hive_command.L2_REG_CONST),
)(initial_input)
output_2 = Dense(
hive_command.RES_LAYER_NODE_COUNT,
use_bias=False,
activation='sigmoid',
kernel_regularizer=regularizers.l2(hive_command.L2_REG_CONST),
)(output_1)
output_3 = BN(axis=1)(output_2)
res_input = LR()(output_3)
res_output = res_input
for i in range(0, hive_command.RES_LAYER_COUNT):
res_output = self.residual_layer(res_input)
res_input = res_output
policy = Dense(11,
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(
hive_command.L2_REG_CONST),
name='policy')(res_output)
model = Model(inputs=[initial_input], outputs=[policy])
model.compile(
loss='mean_squared_error',
optimizer=Adam(lr=hive_command.INITIAL_LEARNING_RATE,
beta_1=0.9,
beta_2=0.999,
epsilon=10E-8),
)
return model
def build_model(self):
# This function constructs the model
initial_input = Input(Param.INPUT_DIM)
output_1 = Conv2D(filters=Param.FILTER_PER_LAYER,
kernel_size=Param.FILTER_KERNEL_SIZE,
data_format="channels_first",
padding='same',
use_bias=False,
activation='linear',
kernel_regularizer=regularizers.l2(
Param.L2_REG_CONST))(initial_input)
output_2 = BN(axis=1)(output_1)
res_input = LR()(output_2)
res_output = res_input # Initialize res_output
# Add residual layers
for i in range(0, Param.RES_LAYER_COUNT):
res_output = self.residual_layer(res_input)
res_input = res_output
value = self.value_head(res_output)
policy = self.policy_head(res_output)
model = Model(inputs=[initial_input], outputs=[value, policy])
# The fit method uses the Adam optimizer because SGD tends to cause the loss function to become unstable
# (nans during training), and overall Adam performs better than SGD.
model.compile(loss={
'value_head': 'mean_squared_error',
'policy_head': loss_function
},
optimizer=Adam(lr=Param.INITIAL_LEARNING_RATE,
beta_1=0.9,
beta_2=0.999,
epsilon=10E-8),
loss_weights={
'value_head': 0.5,
'policy_head': 0.5
})
return model