def GCN_net():
I1 = Input(shape=(no_agents, no_features), name="gcn_input")
Adj = Input(shape=(no_agents, no_agents), name="adj")
gcn = GCNConv(arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1),
use_bias=False,
name="Gcn")([I1, Adj])
concat = tf.keras.layers.Concatenate(axis=2)([I1, gcn])
dense = Dense(arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1),
name="dense_layer")
last_dense = Dense(no_actions,
kernel_initializer=tf.keras.initializers.he_uniform(),
name="last_dense_layer")
split = Lambda(lambda x: tf.squeeze(
tf.split(x, num_or_size_splits=no_agents, axis=1), axis=2))(concat)
outputs = []
for j in list(range(no_agents)):
output = last_dense(dense(split[j]))
output = tf.keras.activations.tanh(output)
outputs.append(output)
V = tf.stack(outputs, axis=1)
model = Model([I1, Adj], V)
model._name = "final_network"
return model
def graph_net(arglist):
I1 = Input(shape=(no_agents, no_features), name="graph_input")
Adj = Input(shape=(no_agents, no_agents), name="adj")
gat = GATConv(
arglist.no_neurons,
activation='relu',
attn_heads=4,
concat_heads=True,
)([I1, Adj])
concat = tf.keras.layers.Concatenate(axis=2)([I1, gat])
dense = Dense(arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1),
name="dense_layer")
last_dense = Dense(no_actions,
kernel_initializer=tf.keras.initializers.he_uniform(),
name="last_dense_layer")
split = Lambda(lambda x: tf.squeeze(
tf.split(x, num_or_size_splits=no_agents, axis=1), axis=2))(concat)
outputs = []
for j in list(range(no_agents)):
output = last_dense(dense(split[j]))
output = tf.keras.activations.tanh(output)
outputs.append(output)
V = tf.stack(outputs, axis=1)
model = Model([I1, Adj], V)
model._name = "final_network"
tf.keras.utils.plot_model(model, show_shapes=True)
return model
def graph_net(arglist):
I1 = Input(shape=(no_agents, no_features), name="gcn_input")
Adj = Input(shape=(no_agents, no_agents), name="adj")
gat = GATConv(
arglist.no_neurons,
activation='relu',
attn_heads=4,
concat_heads=True,
)([I1, Adj])
concat = tf.keras.layers.Concatenate(axis=2)([I1, gat])
outputs = []
dense = Dense(arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1),
name="dense_layer")
dense2 = Dense(arglist.no_neurons / 2,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1),
name="sec_dense_layer")
state_value = Dense(1,
kernel_initializer='he_uniform',
name="value_output")
state_value_lambda = Lambda(lambda s: K.expand_dims(s[:, 0], -1),
output_shape=(no_actions, ))
action_advantage = Dense(no_actions,
name="advantage_output",
kernel_initializer='he_uniform')
action_advantage_lambda = Lambda(
lambda a: a[:, :] - K.mean(a[:, :], keepdims=True),
output_shape=(no_actions, ))
split = Lambda(lambda x: tf.squeeze(
tf.split(x, num_or_size_splits=no_agents, axis=1), axis=2))(concat)
for j in list(range(no_agents)):
V = dense(split[j])
V2 = dense2(V)
if arglist.dueling:
state_value_dense = state_value(V2)
state_value_n = state_value_lambda(state_value_dense)
action_adj_dense = action_advantage(V2)
action_adj_n = action_advantage_lambda(action_adj_dense)
output = Add()([state_value_n, action_adj_n])
output = tf.keras.activations.tanh(output)
outputs.append(output)
else:
outputs.append(V2)
V = tf.stack(outputs, axis=1)
model = Model([I1, Adj], V)
model._name = "final_network"
tf.keras.utils.plot_model(model, show_shapes=True)
return model
def graph_net(arglist):
I = []
for _ in range(no_agents):
I.append(Input(shape=(
arglist.history_size,
no_features,
)))
outputs = []
temporal_state = None
for i in range(no_agents):
if arglist.temporal_mode.lower() == "rnn":
temporal_state = GRU(arglist.no_neurons)(I[i])
elif arglist.temporal_mode.lower() == "attention":
temporal_state = SelfAttention(
activation=tf.keras.layers.LeakyReLU(alpha=0.1))(I[i])
temporal_state = Lambda(lambda x: x[:, -1])(temporal_state)
else:
raise RuntimeError(
"Temporal Information Layer should be rnn or attention but %s found!"
% arglist.temporal_mode)
dense = Dense(
arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1))(temporal_state)
med_dense = Dense(
arglist.no_neurons,
kernel_initializer=tf.keras.initializers.he_uniform(),
activation=tf.keras.layers.LeakyReLU(alpha=0.1))(dense)
last_dense = Dense(
no_actions,
kernel_initializer=tf.keras.initializers.he_uniform())(med_dense)
outputs.append(last_dense)
V = tf.stack(outputs, axis=1)
model = Model(I, V)
model._name = "final_network"
tf.keras.utils.plot_model(model, show_shapes=True)
return model
def _model(self):
from keras.models import Model
from keras.layers import Dense, Input, Conv2D, add, Flatten, BatchNormalization, ReLU
from keras.optimizers import Adam, SGD
from keras.losses import mean_squared_error, binary_crossentropy
from keras.regularizers import l2
board_input = Input(shape=(6, 7, 3))
# Start conv block
conv_1 = Conv2D(self.filters, (4, 4),
padding='same',
kernel_regularizer=l2(self.l2_reg))(board_input)
norm_1 = BatchNormalization()(conv_1)
relu_1 = ReLU()(norm_1)
# Residual convolution blocks
res_1 = ResBlock(relu_1, self.filters, self.l2_reg)
res_2 = ResBlock(res_1, self.filters, self.l2_reg)
res_3 = ResBlock(res_2, self.filters, self.l2_reg)
res_4 = ResBlock(res_3, self.filters, self.l2_reg)
res_5 = ResBlock(res_4, self.filters, self.l2_reg)
res_6 = ResBlock(res_5, self.filters, self.l2_reg)
# Policy head
policy_conv = Conv2D(32, (1, 1),
use_bias=False,
kernel_regularizer=l2(self.l2_reg))(res_6)
policy_norm = BatchNormalization()(policy_conv)
policy_relu = ReLU()(policy_norm)
policy_flat = Flatten()(policy_relu)
# Policy output
policy = Dense(7,
activation='softmax',
name='policy',
use_bias=False,
kernel_regularizer=l2(self.l2_reg))(policy_flat)
# Value head
value_conv = Conv2D(32, (1, 1),
use_bias=False,
kernel_regularizer=l2(self.l2_reg))(res_5)
value_norm = BatchNormalization()(value_conv)
value_relu_1 = ReLU()(value_norm)
value_flat = Flatten()(value_relu_1)
value_dense = Dense(32,
use_bias=False,
kernel_regularizer=l2(self.l2_reg))(value_flat)
value_relu_2 = ReLU()(value_dense)
# Value output
value = Dense(1,
activation='tanh',
name='value',
use_bias=False,
kernel_regularizer=l2(self.l2_reg))(value_relu_2)
# Final model
model = Model(inputs=[board_input], outputs=[policy, value])
# Compile
model.compile(optimizer=Adam(0.001),
loss={
'value': 'mse',
'policy': softmax_cross_entropy_with_logits
})
# model.compile(optimizer=SGD(0.1, 0.9), loss={'value': 'mse', 'policy': softmax_cross_entropy_with_logits})
# Set the model name
model._name = self.name
return model
def _model(self):
from keras.models import Model
from keras.layers import Dense, Input, Conv2D, add, Flatten, BatchNormalization, ReLU
from keras.optimizers import Adam, SGD
from keras.losses import mean_squared_error, binary_crossentropy
from keras.regularizers import l2
from tensorflow.python.util import deprecation
deprecation._PRINT_DEPRECATION_WARNINGS = False
board_input = Input(shape=(2, 6, 7))
# Start conv block
conv_1 = Conv2D(self.config.n_filters,
(self.config.kernel, self.config.kernel),
padding='same',
data_format='channels_first',
kernel_regularizer=l2(self.config.l2_reg))(board_input)
norm_1 = BatchNormalization(axis=1)(conv_1)
relu_1 = ReLU()(norm_1)
res = relu_1
# Residual convolution blocks
for _ in range(self.config.res_layers):
res = ResBlock(res, self.config.n_filters, self.config.kernel,
self.config.l2_reg)
# Policy head
policy_conv = Conv2D(2, (1, 1),
data_format='channels_first',
kernel_regularizer=l2(self.config.l2_reg))(res)
policy_norm = BatchNormalization(axis=1)(policy_conv)
policy_relu = ReLU()(policy_norm)
policy_flat = Flatten()(policy_relu)
# Policy output
policy = Dense(7,
activation='softmax',
name='policy',
kernel_regularizer=l2(self.config.l2_reg))(policy_flat)
# Value head
value_conv = Conv2D(1, (1, 1),
data_format='channels_first',
kernel_regularizer=l2(self.config.l2_reg))(res)
value_norm = BatchNormalization(axis=1)(value_conv)
value_relu_1 = ReLU()(value_norm)
value_flat = Flatten()(value_relu_1)
value_dense = Dense(self.config.value_dense,
kernel_regularizer=l2(
self.config.l2_reg))(value_flat)
value_relu_2 = ReLU()(value_dense)
# Value output
value = Dense(1,
activation='tanh',
name='value',
kernel_regularizer=l2(self.config.l2_reg))(value_relu_2)
# Final model
model = Model(inputs=[board_input], outputs=[policy, value])
# Compile
model.compile(optimizer=SGD(0.001, momentum=0.9),
loss={
'value': objective_function_for_value,
'policy': objective_function_for_policy
},
metrics={'value': [self.mean]})
# Set the model name
model._name = self.config.model
return model