class Actor:
def __init__(self, game, layers=[], checkpoint=None, format='one_hot', optimizer='adam'):
self.game = game
self.format = format
self.layers = layers
self.optimizer = optimizer
self.network = Network(
[game.state_size(format)] + layers + [game.num_possible_moves()],
[],
minibatch_size=50,
steps=1,
loss_function='cross_entropy',
validation_fraction=0,
test_fraction=0,
learning_rate=0.001,
optimizer=optimizer,
output_functions=[tf.nn.softmax]
)
self.network.build()
if checkpoint:
self.load_checkpoint(checkpoint)
def select_move(self, state, stochastic=False):
possible_moves = self.game.get_moves(state)
formatted_state = self.game.format_for_nn(state, format=self.format)
predictions = self.network.predict([formatted_state])[0]
predictions = predictions[:len(possible_moves)]
if not stochastic:
move = np.argmax(predictions)
return possible_moves[move]
predictions = np.array(predictions)
ps = predictions.sum()
if predictions.sum() == 0:
move = np.random.choice(np.arange(0, len(predictions)))
else:
predictions = predictions / predictions.sum()
move = np.random.choice(np.arange(0, len(predictions)), p=predictions)
return possible_moves[move]
def save_checkpoint(self, checkpoint):
self.network.save(checkpoint)
def load_checkpoint(self, checkpoint):
self.network.load(checkpoint)
y_train = tf.keras.utils.to_categorical(y_train)
x_train = (x_train / 255).astype('float32')
x_test = (x_test / 255).astype('float32')
net = Network()
net.init(input_dimension=784,
loss_function="cross entropy",
layers=[{
"units": 128,
"activation": "relu",
"type": "dense"
}, {
"units": 64,
"activation": "relu",
"type": "dense"
}, {
"units": 10,
"activation": "softmax",
"type": "dense"
}])
net.fit(x_train, y_train, epochs=10)
y_pred = net.predict(x_test)
y_pred = np.argmax(y_pred, axis=1)
cmatrix = confusion_matrix(y_test, y_pred)
print(cmatrix)
print(f"Accuracy score: {metrics.accuracy_score(y_test, y_pred):10.5}")