def test_create(self):
self.assertEqual(2, len(
self.net(tf.convert_to_tensor(
encoders.prepareForNetwork([self.env.board], [self.env.isPlaying], [self.env.moveNeeded],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0]],
[self.env.selected]).reshape((-1, 24, 4))))))
print(self.net(tf.convert_to_tensor(
encoders.prepareForNetwork([self.env.board, self.env.board], [self.env.isPlaying, self.env.isPlaying],
[2, 1],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0], 0],
[self.env.selected, None]).reshape((-1, 24, 4)))))
def train_net(in_path, out_path, train_data, tensorboard_path):
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
physical_devices = tf.config.list_physical_devices('GPU')
for gpu_instance in physical_devices:
tf.config.experimental.set_memory_growth(gpu_instance, True)
tensorboard_callback = keras.callbacks.TensorBoard(tensorboard_path,
update_freq=10)
current_Network = Network.get_net(configs.FILTERS, configs.HIDDEN_SIZE,
configs.OUT_FILTERS, configs.NUM_ACTIONS,
configs.INPUT_SIZE, None,
configs.NUM_RESIDUAL)
current_Network.load_weights(in_path)
current_Network.compile(optimizer='adam',
loss={
'policy_output':
Network.cross_entropy_with_logits,
'value_output': 'mse'
},
loss_weights=[0.5, 0.5],
metrics=['accuracy'])
current_Network.fit(encoders.prepareForNetwork(train_data[0],
train_data[1],
train_data[2],
train_data[3],
train_data[4]),
{
'policy_output': train_data[5],
'value_output': train_data[6]
},
epochs=configs.EPOCHS,
batch_size=configs.BATCH_SIZE,
callbacks=[tensorboard_callback])
current_Network.save_weights(out_path)
def test_fit(self):
import Network
self.net.compile(optimizer='adam',
loss={'policy_output': Network.cross_entropy_with_logits, 'value_output': 'mse'},
loss_weights=[0.5, 0.5],
metrics=['accuracy'])
self.net.fit(
encoders.prepareForNetwork([self.env.board, self.env.board], [self.env.isPlaying, self.env.isPlaying],
[2, 1],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0], 0],
[self.env.selected, None]).reshape((-1, 24, 4))
,
{'policy_output': np.array(
[[.25, 0., 0.25, 0., 0., 0., 0., 0., 0., 0., 0.5, 0., 0., 0., 0., 0., 0., 0., 0., 0., .0, 0.,
-1., 0.], [.25, 0., 0.25, 0., 0., 0., 0., 0., .1, 0., 0.4, 0., 0., 0., 0., 0., 0., 0., 0., 0., .0, 0.,
-1., 0.]]),
'value_output': np.array([[0.5], [-0.25]])}, epochs=4,
batch_size=2)
def test_fit_generated(self):
from tensorflow import keras
memory = np.zeros((48000, 7), dtype=object)
val = mp.Value("L")
self.mcts.generatePlay(memory, self.nnet, val, 1, 1)
policy_out, board_in = np.zeros((val.value, 24), dtype=np.float32), np.zeros((val.value, 8, 3),
dtype=np.float32)
for idx in range(val.value):
policy_out[idx], board_in[idx] = memory[idx, 5], memory[idx, 0]
tensorboard_callback = keras.callbacks.TensorBoard("TensorBoard", update_freq=2, profile_batch=0)
self.nnet.fit(
encoders.prepareForNetwork(board_in, memory[:val.value, 1], memory[:val.value, 4],
memory[:val.value, 3],
memory[:val.value, 2]),
{'policy_output': policy_out,
'value_output': memory[:val.value, 6].astype(np.float32)}, epochs=5,
batch_size=128, callbacks=[tensorboard_callback])
self.nnet.save("models/test_save")
def setValAndPriors(self, nnet):
if self.is_terminal_node() == 0:
self.priors, val = nnet(
convert_to_tensor(
encoders.prepareForNetwork(
[self.state[0]], [self.state[1]], [self.state[3]],
[self.state[2][1 if self.state[1] == 1 else 0]],
[self.state[7]])))
mask = np.ones(self.priors.shape, dtype=bool)
mask[0, self.valid_moves] = False
self.priors = np.array(self.priors)
self.priors[mask] = -100.
self.priors = softmax(convert_to_tensor(self.priors)).numpy()
else:
if self.is_terminal_node() == 2:
val = 0
else:
val = self.is_terminal_node()
self.priors = np.zeros((1, 24))
return float(val)
def test_selected_reschape(self):
while self.env.moveNeeded != 2:
self.env.makeMove(self.random_gen.choice(self.env.getValidMoves()))
self.assertEqual(encoders.prepareForNetwork([self.env.board], [self.env.isPlaying], [self.env.moveNeeded],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0]],
[self.env.selected])[0, 4, 0, 0], 2)
def test_reshape(self):
self.env.makeMove(0)
self.assertEqual(1, encoders.prepareForNetwork([self.env.board], [self.env.isPlaying], [self.env.moveNeeded],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0]],
[self.env.selected])[0, 0, 0, 1])
def test_shape(self):
self.assertEqual((1, 8, 3, 4),
encoders.prepareForNetwork([self.env.board], [self.env.isPlaying], [self.env.moveNeeded],
[self.env.gamePhase[1 if self.env.isPlaying == 1 else 0]],
[self.env.selected]).shape)