def DNN_auto(x_train):
encoding_dim = 128 #128 original
input_img = Input(shape=(673, ))
encoded = Dense(450,
activation='relu')(input_img) # 450 - output (input layer)
#encoded = Dense(250, activation='relu')(encoded) # 200 - output (hidden layer1)
encoded = Dense(250,
activation='relu')(encoded) # 100 - output (hidden layer2)
encoder_output = Dense(encoding_dim)(
encoded) # 128 - output (encoding layer)
print()
# decoder layers
decoded = Dense(250, activation='relu')(encoder_output)
#decoded = Dense(250, activation='relu')(decoded)
decoded = Dense(450, activation='relu')(decoded)
decoded = Dense(673, activation='tanh')(decoded)
autoencoder = Model(input=input_img, output=decoded)
encoder = Model(input=input_img, output=encoder_output)
autoencoder.compile(optimizer='adam', loss='mse')
autoencoder.fit(
x_train, x_train, epochs=20, batch_size=100, shuffle=True
) # second x_train is given instead of train labels in DNN, ie here, i/p=o/p
#batch_size=100 original
encoded_imgs = encoder.predict(x_train)
return encoder_output, encoded_imgs
def test_sparse_input_validation_split():
test_input = sparse.random(6, 3, density=0.25).tocsr()
in1 = Input(shape=(3,), sparse=True)
out1 = Dense(4)(in1)
test_output = np.random.random((6, 4))
model = Model(in1, out1)
model.compile('rmsprop', 'mse')
model.fit(test_input, test_output, epochs=1, batch_size=2, validation_split=0.2)
def test_sparse_input_validation_split():
test_input = sparse.random(6, 3, density=0.25).tocsr()
in1 = Input(shape=(3,), sparse=True)
out1 = Dense(4)(in1)
test_output = np.random.random((6, 4))
model = Model(in1, out1)
model.compile('rmsprop', 'mse')
model.fit(test_input, test_output, epochs=1, batch_size=2, validation_split=0.2)
def test_sparse_placeholder_fit():
test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
in1 = Input(shape=(3,))
in2 = Input(shape=(3,), sparse=True)
out1 = Dropout(0.5, name='dropout')(in1)
out2 = Dense(4, name='dense_1')(in2)
model = Model([in1, in2], [out1, out2])
model.compile('rmsprop', 'mse')
model.fit(test_inputs, test_outputs, epochs=1, batch_size=2, validation_split=0.2)
def test_sparse_placeholder_fit():
test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
in1 = Input(shape=(3,))
in2 = Input(shape=(3,), sparse=True)
out1 = Dropout(0.5, name='dropout')(in1)
out2 = Dense(4, name='dense_1')(in2)
model = Model([in1, in2], [out1, out2])
model.predict(test_inputs, batch_size=2)
model.compile('rmsprop', 'mse')
model.fit(test_inputs, test_outputs, epochs=1, batch_size=2, validation_split=0.5)
model.evaluate(test_inputs, test_outputs, batch_size=2)
class MyModel:
model = None
def __init__(self, input_shape, use_dropout=True, use_vat=True):
self.input_shape = input_shape
self.use_dropout = use_dropout
self.use_vat = use_vat
def build(self):
input_layer = Input(self.input_shape)
output_layer = self.core_data_flow(input_layer)
if self.use_vat:
self.model = VATModel(input_layer, output_layer).setup_vat_loss()
else:
self.model = Model(input_layer, output_layer)
return self
def core_data_flow(self, input_layer):
x = Convolution2D(nb_filters,
kernel_size[0],
kernel_size[1],
border_mode='valid')(input_layer)
x = Activation('relu')(x)
x = Convolution2D(nb_filters, kernel_size[0], kernel_size[1])(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=pool_size)(x)
if self.use_dropout:
x = Dropout(0.25)(x)
x = Flatten()(x)
x = Dense(128, activation="relu")(x)
if self.use_dropout:
x = Dropout(0.5)(x)
x = Dense(nb_classes, activation='softmax')(x)
return x
def training(self, X_train, y_train, X_test, y_test):
self.model.compile(loss=K.categorical_crossentropy,
optimizer='adadelta',
metrics=['accuracy'])
np.random.seed(1337) # for reproducibility
self.model.fit(X_train,
y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=1,
validation_data=(X_test, y_test))
def test_trainable_weights_count_consistency():
"""Tests the trainable weights consistency check of Model.
This verifies that a warning is shown if model.trainable is modified
and the model is summarized/run without a new call to .compile()
Reproduce issue #8121
"""
a = Input(shape=(3, ), name='input_a')
model1 = Model(inputs=a, outputs=Dense(1)(a))
model1.trainable = False
b = Input(shape=(3, ), name='input_b')
y = model1(b)
model2 = Model(inputs=b, outputs=Dense(1)(y))
model2.compile(optimizer='adam', loss='mse')
model1.trainable = True
# Should warn on .summary()
with pytest.warns(UserWarning) as w:
model2.summary()
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And on .fit()
with pytest.warns(UserWarning) as w:
model2.fit(x=np.zeros((5, 3)), y=np.zeros((5, 1)))
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And shouldn't warn if we recompile
model2.compile(optimizer='adam', loss='mse')
with pytest.warns(None) as w:
model2.summary()
assert len(
w
) == 0, "Warning raised even when .compile() is called after modifying .trainable"
def test_model_with_partial_loss():
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dropout': 'mse'}
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
# Same without dropout.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dense_2': 'mse'}
model.compile(optimizer, loss, metrics={'dense_1': 'mae'})
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
def test_model_with_partial_loss():
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dropout': 'mse'}
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
# Same without dropout.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dense_2': 'mse'}
model.compile(optimizer, loss, metrics={'dense_1': 'mae'})
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
def test_trainable_weights_count_consistency():
"""Tests the trainable weights consistency check of Model.
This verifies that a warning is shown if model.trainable is modified
and the model is summarized/run without a new call to .compile()
Reproduce issue #8121
"""
a = Input(shape=(3,), name='input_a')
model1 = Model(inputs=a, outputs=Dense(1)(a))
model1.trainable = False
b = Input(shape=(3,), name='input_b')
y = model1(b)
model2 = Model(inputs=b, outputs=Dense(1)(y))
model2.compile(optimizer='adam', loss='mse')
model1.trainable = True
# Should warn on .summary()
with pytest.warns(UserWarning) as w:
model2.summary()
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And on .fit()
with pytest.warns(UserWarning) as w:
model2.fit(x=np.zeros((5, 3)), y=np.zeros((5, 1)))
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And shouldn't warn if we recompile
model2.compile(optimizer='adam', loss='mse')
with pytest.warns(None) as w:
model2.summary()
assert len(w) == 0, "Warning raised even when .compile() is called after modifying .trainable"
class Network():
def __init__(self, conf):
# Some Hyperparameters
self._board_size = conf['board_size'] # the size of the playing board
self._lr = conf['learning_rate'] # learning rate of SGD (2e-3)
self._momentum = conf['momentum'] # nesterov momentum (1e-1)
self._l2_coef = conf['l2'] # coefficient of L2 penalty (1e-4)
# Define Network
self._build_network()
# File Location
self._net_para_file = conf['net_para_file']
# If we use previous model or not
self._use_previous_model = conf['use_previous_model']
if self._use_previous_model:
net_para = self._model.load_weights(self._net_para_file)
self._model.set_weights(net_para)
def _build_network(self):
# Input_Layer
init_x = Input((3, self._board_size, self._board_size))
x = init_x
# Convolutional Layer
x = Conv2D(filters=32,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
data_format='channels_first',
kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
# Residual Layer
x = self._residual_block(x)
x = self._residual_block(x)
x = self._residual_block(x)
# Policy Head
policy = Conv2D(filters=2,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
data_format='channels_first',
kernel_regularizer=l2(self._l2_coef))(x)
policy = BatchNormalization()(policy)
policy = Activation('relu')(policy)
policy = Flatten()(policy)
policy = Dense(self._board_size * self._board_size,
kernel_regularizer=l2(self._l2_coef))(policy)
self._policy = Activation('softmax')(policy)
# Value Head
value = Conv2D(filters=1,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
data_format="channels_first",
kernel_regularizer=l2(self._l2_coef))(x)
value = BatchNormalization()(value)
value = Activation('relu')(value)
value = Flatten()(value)
value = Dense(32, kernel_regularizer=l2(self._l2_coef))(value)
value = Activation('relu')(value)
value = Dense(1, kernel_regularizer=l2(self._l2_coef))(value)
self._value = Activation('tanh')(value)
# Define Network
self._model = Model(inputs=init_x, outputs=[self._policy, self._value])
# Define the Loss Function
opt = SGD(lr=self._lr, momentum=self._momentum, nesterov=True)
losses_type = ['categorical_crossentropy', 'mean_squared_error']
self._model.compile(optimizer=opt, loss=losses_type)
def _residual_block(self, x):
x_shortcut = x
x = Conv2D(filters=32,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
data_format="channels_first",
kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Conv2D(filters=32,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
data_format="channels_first",
kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = add([x, x_shortcut]) # Skip Connection
x = Activation('relu')(x)
return x
def predict(self, board, color, random_flip=False):
if random_flip:
b_t, method_index = input_transform(board)
tensor_t = board2tensor(b_t, color, reshape_flag=True)
prob_tensor_t, value_tensor = self._model.predict_on_batch(
tensor_t)
policy = output_decode(prob_tensor_t, method_index, board.shape[0])
value = value_tensor[0][0]
return policy, value
else:
tensor = board2tensor(board, color)
policy, value_tensor = self._model.predict_on_batch(tensor)
value = value_tensor[0][0]
return policy, value
def train(self, board_list, color_list, pi_list, z_list):
# Reguliza Data
tensor_list = np.array([
board2tensor(board_list[i], color_list[i], reshape_flag=False)
for i in range(len(board_list))
])
pi_list = np.array(pi_list)
z_list = np.array(z_list)
# Training
self._model.fit(tensor_list, [pi_list, z_list],
epochs=20,
batch_size=len(color_list),
verbose=1)
# Calculate Loss Explicitly
loss = self._model.evaluate(tensor_list, [pi_list, z_list],
batch_size=len(board_list),
verbose=0)
loss = loss[0]
return loss
def get_para(self):
net_para = self._model.get_weights()
return net_para
def save_model(self):
""" save model para to file """
self._model.save_weights(self._net_para_file)
def load_model(self):
self._model.load_weights(self._net_para_file)
class QNetwork:
def __init__(self, config: Config) -> None:
self.config = config
self.digest = None
def build(self) -> None:
mc = self.config.model
in_x = x = Input((4, 5, 5))
x = Conv2D(filters=mc.cnn_filter_num,
kernel_size=mc.cnn_filter_size,
padding="same",
data_format="channels_first",
kernel_regularizer=l2(mc.l2_reg))(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
for _ in range(mc.res_layer_num):
x = self._build_residual_block(x)
res_out = x
# for policy output
x = Conv2D(filters=2,
kernel_size=1,
data_format="channels_first",
kernel_regularizer=l2(mc.l2_reg))(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
# no output for 'pass'
out = Dense(100,
kernel_regularizer=l2(mc.l2_reg),
activation="softmax",
name="out")(x)
# x = Dense(mc.value_fc_size, kernel_regularizer=l2(mc.l2_reg),
# activation="relu")(x)
# value_out = Dense(1, kernel_regularizer=l2(mc.l2_reg),
# activation="tanh", name="value_out")(x)
self.model = Model(in_x, out, name="slipe_model")
self.model.compile(loss='mse', optimizer=Adam(lr=mc.learning_rate))
self.model.summary()
def _build_residual_block(self, x):
mc = self.config.model
in_x = x
x = Conv2D(filters=mc.cnn_filter_num,
kernel_size=mc.cnn_filter_size,
padding="same",
data_format="channels_first",
kernel_regularizer=l2(mc.l2_reg))(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Conv2D(filters=mc.cnn_filter_num,
kernel_size=mc.cnn_filter_size,
padding="same",
data_format="channels_first",
kernel_regularizer=l2(mc.l2_reg))(x)
x = BatchNormalization(axis=1)(x)
x = Add()([in_x, x])
x = Activation("relu")(x)
return x
# 重みの学習
def replay(self, memory: Memory, batch_size: int, gamma: float,
targetQN: 'QNetwork') -> None:
inputs = np.zeros((batch_size, 4, 5, 5))
targets = np.zeros((batch_size, 100))
mini_batch = memory.sample(batch_size)
for i, (state_b, action_b, reward_b,
next_state_b) in enumerate(mini_batch):
inputs[i] = state_b # shape=(4, 5, 5)
target = reward_b # type: int
# if not (next_state_b == 0).all():
# 価値計算（DDQNにも対応できるように、行動決定のQネットワークと価値関数のQネットワークは分離）
retmainQs = self.model.predict(next_state_b)
next_action = np.argmax(retmainQs) # 最大の報酬を返す行動を選択する
target = reward_b + gamma * \
targetQN.model.predict(next_state_b)[0][next_action]
targets[i] = self.model.predict(state_b)[0][0] # Qネットワークの出力
# 教師信号 action_b: int <= 100
targets[i, action_b] = target
# epochsは訓練データの反復回数、verbose=0は表示なしの設定
self.model.fit(inputs, targets, epochs=1, verbose=0)
@staticmethod
def fetch_digest(weight_path: str):
if os.path.exists(weight_path):
m = hashlib.sha256()
with open(weight_path, "rb") as f:
m.update(f.read())
return m.hexdigest()
def load(self, config_path: str, weight_path: str) -> bool:
if os.path.exists(weight_path): # os.path.exists(config_path) and
logger.debug(f"loading model from {config_path}")
with open(config_path, "rt") as f:
self.model = Model.from_config(json.load(f))
self.model.load_weights(weight_path)
self.model.compile(
loss='mse', optimizer=Adam(lr=self.config.model.learning_rate))
self.model.summary()
self.digest = self.fetch_digest(weight_path)
logger.debug(f"loaded model digest = {self.digest}")
return True
else:
logger.debug(
f"model files does not exist at {config_path} and {weight_path}"
)
return False
def save(self, config_path: str, weight_path: str) -> None:
logger.debug(f"save model to {config_path}")
with open(config_path, "wt") as f:
json.dump(self.model.get_config(), f)
self.model.save_weights(weight_path)
self.digest = self.fetch_digest(weight_path)
logger.debug(f"saved model digest {self.digest}")
class AIPlayer(Player):
def __init__(self, buffer_size, sim_count, train=True, model="", tau = 1, compile=False):
self.buffer = ReplayBuffer(buffer_size)
self.temp_state = deque()
self.train = train
self.loss = 0
self.acc = 0
self.batch_count = 0
self.sim_count = sim_count
if model != "":
self.load(model, compile)
else:
self.create_network()
self.tau = tau
@staticmethod
def create_if_nonexistant(config):
models = glob.glob(config.data.model_location + "*.h5")
if len(models) == 0:
ai = AIPlayer(config.buffer_size, config.game.simulation_num_per_move)
ai.save(config.data.model_location+"model_0.h5")
del ai
def set_training(self, train):
self.train = train
@staticmethod
def clear():
K.clear_session()
def load(self, file, compile=False):
try:
del self.network
except Exception:
pass
self.network = load_model(file, custom_objects={"objective_function_for_policy":AIPlayer.objective_function_for_policy,
"objective_function_for_value":AIPlayer.objective_function_for_value}, compile=compile)
def save(self, file):
self.network.save(file)
def create_network(self):
x_in = Input((3, 8, 8))
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x_in)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
for _ in range(10):
x = self._build_residual_block(x)
res_out = x
x = Conv2D(filters=2, kernel_size=1, data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
policy_out = Dense(8*8+1, activation="softmax", name="policy_out")(x)
x = Conv2D(filters=1, kernel_size=1, data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
x = Dense(64, activation="relu")(x)
value_out = Dense(1, activation="tanh", name="value_out")(x)
self.network = Model(x_in, [policy_out, value_out], name="reversi_model")
self.compile()
def _build_residual_block(self, x):
in_x = x
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Add()([in_x, x])
x = Activation("relu")(x)
return x
def compile(self):
losses = [AIPlayer.objective_function_for_policy, AIPlayer.objective_function_for_value]
self.network.compile(optimizer=optimizers.SGD(lr=1e-3, momentum=0.9), loss=losses)
def update_lr(self, lr):
K.set_value(self.network.optimizer.lr, lr)
@staticmethod
def objective_function_for_policy(y_true, y_pred):
# can use categorical_crossentropy??
return K.sum(-y_true * K.log(y_pred + K.epsilon()), axis=-1)
@staticmethod
def objective_function_for_value(y_true, y_pred):
return mean_squared_error(y_true, y_pred)
def update_buffer(self, winner):
if self.train:
while len(self.temp_state) > 0:
t = self.temp_state.pop()
self.buffer.add((t[0], t[1], winner))
def train_batches(self, batch_size, batches=-1, verbose=2):
if batches == -1:
s_buffer = np.array([_[0] for _ in self.buffer.buffer])
p_buffer = np.array([_[1] for _ in self.buffer.buffer])
v_buffer = np.array([_[2] for _ in self.buffer.buffer])
else:
sample_size = batch_size*batches
sample = []
while sample_size > 0:
sample += self.buffer.sample(sample_size)
sample_size -= self.buffer.size()
s_buffer = np.array([_[0] for _ in sample])
p_buffer = np.array([_[1] for _ in sample])
v_buffer = np.array([_[2] for _ in sample])
history = self.network.fit(s_buffer, [p_buffer, v_buffer], batch_size=batch_size, epochs=1, verbose=verbose)
return history
def preprocess_input(self, board, side):
state = np.zeros((3, 8, 8), dtype=np.int)
for i in range(8):
for j in range(8):
if board[i,j] == 1:
state[0,i,j] = 1
elif board[i,j] == -1:
state[1,i,j] = 1
if side == 1:
state[2,i,j] = 1
return state
def evaluate(self, game, side):
current_input = self.preprocess_input(game.board, side)
pred = self.network.predict(current_input[np.newaxis,:])
return pred[1][0]
def pick_move(self, game, side):
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
monte_prob = self.monte_carlo(game, side)
if self.train:
self.temp_state.append((self.preprocess_input(game.board, side), np.divide(monte_prob, np.sum(monte_prob))))
monte_prob = np.float_power(monte_prob, 1/self.tau)
monte_prob = np.divide(monte_prob, np.sum(monte_prob))
r = random()
for i, move in enumerate(possible_moves):
r -= monte_prob[Othello.move_id(move)]
if r <= 0:
return move
return possible_moves[-1]
def monte_carlo(self, game, side):
N = defaultdict(lambda: 0)
W = defaultdict(lambda: 0)
Q = defaultdict(lambda: 0)
P = defaultdict(lambda: 0)
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
policy = np.zeros((65))
policy[64] = 1
return policy
elif len(possible_moves) == 1:
policy = np.zeros((65))
policy[Othello.move_id(possible_moves[0])] = 1
return policy
current_input = self.preprocess_input(game.board, side)
sid = Othello.state_id(game.board)
pred = self.network.predict(current_input[np.newaxis,:])
policy = pred[0][0]
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move))] = policy[Othello.move_id(move)]/total
for i in range(self.sim_count):
#print("Sim #%d"% i)
clone = deepcopy(game)
current_side = side
visited = deque()
while True:
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
best_move = None
best_move_value = -2
sid = Othello.state_id(clone.board)
for move in possible_moves:
mid = Othello.move_id(move)
qu_val = Q[(sid, mid)] + P[(sid, mid)]/(N[(sid, mid)]+1)
if qu_val > best_move_value:
best_move_value = qu_val
best_move = move
#print(best_move)
if N[(sid, Othello.move_id(best_move))] == 0:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner()*side
Q[node] = W[node]/N[node]
break
current_input = self.preprocess_input(clone.board, current_side)
sid = Othello.state_id(clone.board)
pred = self.network.predict(current_input[np.newaxis,:])
policy = pred[0][0]
value = pred[1][0]
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move))] = policy[Othello.move_id(move)]/total
for node in visited:
N[node] += 1
W[node] += value*side
Q[node] = W[node]/N[node]
#print()
break
else:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner()*side
Q[node] = W[node]/N[node]
break
policy = np.zeros((65))
possible_moves = game.possible_moves(side)
sid = Othello.state_id(game.board)
for move in possible_moves:
mid = Othello.move_id(move)
policy[mid] = N[(sid,mid)]
return policy
class PolicyValueNet():
"""策略价值网络"""
#def __init__(self, board_width, board_height, model_file=None):
def __init__(self, policy_infer_size, model_file=None):
#self.board_width = board_width
#self.board_height = board_height
self.policy_infer_size = policy_infer_size
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
self._loss_train_op()
self.load_model_done = True
if model_file and os.path.exists(model_file):
self.load_model_done = False
self.load_model(model_file)
def load_model(self, model_file):
"""重新加载模型(仅用于selfplay时load new model)"""
try:
#net_params = pickle.load(open(model_file, 'rb'), encoding='bytes') #iso-8859-1')
net_params = utils.pickle_load(model_file)
self.model.set_weights(net_params)
self.load_model_done = True
except:
logging.error("load_model fail! {}\t{}".format(
model_file, utils.get_trace()))
self.load_model_done = False
if os.path.exists(
model_file
) and self.load_model_done is False: #鏂囦欢瀛樺湪鍗村鍦ㄥけ璐ユ椂缁堟杩愯
exit(-1)
return self.load_model_done
def create_policy_value_net(self):
"""创建policy-value网络"""
# 输入层
#in_x = network = Input((4, self.board_width, self.board_height))
in_x = network = Input((4, 1, self.policy_infer_size))
# conv layers
network = Conv2D(filters=32,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
# 走子策略 action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
# infer self.board_width * self.board_height action_probs
#self.policy_net = Dense(self.board_width * self.board_height, activation="softmax", kernel_regularizer=l2(self.l2_const))(policy_net)
self.policy_net = Dense(self.policy_infer_size,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# 盘面价值 state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
# infer one current state score
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
# 创建网络模型
self.model = Model(in_x, [self.policy_net, self.value_net])
# 返回走子策略和价值概率
def policy_value(state_input):
state_input_union = np.array(state_input)
#print(state_input_union)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""使用模型预测棋盘所有actionid的价值概率"""
# 棋盘所有可移动action_ids
legal_positions = board.availables
#print(legal_positions)
# 当前玩家角度的actions过程
current_actions = board.current_actions()
#print(current_actions)
# 使用模型预测走子策略和价值概率
#print(self.policy_infer_size)
#act_probs, value = self.policy_value(current_actions.reshape(-1, 4, self.board_width, self.board_height))
act_probs, value = self.policy_value(
current_actions.reshape(-1, 4, 1, self.policy_infer_size))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
# 返回[(action, 概率)] 以及当前玩家的后续走子value
return act_probs, value[0][0]
def _loss_train_op(self):
"""初始化损失
3个损失函数因子
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
loss = value损失函数 + policy损失函数 + 惩罚项
"""
# 定义优化器和损失函数
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
"""输出训练过程中的结果"""
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
# 评估
loss = self.model.evaluate(state_input_union,
[mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
# 预测
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
"""获得模型参数"""
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
"""保存模型参数到文件"""
net_params = self.get_policy_param()
#pickle.dump(net_params, open(model_file, 'wb'), protocol=4)
utils.pickle_dump(net_params, model_file)
class PolicyValueNet():
"""policy-value network """
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
if model_file:
# net_params = pickle.load(open(model_file, 'rb'))
# self.model.set_weights(net_params)
self.model = load_model(model_file)
else:
# self.create_policy_value_net()
self.create_policy_value_resnet()
self._loss_train_op()
def create_policy_value_resnet(self):
def _conv_bn_relu(filters=128, kernel_size=(3, 3)):
def f(input):
conv = Conv2D(kernel_size=kernel_size,
filters=filters,
padding="same",
data_format="channels_first",
kernel_regularizer=l2(self.l2_const))(input)
norm = BatchNormalization(axis=1)(conv)
return Activation("relu")(norm)
return f
def _conv_bn(filters=128, kernel_size=(3, 3)):
def f(input):
conv = Conv2D(kernel_size=kernel_size,
filters=filters,
padding="same",
data_format="channels_first",
kernel_regularizer=l2(self.l2_const))(input)
norm = BatchNormalization(axis=1)(conv)
return norm
return f
def _basic_block(nb_filters):
def f(input):
conv1 = _conv_bn_relu(nb_filters, (3, 3))(input)
conv2 = _conv_bn(nb_filters, (3, 3))(conv1)
shortcut = keras.layers.add([conv1, conv2])
return Activation("relu")(shortcut)
return f
in_x = network = Input((4, self.board_width, self.board_height))
network = _basic_block(64)(network)
network = _basic_block(128)(network)
'''
layer1 = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first",
activation="relu", kernel_regularizer=l2(self.l2_const))(network)
layer2 = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first",
activation="relu", kernel_regularizer=l2(self.l2_const))(layer1)
network = Conv2D(filters=128, kernel_size=(3, 3), padding="same", data_format="channels_first",
activation="relu", kernel_regularizer=l2(self.l2_const))(network)
'''
# action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width * self.board_height,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def create_policy_value_net(self):
"""create the policy value network """
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
'''
network = Conv2D(filters=32, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
'''
layer1 = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
layer2 = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(layer1)
network = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
# action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width * self.board_height,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
'''
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
'''
def policy_value_fn(self, board):
"""
input: board
output: a list of (action, probability) tuples for each available action and the score of the board state
"""
legal_positions = board.availables
current_state = board.current_state()
act_probs, value = self.policy_value(
current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
return act_probs, value[0][0]
def _loss_train_op(self):
"""
Three loss terms：
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
"""
# get the train op
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
loss = self.model.evaluate(state_input_union,
[mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
""" save model params to file """
# net_params = self.get_policy_param()
# pickle.dump(net_params, open(model_file, 'wb'), protocol=2)
# self.model.save_weights(model_file)
self.model.save(model_file)
@staticmethod
def _shortcut(self, input, residual):
stride_width = input._keras_shape[2] / residual._keras_shape[2]
stride_height = input._keras_shape[3] / residual._keras_shape[3]
equal_channels = residual._keras_shape[1] == input._keras_shape[1]
shortcut = input
if stride_width > 1 or stride_height > 1 or not equal_channels:
shortcut = Conv2D(nb_filter=residual._keras_shape[1],
nb_row=1,
nb_col=1,
subsample=(stride_width, stride_height),
init="he_normal",
border_mode="valid")(input)
return merge([shortcut, residual], mode="sum")
@staticmethod
def _residual_block(self,
block_function,
nb_filters,
repetations,
is_first_layer=False):
def f(input):
for i in range(repetations):
init_subsample = (1, 1)
if i == 0 and not is_first_layer:
init_subsample = (2, 2)
input = block_function(nb_filters=nb_filters,
init_subsample=init_subsample)(input)
return input
return f
def resnet(self):
from keras.layers.convolutional import MaxPooling2D, AveragePooling2D
input = Input(shape=(3, 224, 224))
conv1 = self._conv_bn_relu(nb_filter=64,
nb_row=7,
nb_col=7,
subsample=(2, 2))(input)
pool1 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
border_mode="same")(conv1)
# Build residual blocks..
block_fn = self._basic_block
block1 = self._residual_block(block_fn,
nb_filters=64,
repetations=3,
is_first_layer=True)(pool1)
block2 = self._residual_block(block_fn, nb_filters=128,
repetations=4)(block1)
block3 = self._residual_block(block_fn, nb_filters=256,
repetations=6)(block2)
block4 = self._residual_block(block_fn, nb_filters=512,
repetations=3)(block3)
# Classifier block
pool2 = AveragePooling2D(pool_size=(7, 7),
strides=(1, 1),
border_mode="same")(block4)
flatten1 = Flatten()(pool2)
dense = Dense(output_dim=1000, init="he_normal",
activation="softmax")(flatten1)
model = Model(input=input, output=dense)
return model
cnn = Convolution1D(filters=50, kernel_size=3, activation='tanh')(dropouted)
cnn = Convolution1D(filters=50, kernel_size=3, activation='tanh')(cnn)
flattened = Flatten()(cnn)
dense = Dense(100, activation='tanh')(flattened)
predict = Dense(2, activation='softmax')(dense)
model = Model(input=[word, distance_e1, distance_e2], output=predict)
# opt = RMSprop(lr=0.001, rho=0.9, epsilon=1e-06)
# opt = Adagrad(lr=0.01, epsilon=1e-06)
# opt = Adadelta(lr=1.0, rho=0.95, epsilon=1e-06)
# opt = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
opt = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=opt)
train_instances = [line.strip() for line in lines]
label_array_t, word_array_t, dis_e1_array_t, dis_e2_array_t = rep.represent_instances(
train_instances)
model.fit([word_array_t, dis_e1_array_t, dis_e2_array_t],
label_array_t,
batch_size=128,
epochs=epoch_size)
model.save(output_file)
label_array_ans = model.predict([word_array_t, dis_e1_array_t, dis_e2_array_t],
batch_size=128)
print(label_array_ans)
print("训练完成！！")
eval_mulclass(label_array_t, label_array_ans)
def build_CNN_model(inputType, do_training=False, model_inputs=None, loss_func='binary_crossentropy',
optimize_proc='adam', is_IntermediateModel=False, load_weight_path=None, **kwargs):
"""
:param inputType:
:param do_training:
:param model_inputs:
:param loss_func:
:param optimize_proc:
:param is_IntermediateModel:
:param load_weight_path:
:param kwargs:
:return:
"""
# assert not do_training and model_inputs, "if do_training then must pass in model_inputs dictionary"
EMBEDDING_TYPE = 'embeddingMatrix'
ONEHOT_TYPE = '1hotVector'
defined_input_types = {EMBEDDING_TYPE, ONEHOT_TYPE}
assert inputType in defined_input_types, "unknown input type {0}".format(inputType)
if inputType is ONEHOT_TYPE:
review_input = Input(shape=(modelParameters.MaxLen_w,), dtype='float32',
name="ONEHOT_INPUT")
layer = Embedding(modelParameters.VocabSize_w + modelParameters.INDEX_FROM, embedding_dims,
embeddings_initializer=embedding_init, embeddings_regularizer=embedding_reg,
input_length=modelParameters.MaxLen_w, name='1hot_embeddingLayer')(review_input)
layer = SpatialDropout1D(0.50)(layer)
elif inputType is EMBEDDING_TYPE:
review_input = Input(shape=(modelParameters.MaxLen_w, embedding_dims), dtype="float32", name="EMBEDDING_INPUT")
layer = review_input
else:
raise ValueError("Bad inputType arg to build_CNN_model")
layer = Convolution1D(filters=num_filters1,
kernel_size=filter_length1,
padding=region,
strides=1,
activation=conv_activation1,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=conv_reg1,
dilation_rate=1,
name='ConvLayer1')(layer)
layer = SpatialDropout1D(0.50)(layer)
layer = MaxPooling1D(pool_size=pool_len1)(layer)
# layer = Convolution1D(filters=num_filters2,
# kernel_size=filter_length2,
# padding=region,
# strides=1,
# activation=conv_activation2,
# kernel_initializer=conv_init2,
# kernel_regularizer=conv_reg2,
# dilation_rate=1,
# name='ConvLayer2')(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len2)(layer)
# layer = Convolution1D(filters=num_filters3,
# kernel_size=filter_length3,
# padding=region,
# activation=conv_activation3,
# kernel_initializer=conv_init3,
# kernel_regularizer=conv_reg3,
# dilation_rate=1,
# name='ConvLayer3')(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len3)(layer)
# #layer = GlobalMaxPool1D()(layer)
#
# layer = Convolution1D(filters=num_filters4,
# kernel_size=filter_length4,
# padding=region,
# activation=conv_activation4,
# kernel_initializer=conv_init4,
# kernel_regularizer=conv_reg4,
# dilation_rate=1,
# name='ConvLayer4')(layer)
#
# #layer = leaky_relu(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len4)(layer)
# #layer = GlobalMaxPool1D()(layer)
#
# # layer = BatchNormalization()(layer)
layer = Flatten()(layer)
layer = Dense(dense_dims0, activation=dense_activation0, kernel_regularizer=dense_reg0,
kernel_initializer='glorot_normal', bias_initializer='zeros',
name='dense0')(layer)
layer = Dropout(0.50)(layer)
layer = Dense(dense_dims1, activation=dense_activation1, kernel_regularizer=dense_reg1,
kernel_initializer='glorot_normal', bias_initializer='zeros',
name='dense1')(layer)
layer = Dropout(0.50)(layer)
# layer = Dense(dense_dims2, activation=dense_activation2, kernel_regularizer=dense_reg2,
# kernel_initializer=dense_init2,
# name='dense2')(layer)
#
#
# layer = Dropout(0.50)(layer)
#
# layer = Dense(dense_dims3, activation=dense_activation3, kernel_regularizer=dense_reg3,
# kernel_initializer=dense_init3,
# name='dense3_outA')(layer)
# #layer = leaky_relu(layer)
#
if is_IntermediateModel:
return Model(inputs=[review_input], outputs=[layer], name="CNN_model")
#
# layer = Dropout(0.5)(layer)
layer = Dense(dense_dims_final, activation=dense_activation_final, kernel_initializer=dense_init_final,
kernel_regularizer=dense_reg0,
name='output_Full')(layer)
CNN_model = Model(inputs=[review_input], outputs=[layer], name="CNN_model")
CNN_model.compile(optimizer=Adam(lr=0.001, decay=0.0), loss=loss_func, metrics=[binary_accuracy])
if load_weight_path is not None:
CNN_model.load_weights(load_weight_path)
hist = ""
if do_training:
weightPath = os.path.join(modelParameters.WEIGHT_PATH, filename)
configPath = os.path.join(modelParameters.WEIGHT_PATH, filename_config)
with open(configPath + ".json", 'wb') as f:
f.write(CNN_model.to_json())
checkpoint = ModelCheckpoint(weightPath + '_W.{epoch:02d}-{val_loss:.4f}.hdf5',
verbose=1, save_best_only=True, save_weights_only=False, monitor='val_loss')
earlyStop = EarlyStopping(patience=3, verbose=1, monitor='val_loss')
LRadjuster = ReduceLROnPlateau(monitor='val_loss', factor=0.30, patience=0, verbose=1, cooldown=1,
min_lr=0.00001, epsilon=1e-2)
call_backs = [checkpoint, earlyStop, LRadjuster]
CNN_model.summary()
hist = CNN_model.fit(*model_inputs['training'],
batch_size=batch_size,
epochs=nb_epoch, verbose=1,
validation_data=model_inputs['dev'],
callbacks=call_backs)
return {"model": CNN_model, "hist": hist}
def test_pandas_dataframe():
input_a = Input(shape=(3, ), name='input_a')
input_b = Input(shape=(3, ), name='input_b')
x = Dense(4, name='dense_1')(input_a)
y = Dense(3, name='desne_2')(input_b)
model_1 = Model(inputs=input_a, outputs=x)
model_2 = Model(inputs=[input_a, input_b], outputs=[x, y])
optimizer = 'rmsprop'
loss = 'mse'
model_1.compile(optimizer=optimizer, loss=loss)
model_2.compile(optimizer=optimizer, loss=loss)
input_a_df = pd.DataFrame(np.random.random((10, 3)))
input_b_df = pd.DataFrame(np.random.random((10, 3)))
output_a_df = pd.DataFrame(np.random.random((10, 4)))
output_b_df = pd.DataFrame(np.random.random((10, 3)))
model_1.fit(input_a_df, output_a_df)
model_2.fit([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.fit([input_a_df], [output_a_df])
model_1.fit({'input_a': input_a_df}, output_a_df)
model_2.fit({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.predict(input_a_df)
model_2.predict([input_a_df, input_b_df])
model_1.predict([input_a_df])
model_1.predict({'input_a': input_a_df})
model_2.predict({'input_a': input_a_df, 'input_b': input_b_df})
model_1.predict_on_batch(input_a_df)
model_2.predict_on_batch([input_a_df, input_b_df])
model_1.predict_on_batch([input_a_df])
model_1.predict_on_batch({'input_a': input_a_df})
model_2.predict_on_batch({'input_a': input_a_df, 'input_b': input_b_df})
model_1.evaluate(input_a_df, output_a_df)
model_2.evaluate([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.evaluate([input_a_df], [output_a_df])
model_1.evaluate({'input_a': input_a_df}, output_a_df)
model_2.evaluate({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.train_on_batch(input_a_df, output_a_df)
model_2.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.train_on_batch([input_a_df], [output_a_df])
model_1.train_on_batch({'input_a': input_a_df}, output_a_df)
model_2.train_on_batch({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.test_on_batch(input_a_df, output_a_df)
model_2.test_on_batch([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.test_on_batch([input_a_df], [output_a_df])
model_1.test_on_batch({'input_a': input_a_df}, output_a_df)
model_2.test_on_batch({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
class FinancialNewsAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
news_input = Input(shape=(nb_time_step, dim_data))
lstm = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh')
bi_lstm = Bidirectional(lstm, input_shape=(nb_time_step, dim_data), merge_mode='concat')
all_news_rep = bi_lstm(news_input)
news_predictions = Dense(1, activation='linear')(all_news_rep)
self.model = Model(news_input, news_predictions, name="deep rnn for financial news analysis")
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, loss_weights=0.1):
optimizer = Adam(lr=lr)
loss = 'mse'
# loss = custom_objective
self.model.compile(optimizer=optimizer, loss=loss)
#metrics=['mse'])
plot(self.model, to_file='model.png')
def fit_model(self, X, y, X_val=None, y_val=None, epoch=500):
early_stopping = EarlyStopping(monitor='val_loss', patience=100, verbose=0)
if X_val is None:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_split=0.2,
shuffle=True, callbacks=[early_stopping])
else:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_data=(X_val, y_val),
shuffle=True, callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)
count_true = 0
count_all = y.shape[0]
for i in range(y.shape[0]):
count_true = count_true + 1 if y[i,0]*y_hat[i,0]>0 else count_true
print y[i,0],y_hat[i,0]
print count_all,count_true
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np], [output_a_np, output_b_np])
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4, validation_split=0.5,
validation_data=(
{'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
trained_batches = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
def on_batch_begin(batch, logs):
trained_batches.append(batch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin,
on_batch_begin=on_batch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), steps_per_epoch=3, epochs=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# enable verbose for evaluate_generator
out = model.evaluate_generator(gen_data(4), steps=3, verbose=1)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np],))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer, loss='mse', loss_weights={'dense_1': 0.2, 'dropout': 0.8})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': None, 'dropout': 'temporal'})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3), steps_per_epoch=3, epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
validation_steps=3, callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(3)) * 5
# steps_per_epoch will be equal to len of sequence if it's unspecified
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3), epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(12)) * 5
# fit_generator will throw an exception if steps is unspecified for regular generator
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.fit_generator(generator=gen_data(), epochs=5,
initial_epoch=0, validation_data=gen_data(),
callbacks=[tracker_cb])
# Check if generator is only accessed an expected number of times
gen_counters = [0, 0]
def gen_data(i):
while True:
gen_counters[i] += 1
yield ([np.random.random((1, 3)), np.random.random((1, 3))],
[np.random.random((1, 4)), np.random.random((1, 3))])
out = model.fit_generator(generator=gen_data(0), epochs=3,
steps_per_epoch=2,
validation_data=gen_data(1),
validation_steps=1,
max_queue_size=2,
workers=2)
# Need range check here as filling of the queue depends on sleep in the enqueuers
assert 6 <= gen_counters[0] <= 8
# 12 = (epoch * workers * validation steps * max_queue_size)
assert 3 <= gen_counters[1] <= 12
gen_counters = [0]
out = model.fit_generator(generator=RandomSequence(3), epochs=3,
validation_data=gen_data(0),
validation_steps=1,
max_queue_size=2,
workers=2)
# 12 = (epoch * workers * validation steps * max_queue_size)
# Need range check here as filling of the queue depends on sleep in the enqueuers
assert 3 <= gen_counters[0] <= 12
# predict_generator output shape behavior should be consistent
def expected_shape(batch_size, n_batches):
return (batch_size * n_batches, 4), (batch_size * n_batches, 3)
# Multiple outputs and one step.
batch_size = 5
sequence_length = 1
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Multiple outputs and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Create a model with a single output.
single_output_model = Model([a, b], a_2)
single_output_model.compile(optimizer, loss, metrics=[], sample_weight_mode=None)
# Single output and one step.
batch_size = 5
sequence_length = 1
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out) == shape_0
# Single output and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out) == shape_0
class PolicyValueNet():
"""策略价值网络"""
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
self._loss_train_op()
if model_file:
if platform.python_version().split('.')[0] == '3': #python3
net_params = pickle.load(open(model_file, 'rb'), encoding='iso-8859-1')
else:
net_params = pickle.load(open(model_file, 'rb'))
self.model.set_weights(net_params)
def create_policy_value_net(self):
"""创建policy-value网络"""
# 输入层
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
network = Conv2D(filters=32, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
# 走子策略 action policy layers
policy_net = Conv2D(filters=4, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width * self.board_height, activation="softmax", kernel_regularizer=l2(self.l2_const))(policy_net)
# 盘面价值 state value layers
value_net = Conv2D(filters=2, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1, activation="tanh", kernel_regularizer=l2(self.l2_const))(value_net)
# 创建网络模型
self.model = Model(in_x, [self.policy_net, self.value_net])
# 返回走子策略和价值概率
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""使用模型预测棋盘所有可落子位置价值概率"""
# 棋盘所有可落子位置
legal_positions = board.availables
# 当前玩家角度的棋盘方格状态
current_state = board.current_state()
# 使用模型预测走子策略和价值概率
act_probs, value = self.policy_value(current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
# 返回[(action, 概率)] 以及当前玩家的后续走子value
return act_probs, value[0][0]
def _loss_train_op(self):
"""初始化损失
3个损失函数因子
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
loss = value损失函数 + policy损失函数 + 惩罚项
"""
# 定义优化器和损失函数
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
"""输出训练过程中的结果"""
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
# 评估
loss = self.model.evaluate(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
# 预测
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
"""获得模型参数"""
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
"""保存模型参数到文件"""
net_params = self.get_policy_param()
pickle.dump(net_params, open(model_file, 'wb'), protocol=2)
class FinancialTimeSeriesAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
financial_time_series_input = Input(shape=(nb_time_step, dim_data))
lstm_layer_1 = LSTM(output_dim=nb_hidden_units,
dropout_U=dropout,
dropout_W=dropout,
inner_activation='sigmoid',
W_regularizer=l2(l2_norm_alpha),
b_regularizer=l2(l2_norm_alpha),
activation='tanh',
return_sequences=True)
lstm_layer_2 = LSTM(output_dim=nb_hidden_units,
dropout_U=dropout,
dropout_W=dropout,
inner_activation='sigmoid',
W_regularizer=l2(l2_norm_alpha),
b_regularizer=l2(l2_norm_alpha),
activation='tanh',
return_sequences=True)
h1 = lstm_layer_1(financial_time_series_input)
h2 = lstm_layer_2(h1)
time_series_predictions = TimeDistributedDense(1)(h2)
self.model = Model(
financial_time_series_input,
time_series_predictions,
name="deep rnn for financial time series forecasting")
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, arg_weight=1.):
optimizer = Adam(lr=lr)
loss = 'mse'
self.model.compile(optimizer=optimizer, loss=loss)
def fit_model(self, X, y, X_val=None, y_val=None, epoch=3):
early_stopping = EarlyStopping(monitor='val_loss',
patience=3,
verbose=0)
if X_val is None:
self.model.fit(X,
y,
batch_size=self.batch_size,
nb_epoch=epoch,
validation_split=0.2,
shuffle=True,
callbacks=[early_stopping])
else:
self.model.fit(X,
y,
batch_size=self.batch_size,
nb_epoch=epoch,
validation_data=(X_val, y_val),
shuffle=True,
callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" %
(w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)
count_true = 0
count_all = y.shape[1]
for i in range(y.shape[1]):
count_true = count_true + 1 if y[0, i, 0] * y_hat[
0, i, 0] > 0 else count_true
print(y[0, i, 0], y_hat[0, i, 0])
print(count_all, count_true)
class FinancialNewsAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
news_input = Input(shape=(nb_time_step, dim_data), name='x1')
lstm = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha),
activation='tanh', name='h1')
bi_lstm = Bidirectional(lstm, input_shape=(nb_time_step, dim_data), merge_mode='concat', name='h1')
all_news_rep = bi_lstm(news_input)
news_predictions = Dense(1, activation='linear')(all_news_rep)
self.model = Model(news_input, news_predictions, name="deep rnn for financial news analysis")
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, loss_weights=0.1):
optimizer = Adam(lr=lr)
loss = 'mse'
# loss = custom_objective
self.model.compile(optimizer=optimizer, loss=loss)
#metrics=['mse'])
plot(self.model, to_file='model.png')
def fit_model(self, X, y, X_val=None, y_val=None, epoch=500):
early_stopping = EarlyStopping(monitor='val_loss', patience=3, verbose=0)
if X_val is None:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_split=0.2,
shuffle=True, callbacks=[early_stopping])
else:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_data=(X_val, y_val),
shuffle=True, callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)
count_true = 0
count_all = y.shape[0]
for i in range(y.shape[0]):
count_true = count_true + 1 if y[i,0]*y_hat[i,0]>0 else count_true
print y[i,0],y_hat[i,0]
print count_all,count_true
def test_model_methods():
a = Input(shape=(3, ), name='input_a')
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# test fit
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
epochs=1,
batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_data=([input_a_np,
input_b_np], [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
epochs=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10, ))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer,
loss,
metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer,
loss,
metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
trained_batches = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
def on_batch_begin(batch, logs):
trained_batches.append(batch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin,
on_batch_begin=on_batch_begin)
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=5,
batch_size=4,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([
np.random.random((batch_sz, 3)),
np.random.random((batch_sz, 3))
], [
np.random.random((batch_sz, 4)),
np.random.random((batch_sz, 3))
])
out = model.fit_generator(gen_data(4),
steps_per_epoch=3,
epochs=5,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4,
epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch(
[input_a_np, input_b_np], [output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_data=([input_a_np, input_b_np], ))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer,
loss='mse',
sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer,
loss='mse',
sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer,
loss='mse',
loss_weights={
'dense_1': 0.2,
'dropout': 0.8
})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch(
[input_a_np, input_b_np], [output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer,
loss='mse',
sample_weight_mode={
'dense_1': None,
'dropout': 'temporal'
})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3),
steps_per_epoch=3,
epochs=5,
initial_epoch=0,
validation_data=RandomSequence(4),
validation_steps=3,
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(3)) * 5
# steps_per_epoch will be equal to len of sequence if it's unspecified
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3),
epochs=5,
initial_epoch=0,
validation_data=RandomSequence(4),
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(12)) * 5
# fit_generator will throw an exception if steps is unspecified for regular generator
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.fit_generator(generator=gen_data(),
epochs=5,
initial_epoch=0,
validation_data=gen_data(),
callbacks=[tracker_cb])
# predict_generator output shape behavior should be consistent
def expected_shape(batch_size, n_batches):
return (batch_size * n_batches, 4), (batch_size * n_batches, 3)
# Multiple outputs and one step.
batch_size = 5
sequence_length = 1
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Multiple outputs and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Create a model with a single output.
single_output_model = Model([a, b], a_2)
single_output_model.compile(optimizer,
loss,
metrics=[],
sample_weight_mode=None)
# Single output and one step.
batch_size = 5
sequence_length = 1
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out) == shape_0
# Single output and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out) == shape_0
cos_inner = np.pi * (t % (self.T // self.M)
) # t - 1 is used when t has 1-based indexing.
cos_inner /= self.T // self.M
cos_out = np.cos(cos_inner) + 1
return float(self.alpha_zero / 2 * cos_out)
# In[ ]:
from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
check_point = ModelCheckpoint('model.hdf5', verbose=True, save_best_only=True)
early_stop = EarlyStopping(patience=5, verbose=True)
model.fit(X_train,
y_train.astype(int),
validation_data=(X_valid, y_valid.astype(int)),
epochs=5,
verbose=True,
callbacks=[early_stop, check_point])
# In[ ]:
from sklearn.metrics import accuracy_score
# distribution of confidence that will be used as submission
model.load_weights('model.hdf5')
confidence_valid = model.predict(X_valid)[:, 0] * 2 - 1
print(accuracy_score(confidence_valid > 0, y_valid))
plt.hist(confidence_valid, bins='auto')
plt.title("predicted confidence")
plt.show()
# In[ ]:
class Network:
def __init__(self, conf):
# All hyperparameters used in the model
self._board_size = conf['board_size'] # the size of the playing board
self._lr = conf['learning_rate'] # learning rate of SGD (2e-3)
self._momentum = conf['momentum'] # nesterov momentum (1e-1)
self._l2_coef = conf['l2'] # coefficient of L2 penalty (1e-4)
self._mini_batch_size = conf['mini_batch_size'] # the size of batch when training the network
self._fit_epochs = conf['fit_epochs'] # the number of iteration
# Define Network
self._build_network()
# The location of the file which stores the parameters of the network
self._net_para_file = conf['net_para_file']
self._fit_history_file = conf['fit_history_file']
# Whether we use previous model or not
self._use_previous_model = conf['use_previous_model']
if self._use_previous_model:
if os.path.exists(self._net_para_file):
self._model.load_weights(self._net_para_file)
else:
print('> error: [use_previous_model] = True, ' + self._net_para_file + ' not found')
@log
def _build_network(self):
# Input_Layer
init_x = Input((3, self._board_size, self._board_size)) # the input is a tensor with the shape 3*(15*15)
x = init_x
# First Convolutional Layer with 32 filters
x = Conv2D(filters=32, kernel_size=(3, 3), strides=(1, 1), padding='same',
data_format='channels_first', kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
# Two Residual Blocks
x = self._residual_block(x)
x = self._residual_block(x)
x = self._residual_block(x)
# Policy Head for generating prior probability vector for each action
policy = Conv2D(filters=2, kernel_size=(1, 1), strides=(1, 1), padding='same',
data_format='channels_first', kernel_regularizer=l2(self._l2_coef))(x)
policy = BatchNormalization()(policy)
policy = Activation('relu')(policy)
policy = Flatten()(policy)
policy = Dense(self._board_size*self._board_size, kernel_regularizer=l2(self._l2_coef))(policy)
self._policy = Activation('softmax')(policy)
# Value Head for generating value of each action
value = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding='same',
data_format="channels_first", kernel_regularizer=l2(self._l2_coef))(x)
value = BatchNormalization()(value)
value = Activation('relu')(value)
value = Flatten()(value)
value = Dense(32, kernel_regularizer=l2(self._l2_coef))(value)
value = Activation('relu')(value)
value = Dense(1, kernel_regularizer=l2(self._l2_coef))(value)
self._value = Activation('tanh')(value)
# Define Network
self._model = Model(inputs=init_x, outputs=[self._policy, self._value])
# Define the Loss Function
opt = SGD(lr=self._lr, momentum=self._momentum, nesterov=True) # stochastic gradient descend with momentum
losses_type = ['categorical_crossentropy', 'mean_squared_error'] # cross-entrophy and MSE are weighted equally
self._model.compile(optimizer=opt, loss=losses_type)
def _residual_block(self, x):
x_shortcut = x
x = Conv2D(filters=32, kernel_size=(3, 3), strides=(1, 1), padding='same',
data_format="channels_first", kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Conv2D(filters=32, kernel_size=(3, 3), strides=(1, 1), padding='same',
data_format="channels_first", kernel_regularizer=l2(self._l2_coef))(x)
x = BatchNormalization()(x)
x = add([x, x_shortcut]) # Skip Connection
x = Activation('relu')(x)
return x
def predict(self, board, color, last_move):
if sum(sum(board)) == 0 and color == WHITE:
print('error: network.predict')
if sum(sum(board)) == 1 and color == BLACK:
print('error: network.predict')
tensor = board2tensor(board, color, last_move)
policy, value_tensor = self._model.predict_on_batch(tensor)
value = value_tensor[0][0]
return policy, value
def train(self, board_list, color_list, last_move_list, pi_list, z_list):
size = len(color_list)
for i in range(size):
if sum(sum(board_list[i])) == 0 and color_list[i] == WHITE:
print('error: network.train')
if sum(sum(board_list[i])) == 1 and color_list[i] == BLACK:
print('error: network.train')
# Data Augmentation through symmetric and self-rotation transformation
board_aug = []
color_aug = []
last_move_aug = []
pi_aug = []
z_aug = []
for i in range(len(board_list)):
new_board, new_color, new_last_move, new_pi, new_z = \
data_augmentation(board_list[i], color_list[i], last_move_list[i], pi_list[i], z_list[i])
board_aug.extend(new_board)
color_aug.extend(new_color)
last_move_aug.extend(new_last_move)
pi_aug.extend(new_pi)
z_aug.extend(new_z)
board_list.extend(board_aug)
color_list.extend(color_aug)
last_move_list.extend(last_move_aug)
pi_list.extend(pi_aug)
z_list.extend(z_aug)
# Regularize Data
board_list = np.array([board2tensor(board_list[i], color_list[i], last_move_list[i], reshape_flag=False)
for i in range(len(board_list))])
pi_list = np.array(pi_list)
z_list = np.array(z_list)
# Training
hist = self._model.fit(board_list, [pi_list, z_list], epochs=self._fit_epochs, batch_size=self._mini_batch_size, verbose=1)
hist_path = self._fit_history_file + '_' + str(self._fit_epochs) + '_' + str(self._mini_batch_size) + '.txt'
with open(hist_path, 'a') as f:
f.write(str(hist.history))
return hist.history['loss'][0] # only sample loss of first epoch
def get_para(self):
net_para = self._model.get_weights()
return net_para
def save_model(self):
""" save model para to file """
self._model.save_weights(self._net_para_file)
def load_model(self):
if os.path.exists(self._net_para_file):
self._model.load_weights(self._net_para_file)
else:
print('> error: ' + self._net_para_file + ' not found')
def test_model_with_input_feed_tensor():
"""We test building a model with a TF variable as input.
We should be able to call fit, evaluate, predict,
by only passing them data for the placeholder inputs
in the model.
"""
import tensorflow as tf
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=['mean_squared_error'],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch(input_b_np,
[output_a_np, output_b_np])
out = model.train_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.predict_on_batch({'input_b': input_b_np})
# test fit
out = model.fit({'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=10)
out = model.fit(input_b_np,
[output_a_np, output_b_np], epochs=1, batch_size=10)
# test evaluate
out = model.evaluate({'input_b': input_b_np},
[output_a_np, output_b_np], batch_size=10)
out = model.evaluate(input_b_np,
[output_a_np, output_b_np], batch_size=10)
# test predict
out = model.predict({'input_b': input_b_np}, batch_size=10)
out = model.predict(input_b_np, batch_size=10)
assert len(out) == 2
# Now test a model with a single input
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None,
output_a_np)
out = model.train_on_batch(None,
output_a_np)
out = model.test_on_batch(None,
output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([],
output_a_np)
out = model.train_on_batch({},
output_a_np)
# test fit
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None,
output_a_np, batch_size=10)
out = model.evaluate(None,
output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Same, without learning phase
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None,
output_a_np)
out = model.train_on_batch(None,
output_a_np)
out = model.test_on_batch(None,
output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([],
output_a_np)
out = model.train_on_batch({},
output_a_np)
# test fit
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None,
output_a_np, batch_size=10)
out = model.evaluate(None,
output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
def test_model_with_input_feed_tensor():
"""We test building a model with a TF variable as input.
We should be able to call fit, evaluate, predict,
by only passing them data for the placeholder inputs
in the model.
"""
import tensorflow as tf
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer,
loss,
metrics=['mean_squared_error'],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch(input_b_np, [output_a_np, output_b_np])
out = model.train_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.predict_on_batch({'input_b': input_b_np})
# test fit
out = model.fit({'input_b': input_b_np}, [output_a_np, output_b_np],
epochs=1,
batch_size=10)
out = model.fit(input_b_np, [output_a_np, output_b_np],
epochs=1,
batch_size=10)
# test evaluate
out = model.evaluate({'input_b': input_b_np}, [output_a_np, output_b_np],
batch_size=10)
out = model.evaluate(input_b_np, [output_a_np, output_b_np], batch_size=10)
# test predict
out = model.predict({'input_b': input_b_np}, batch_size=10)
out = model.predict(input_b_np, batch_size=10)
assert len(out) == 2
# Now test a model with a single input
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, output_a_np)
out = model.train_on_batch(None, output_a_np)
out = model.test_on_batch(None, output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([], output_a_np)
out = model.train_on_batch({}, output_a_np)
# test fit
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None, output_a_np, batch_size=10)
out = model.evaluate(None, output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Same, without learning phase
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, output_a_np)
out = model.train_on_batch(None, output_a_np)
out = model.test_on_batch(None, output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([], output_a_np)
out = model.train_on_batch({}, output_a_np)
# test fit
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None, output_a_np, batch_size=10)
out = model.evaluate(None, output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
def test_model_with_external_loss():
# None loss, only regularization loss.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1',
kernel_regularizer='l1',
bias_regularizer='l2')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# No dropout, external loss.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a)
model = Model(a, [a_2, a_3])
model.add_loss(K.mean(a_3 + a_2))
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# Test fit with no external data at all.
if K.backend() == 'tensorflow':
import tensorflow as tf
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None, None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None, None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Test multi-output model without external data.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_1 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_1)
model = Model(a, [a_1, a_2])
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None, None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None, None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert len(out) == 2
assert out[0].shape == (10 * 3, 4)
assert out[1].shape == (10 * 3, 4)
def test_model_with_external_loss():
# None loss, only regularization loss.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4,
name='dense_1',
kernel_regularizer='l1',
bias_regularizer='l2')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# No dropout, external loss.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a)
model = Model(a, [a_2, a_3])
model.add_loss(K.mean(a_3 + a_2))
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# Test fit with no external data at all.
if K.backend() == 'tensorflow':
import tensorflow as tf
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Test multi-output model without external data.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_1 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_1)
model = Model(a, [a_1, a_2])
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert len(out) == 2
assert out[0].shape == (10 * 3, 4)
assert out[1].shape == (10 * 3, 4)
class PolicyValueNet():
"""policy-value network """
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
if model_file:
print("[Notice] load model from file")
self.model = load_model(model_file)
else:
print("[Notice] create model")
self.create_policy_value_net()
self._loss_train_op()
def create_policy_value_net(self):
"""create the policy value network """
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
network = Conv2D(filters=32,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
# action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width * self.board_height,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""
input: board
output: a list of (action, probability) tuples for each available action and the score of the board state
"""
legal_positions = board.availables
current_state = board.current_state()
act_probs, value = self.policy_value(
current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
return act_probs, value[0][0]
def _loss_train_op(self):
"""
Three loss terms：
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
"""
# get the train op
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
loss = self.model.evaluate(state_input_union,
[mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
""" save model to file """
print("save model file")
self.model.save(model_file)
def test_pandas_dataframe():
input_a = Input(shape=(3,), name='input_a')
input_b = Input(shape=(3,), name='input_b')
x = Dense(4, name='dense_1')(input_a)
y = Dense(3, name='desne_2')(input_b)
model_1 = Model(inputs=input_a, outputs=x)
model_2 = Model(inputs=[input_a, input_b], outputs=[x, y])
optimizer = 'rmsprop'
loss = 'mse'
model_1.compile(optimizer=optimizer, loss=loss)
model_2.compile(optimizer=optimizer, loss=loss)
input_a_df = pd.DataFrame(np.random.random((10, 3)))
input_b_df = pd.DataFrame(np.random.random((10, 3)))
output_a_df = pd.DataFrame(np.random.random((10, 4)))
output_b_df = pd.DataFrame(np.random.random((10, 3)))
model_1.fit(input_a_df,
output_a_df)
model_2.fit([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.fit([input_a_df],
[output_a_df])
model_1.fit({'input_a': input_a_df},
output_a_df)
model_2.fit({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.predict(input_a_df)
model_2.predict([input_a_df, input_b_df])
model_1.predict([input_a_df])
model_1.predict({'input_a': input_a_df})
model_2.predict({'input_a': input_a_df, 'input_b': input_b_df})
model_1.predict_on_batch(input_a_df)
model_2.predict_on_batch([input_a_df, input_b_df])
model_1.predict_on_batch([input_a_df])
model_1.predict_on_batch({'input_a': input_a_df})
model_2.predict_on_batch({'input_a': input_a_df, 'input_b': input_b_df})
model_1.evaluate(input_a_df,
output_a_df)
model_2.evaluate([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.evaluate([input_a_df],
[output_a_df])
model_1.evaluate({'input_a': input_a_df},
output_a_df)
model_2.evaluate({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.train_on_batch(input_a_df,
output_a_df)
model_2.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.train_on_batch([input_a_df],
[output_a_df])
model_1.train_on_batch({'input_a': input_a_df},
output_a_df)
model_2.train_on_batch({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.test_on_batch(input_a_df,
output_a_df)
model_2.test_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.test_on_batch([input_a_df],
[output_a_df])
model_1.test_on_batch({'input_a': input_a_df},
output_a_df)
model_2.test_on_batch({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
return mean_square_loss + K.mean(loss_tensor, axis=-1)
# return K.mean(y_pred - tf.reshape(loss_tensor, [-1, 3]), axis=-1)
# return mean_square_loss
# return mean_squared_error(y_true, y_pred)
# Compiling the model using 'adam' optimizer and MSE as loss function
# sgd = SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False)
# model.compile(optimizer=sgd, loss=my_loss_function, metrics=['mse', 'mae', 'mape'], loss_weights=[1.0, 1.0, 1.0])
model.compile(optimizer='adam', loss=my_loss_function, metrics=['mse', 'mae', 'mape'], loss_weights=[1.0, 1.0, 1.0])
# model.compile(optimizer='adam', loss='mean_squared_error', metrics=['mse', 'mae', 'mape'], loss_weights=[1.0, 1.0, 1.0])
#muti_outputs shape= tasks x train_samples
callbacks = []
model.fit(x=dat_train, y=[label_train_1, label_train_2, label_train_3], epochs=5000, batch_size=32)
pred1, pred2, pred3 = model.predict(dat_test)
# inv_y = scaler.inverse_transform(inv_y)
plot_x = dat_test[:, 0]
plot_y = pred1.flatten() - label_test_1
# plt.scatter(x=plot_x, y=plot_y)
final_data = np.array([
grid_cells_test,
label_test_1,
pred1.flatten(),
grid_cells_id_test
]).transpose()
class AIPlayer(Player):
def __init__(self,
buffer_size,
sim_count,
train=True,
model="",
tau=1,
compile=False):
self.buffer = ReplayBuffer(buffer_size)
self.temp_state = deque()
self.train = train
self.loss = 0
self.acc = 0
self.batch_count = 0
self.sim_count = sim_count
if model != "":
self.load(model, compile)
else:
self.create_network()
self.tau = tau
@staticmethod
def create_if_nonexistant(config):
models = glob.glob(config.data.model_location + "*.h5")
if len(models) == 0:
ai = AIPlayer(config.buffer_size,
config.game.simulation_num_per_move)
ai.save(config.data.model_location + "model_0.h5")
del ai
def set_training(self, train):
self.train = train
@staticmethod
def clear():
K.clear_session()
def load(self, file, compile=False):
try:
del self.network
except Exception:
pass
self.network = load_model(file,
custom_objects={
"objective_function_for_policy":
AIPlayer.objective_function_for_policy,
"objective_function_for_value":
AIPlayer.objective_function_for_value
},
compile=compile)
def save(self, file):
self.network.save(file)
def create_network(self):
x_in = Input((3, 8, 8))
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x_in)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
for _ in range(10):
x = self._build_residual_block(x)
res_out = x
x = Conv2D(filters=2, kernel_size=1,
data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
policy_out = Dense(8 * 8 + 1, activation="softmax",
name="policy_out")(x)
x = Conv2D(filters=1, kernel_size=1,
data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
x = Dense(64, activation="relu")(x)
value_out = Dense(1, activation="tanh", name="value_out")(x)
self.network = Model(x_in, [policy_out, value_out],
name="reversi_model")
self.compile()
def _build_residual_block(self, x):
in_x = x
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Add()([in_x, x])
x = Activation("relu")(x)
return x
def compile(self):
losses = [
AIPlayer.objective_function_for_policy,
AIPlayer.objective_function_for_value
]
self.network.compile(optimizer=optimizers.SGD(lr=1e-3, momentum=0.9),
loss=losses)
def update_lr(self, lr):
K.set_value(self.network.optimizer.lr, lr)
@staticmethod
def objective_function_for_policy(y_true, y_pred):
# can use categorical_crossentropy??
return K.sum(-y_true * K.log(y_pred + K.epsilon()), axis=-1)
@staticmethod
def objective_function_for_value(y_true, y_pred):
return mean_squared_error(y_true, y_pred)
def update_buffer(self, winner):
if self.train:
while len(self.temp_state) > 0:
t = self.temp_state.pop()
self.buffer.add((t[0], t[1], winner))
def train_batches(self, batch_size, batches=-1, verbose=2):
if batches == -1:
s_buffer = np.array([_[0] for _ in self.buffer.buffer])
p_buffer = np.array([_[1] for _ in self.buffer.buffer])
v_buffer = np.array([_[2] for _ in self.buffer.buffer])
else:
sample_size = batch_size * batches
sample = []
while sample_size > 0:
sample += self.buffer.sample(sample_size)
sample_size -= self.buffer.size()
s_buffer = np.array([_[0] for _ in sample])
p_buffer = np.array([_[1] for _ in sample])
v_buffer = np.array([_[2] for _ in sample])
history = self.network.fit(s_buffer, [p_buffer, v_buffer],
batch_size=batch_size,
epochs=1,
verbose=verbose)
return history
def preprocess_input(self, board, side):
state = np.zeros((3, 8, 8), dtype=np.int)
for i in range(8):
for j in range(8):
if board[i, j] == 1:
state[0, i, j] = 1
elif board[i, j] == -1:
state[1, i, j] = 1
if side == 1:
state[2, i, j] = 1
return state
def evaluate(self, game, side):
current_input = self.preprocess_input(game.board, side)
pred = self.network.predict(current_input[np.newaxis, :])
return pred[1][0]
def pick_move(self, game, side):
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
monte_prob = self.monte_carlo(game, side)
if self.train:
self.temp_state.append((self.preprocess_input(game.board, side),
np.divide(monte_prob, np.sum(monte_prob))))
monte_prob = np.float_power(monte_prob, 1 / self.tau)
monte_prob = np.divide(monte_prob, np.sum(monte_prob))
r = random()
for i, move in enumerate(possible_moves):
r -= monte_prob[Othello.move_id(move)]
if r <= 0:
return move
return possible_moves[-1]
def monte_carlo(self, game, side):
N = defaultdict(lambda: 0)
W = defaultdict(lambda: 0)
Q = defaultdict(lambda: 0)
P = defaultdict(lambda: 0)
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
policy = np.zeros((65))
policy[64] = 1
return policy
elif len(possible_moves) == 1:
policy = np.zeros((65))
policy[Othello.move_id(possible_moves[0])] = 1
return policy
current_input = self.preprocess_input(game.board, side)
sid = Othello.state_id(game.board)
pred = self.network.predict(current_input[np.newaxis, :])
policy = pred[0][0]
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid,
Othello.move_id(move))] = policy[Othello.move_id(move)] / total
for i in range(self.sim_count):
#print("Sim #%d"% i)
clone = deepcopy(game)
current_side = side
visited = deque()
while True:
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
best_move = None
best_move_value = -2
sid = Othello.state_id(clone.board)
for move in possible_moves:
mid = Othello.move_id(move)
qu_val = Q[(sid,
mid)] + P[(sid, mid)] / (N[(sid, mid)] + 1)
if qu_val > best_move_value:
best_move_value = qu_val
best_move = move
#print(best_move)
if N[(sid, Othello.move_id(best_move))] == 0:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner() * side
Q[node] = W[node] / N[node]
break
current_input = self.preprocess_input(
clone.board, current_side)
sid = Othello.state_id(clone.board)
pred = self.network.predict(current_input[np.newaxis, :])
policy = pred[0][0]
value = pred[1][0]
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move)
)] = policy[Othello.move_id(move)] / total
for node in visited:
N[node] += 1
W[node] += value * side
Q[node] = W[node] / N[node]
#print()
break
else:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner() * side
Q[node] = W[node] / N[node]
break
policy = np.zeros((65))
possible_moves = game.possible_moves(side)
sid = Othello.state_id(game.board)
for move in possible_moves:
mid = Othello.move_id(move)
policy[mid] = N[(sid, mid)]
return policy
class FinancialTimeSeriesAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
financial_time_series_input = Input(shape=(nb_time_step, dim_data), name='x1')
lstm_layer_1 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer1')
lstm_layer_21 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss1')
lstm_layer_22 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss2')
lstm_layer_23 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss3')
lstm_layer_24 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss4')
lstm_layer_25 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss5')
h1 = lstm_layer_1(financial_time_series_input)
h21 = lstm_layer_21(h1)
h22 = lstm_layer_22(h1)
h23 = lstm_layer_23(h1)
h24 = lstm_layer_24(h1)
h25 = lstm_layer_25(h1)
time_series_predictions1 = TimeDistributed(Dense(1), name="p1")(h21) # custom 1
time_series_predictions2 = TimeDistributed(Dense(1), name="p2")(h22) # custom 2
time_series_predictions3 = TimeDistributed(Dense(1), name="p3")(h23) # mse
time_series_predictions4 = TimeDistributed(Dense(1, activation='sigmoid'), name="p4")(h24) # logloss
time_series_predictions5 = TimeDistributed(Dense(nb_labels, activation='softmax'), name="p5")(h25) # cross
self.model = Model(input=financial_time_series_input,
output=[time_series_predictions1, time_series_predictions2,
time_series_predictions3, time_series_predictions4,
time_series_predictions5],
name="multi-task deep rnn for financial time series forecasting")
plot(self.model, to_file='model.png')
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, arg_weight=1.):
optimizer = Adam(lr=lr)
loss = [custom_objective1, custom_objective2, 'mse', 'binary_crossentropy', 'categorical_crossentropy']
self.model.compile(optimizer=optimizer, loss=loss)
def fit_model(self, X, y, y_label, epoch=300):
early_stopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0)
self.model.fit(X, [y]*3 + [y > 0] + [y_label], batch_size=self.batch_size, nb_epoch=epoch, validation_split=0.2,
shuffle=True, callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)[0]
count_true = 0
count_all = y.shape[1]
for i in range(y.shape[1]):
count_true = count_true + 1 if y[0,i,0]*y_hat[0,i,0]>0 else count_true
print(y[0,i,0],y_hat[0,i,0])
print(count_all,count_true)
def test_model_methods():
a = Input(shape=(3, ), name='input_a')
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# test fit
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4,
validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_data=([input_a_np,
input_b_np], [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10, ))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer,
loss,
metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer,
loss,
metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=5,
batch_size=4,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([
np.random.random((batch_sz, 3)),
np.random.random((batch_sz, 3))
], [
np.random.random((batch_sz, 4)),
np.random.random((batch_sz, 3))
])
out = model.fit_generator(gen_data(4),
samples_per_epoch=10,
nb_epoch=5,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
def mse_powers(y_true, y_pred):
m = mse(y_true, y_pred)
return {'mse_squared': K.pow(m, 2), 'mse_cubed': K.pow(m, 3)}
model.compile(optimizer,
loss,
metrics=[mse, mse_powers],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * 4 # total loss, per layer: loss + 3 metrics
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4,
nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
class BasicModel(object):
def __init__(self, multi_class=False):
self.multi_class = multi_class
def build(self,
input_shape,
nn_type='Dense',
bidirectional=True,
vat=True):
""" build model
:param input_shape: shape=(number of input rows, 1)
:param nn_type: select 'Dense' or 'RNN' or 'GRU' or 'LSTM'
:param bidirectional: use_flag for Bidirectional rnn
:param vat: use_flag for VAT
:param multi_class: use_flag for multi_class
:return: self
"""
input_layer = Input(input_shape)
output_layer = self.core_data_flow(input_layer, nn_type, bidirectional)
if vat:
self.model = VATModel(input_layer, output_layer).setup_vat_loss()
else:
self.model = Model(input_layer, output_layer)
return self
def core_data_flow(self, input_layer, nn_type, bidirectional):
""" build nn model
:param input_layer: required for Model()
:return: layer
"""
if nn_type == 'Dense':
x = Dense(64, activation='relu')(input_layer)
x = Dropout(0.5)(x)
x = Dense(64, activation='relu')(x)
x = Dropout(0.5)(x)
x = Flatten()(x)
if self.multi_class:
x = Dense(5, activation='softmax')(x)
else:
x = Dense(1, activation='sigmoid')(x)
else:
x = Dense(160)(input_layer)
x = BatchNormalization()(x)
x = LeakyReLU()(x)
if nn_type == 'RNN':
x = Bidirectional(SimpleRNN(256))(x)
elif nn_type == 'GRU':
x = Bidirectional(GRU(256))(x)
elif nn_type == 'LSTM':
x = Bidirectional(LSTM(256))(x)
x = BatchNormalization()(x)
x = LeakyReLU()(x)
if self.multi_class:
x = Dense(5, activation='softmax')(x)
else:
x = Dense(1, activation='sigmoid')(x)
return x
def train(self,
X_train,
X_test,
y_train,
y_test,
batch_size=128,
epochs=100,
early_stop=True):
""" train rnn model
:param X_train, X_test, y_train, y_test: X is feature vectol. y is label
:param batch_size: onece per training size
:param epochs: number of iterations
:param early_stopinput_layer: use_flag for EarlyStopping
:return: history data
"""
if self.multi_class:
self.model.compile(loss='categorical_crossentropy',
optimizer=SGD(lr=0.01,
decay=1e-6,
momentum=0.9,
nesterov=True),
metrics=['accuracy'])
else:
self.model.compile(loss='binary_crossentropy',
optimizer='RMSprop',
metrics=['accuracy'])
np.random.seed(1337) # for reproducibility
if early_stop:
early_stopping = EarlyStopping(monitor='val_loss',
mode='auto',
patience=5)
return self.model.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(X_test, y_test),
callbacks=[early_stopping])
else:
return self.model.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(X_test, y_test))
def predict(self, X):
return self.model.predict(X)
def evaluate(self, X, y):
return self._score(y, self.model.predict_proba(X)[:, 1])
def _score(self, true_label, predicted_prob):
""" calculate the performance score for binary calssification
:param true_label: the ground truth score
:param predicted_label: the predicted probability
:return: a dict of scores
"""
score_dict = dict()
score_dict['AUC'] = metrics.roc_auc_score(true_label, predicted_prob)
predicted_label = [0 if prob < 0.5 else 1 for prob in predicted_prob]
score_dict['Accuracy'] = metrics.accuracy_score(
true_label, predicted_label)
cm = metrics.confusion_matrix(true_label, predicted_label)
score_dict['Confusion Matrix'] = cm
score_dict['TPR'] = cm[1, 1] / float(cm[1, 0] + cm[1, 1])
score_dict['FPR'] = cm[0, 1] / float(cm[0, 0] + cm[0, 1])
return score_dict
class ResiCNN: # My JackNet with residual block
cnn_filter_num = 64
cnn_kernel_size = 3
def __init__(self, channels=3):
self.model = None
self.optimizer = None
self.channels = channels
def bulid(self): # build model
image_in = Input((None, None, self.channels))
conv = Conv2D(filters=self.cnn_filter_num,
kernel_size=self.cnn_kernel_size,
strides=(1, 1),
padding='same',
data_format='channels_last')(image_in)
conv = Activation('relu')(conv)
x = conv
for layers in range(8):
x = self._build_residual_block(x)
conv_out = Conv2D(filters=self.channels,
kernel_size=self.cnn_kernel_size,
strides=(1, 1),
padding='same',
data_format='channels_last')(x)
output = Add()([image_in, conv_out])
self.model = Model(image_in, output, name='model')
def _build_residual_block(self, x): # build residual block
x_in = x
x = Conv2D(filters=self.cnn_filter_num,
kernel_size=self.cnn_kernel_size,
strides=(1, 1),
padding='same',
data_format='channels_last')(x)
x = BatchNormalization(axis=-1)(x)
x = Activation('relu')(x)
x = Conv2D(filters=self.cnn_filter_num,
kernel_size=self.cnn_kernel_size,
strides=(1, 1),
padding='same',
data_format='channels_last')(x)
x = BatchNormalization(axis=-1)(x)
x = Add()([x_in, x])
x = Activation("relu")(x)
return x
def predict(self, x): # denoise on input x
if x.ndim == 3:
x = x.reshape(1, x.shape[0], x.shape[1], self.channels)
return self.model.predict_on_batch(x)
def load(self, config_path, model_path): # load model
print('restore model...')
if os.path.exists(config_path) and os.path.exists(model_path):
with open(config_path, 'r') as fp:
self.model = Model.from_config(json.load(fp))
self.model.load_weights(model_path)
return True
return False
def save(self, config_path, model_path): # save model
with open(config_path, 'w') as fp:
json.dump(self.model.get_config(), fp)
self.model.save_weights(model_path)
def compile(self): # choose adam optimizer and set learning rate
self.optimizer = Adam(lr=1e-2)
self.model.compile(optimizer=self.optimizer, loss=self.loss)
def train_generator(self,
data,
epochs=1,
steps_per_epochs=None,
callbacks=None):
self.model.fit_generator(iter(data),
epochs=epochs,
steps_per_epoch=steps_per_epochs,
callbacks=callbacks)
def train(self, data, epochs=1, callbacks=None):
self.model.fit(x=data[0],
y=data[1],
epochs=epochs,
batch_size=8,
callbacks=callbacks)
@staticmethod
def loss(y_true, y_pred): # loss function, mean square error
return 0.5 * K.sum(K.square(y_pred - y_true), axis=-1)
def training_CNN(model: training.Model,
train,
val,
num_epochs: int,
save=True,
learning_rate=0.5,
verbose=0,
batch_size=32) -> Tuple[History, str]:
'''
cnn.training_CNN()
This function takes the model arcitecture and compiles it using the SDG optimiser and the Binary Cross entropy loss-function.
The function defines the optimzer, the Stocastich Gradient Descent using the learning rate. The Learning rate is described in the thesis.
For information on the two hyper parameters momentum and decay see https://keras.io/api/optimizers/.
The funcition then compiles the model with all the hyper parameters. Here, callback are defined. These are used for plotting, saving ect.
Then, the function is trained on the training data and validataed on the validation data.
Input:
model: the saved model arcitecture from keras. Made using e.g. cnn.s3_model or cnn.s1_model.
train: Training data. Made using e.g. cnn.make_dataset
val: Validataion data. Made using e.g. cnn.make_dataset.
num_epochs: Number of epochs
save[True/false]: If true, the model will be saved to disk.
learning_rate[float]: Value between 0 and 1.
verbose[int]: if verbose =0, nothing will be printed. If verbose=1, information will be printed.
batch_size: hyper parater used in keras.
Output:
history: object containing the training history, the weight, plots and more
weights: the weights as an array.
Example:
history, weights = training_CNN(model, train_data, val_data, 40, save=True,0.001,verbose=1)
history, weights = training_CNN(model, train_data, val_data, 40, save=True,0.9,verbose=1)
Author:
Kristian Soerensen
July 2020
[email protected]
'''
#printing info if verbose >0
if verbose > 0:
print(model.name + ' is being trained.\n')
#setting tensorflow verbosity.
tf.autograph.set_verbosity(0)
#defining learning and loss fucntion.
optimizer1 = keras.optimizers.SGD(lr=learning_rate,
momentum=0.9,
decay=0.01)
#compiling the model with hyper paramters.
model.compile(optimizer=optimizer1,
loss="binary_crossentropy",
metrics=[
"accuracy", "AUC",
tf.keras.metrics.BinaryCrossentropy(),
tf.keras.metrics.Precision(),
tf.keras.metrics.BinaryAccuracy()
])
#path to save weight to..
filepath = 'weights/' + model.name + '.{epoch:02d}-{loss:.2f}.hdf5'
#checkpoint and tensorbord are used in the fitting to get info..
checkpoint = keras.callbacks.ModelCheckpoint(filepath,
monitor='loss',
verbose=verbose,
save_weights_only=True,
save_best_only=True,
mode='auto')
tensor_board = keras.callbacks.TensorBoard(log_dir='logs/',
histogram_freq=0)
#fitting the data to the model!
history = model.fit(train,
batch_size=batch_size,
epochs=num_epochs,
verbose=verbose,
callbacks=[checkpoint, tensor_board],
validation_data=val)
weight_files = glob.glob(os.path.join(os.getcwd(), 'weights/*'))
#making a folder to save models to.
if os.path.exists(model.name) == False:
os.mkdir(model.name)
#saving models and weights..
if save == True:
# convert history dict to pandas to save
hist_df = pd.DataFrame(history.history)
# save to json:
hist_json_file = model.name + '/history.json'
#with open(os.path.join(dirName, '') + 'index.html', 'w') as write_file:
with open(hist_json_file, mode='w') as f:
hist_df.to_json(f)
hist_csv_file = model.name + '/history.csv'
with open(hist_csv_file, mode='w') as f:
hist_df.to_csv(f)
np.save(model.name + '/history.npy', history.history)
#saving entire model! here, we also have the weights and all!!!
model.save(model.name + "/model_try.h5")
print("Model is saved.")
return history, weight_files
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, {'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
class AE(Model):
"""
Autoencoder. This is a simple autoencoder consisting of an encoder and a decoder.
You can use the class like this:
>>> encoder = ...
>>> decoder = ...
>>> ae = Autoencoder(encoder=encoder, decoder=decoder)
>>> ae.compile(...)
>>> ae.fit(...)
"""
def __init__(self, encoder=None, decoder=None, autoencoder=None):
super(AE, self).__init__()
# For calling this as a super-constructor.
parameters = [encoder, decoder]
if all(v is None for v in parameters):
return
# From loading.
if encoder != None and decoder != None and autoencoder != None:
self.encoder = encoder
self.decoder = decoder
self.autoencoder = autoencoder
return
# Check preconditions.
assert len(encoder.outputs) == 1
assert len(decoder.inputs) == 1
assert encoder.outputs[0].shape[1:] == decoder.inputs[0].shape[
1:], str(encoder.outputs[0].shape) + " " + str(
decoder.inputs[0].shape)
self.latent_dim = encoder.outputs[0].shape[1]
self.encoder = encoder
self.decoder = decoder
# Creating the AE.
inputs = self.encoder.inputs[0]
outputs = self.decoder(self.encoder(inputs))
self.autoencoder = Model(inputs, outputs, name='ae')
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
**kwargs):
"""
Compiles the model.
This is the same as compilation in Keras.
"""
assert "reconstruction_loss" not in kwargs, "Not expected to use reconstruction_loss in AE."
self.autoencoder.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics,
**kwargs)
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
**kwargs):
"""
Trains the autoencoder.
"""
return self.autoencoder.fit(x, y, batch_size, epochs, verbose,
callbacks, validation_split,
validation_data, shuffle, class_weight,
sample_weight, initial_epoch,
steps_per_epoch, validation_steps,
**kwargs)
def fit_generator(self,
generator,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0):
"""
Trains the autoencoder with a generator.
"""
return self.autoencoder.fit_generator(
generator,
steps_per_epoch,
epochs,
verbose=verbose,
callbacks=callbacks,
validation_data=validation_data,
validation_steps=validation_steps,
class_weight=class_weight,
max_queue_size=max_queue_size,
workers=workers,
use_multiprocessing=use_multiprocessing,
shuffle=shuffle,
initial_epoch=initial_epoch)
def evaluate(self,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None):
"""
Evaluates the autoencoder.
"""
return self.autoencoder.evaluate(x,
y,
batch_size,
verbose,
sample_weight,
steps=None)
def predict(self, x, batch_size=None, verbose=0, steps=None):
"""
Does a prediction. This is the same as :func:`~ngdlm.models.AE.predict_reconstruct_from_samples`
"""
return self.predict_reconstruct_from_samples(x, batch_size, verbose,
steps)
def predict_reconstruct_from_samples(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Reconstructs samples.
Samples are firstly mapped to latent space using the encoder.
The resulting latent vectors are then mapped to reconstruction space via the decoder.
"""
return self.autoencoder.predict(x, batch_size, verbose, steps)
def predict_embed_samples_into_latent(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Embeds samples into latent space using the encoder.
"""
return self.encoder.predict(x, batch_size, verbose, steps)
def predict_reconstruct_from_latent(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Maps latent vectors to reconstruction space using the decoder.
"""
return self.decoder.predict(x, batch_size, verbose, steps)
def summary(self):
"""
Provides a summary.
"""
print("Encoder:")
self.encoder.summary()
print("Decoder:")
self.decoder.summary()
print("Autoencoder:")
self.autoencoder.summary()
def save(self, path):
"""
Saves the autoencoder.
This includes the whole autoencoder plus the encoder and the decoder.
The encoder and decoder use the path plus a respective annotation.
This code
>>> ae.save("myae.h5")
will create the files *myae.h5*, *myae-encoder.h5*, and *myae-decoder.h5*.
"""
self.autoencoder.save(path)
self.encoder.save(append_to_filepath(path, "-encoder"))
self.decoder.save(append_to_filepath(path, "-decoder"))
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
input_a_df = pd.DataFrame(input_a_np)
input_b_df = pd.DataFrame(input_b_np)
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
output_a_df = pd.DataFrame(output_a_np)
output_b_df = pd.DataFrame(output_b_np)
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np], [output_a_np, output_b_np])
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
out = model.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4)
out = model.fit([input_a_df, input_b_df],
[output_a_df, output_b_df], epochs=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4, validation_split=0.5,
validation_data=(
{'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
out = model.test_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
out = model.predict_on_batch([input_a_df, input_b_df])
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.evaluate([input_a_df, input_b_df], [output_a_df, output_b_df], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
out = model.predict([input_a_df, input_b_df], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), steps_per_epoch=3, epochs=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np],))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer, loss='mse', loss_weights={'dense_1': 0.2, 'dropout': 0.8})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': None, 'dropout': 'temporal'})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
out = model.fit_generator(generator=RandomSequence(3), steps_per_epoch=12, epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
validation_steps=12, callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
class TL(Model):
"""
Triplet-Loss trained Neural Network.
https://arxiv.org/abs/1503.03832
"""
def __init__(self, base=None, siamese=None):
super(TL, self).__init__()
# Store the base model.
assert (base != None)
self.base = base
# For loading.
if base != None and siamese != None:
self.base = base
self.siamese = siamese
self.latent_dim = self.base.outputs[0].shape[1]
return
# Get the latent dimension.
assert len(self.base.outputs) == 1
assert len(self.base.outputs[0].shape) == 2
self.latent_dim = self.base.outputs[0].shape[1]
# Get the input shape.
input_shape = self.base.inputs[0].shape.as_list()[1:]
# Create the anchor.
input_anchor = layers.Input(shape=input_shape)
output_anchor = input_anchor
output_anchor = self.base(output_anchor)
# Create the positive.
input_positive = layers.Input(shape=input_shape)
output_positive = input_positive
output_positive = self.base(output_positive)
# Create the negative.
input_negative = layers.Input(shape=input_shape)
output_negative = input_negative
output_negative = self.base(output_negative)
# Create a dummy output.
output = layers.concatenate(
[output_anchor, output_positive, output_negative])
# Create the model.
self.siamese = Model([input_anchor, input_positive, input_negative],
output,
name="triplet_model")
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
triplet_loss="mse",
**kwargs):
"""
Compiles the TL.
Additionally to the default functionality of *compile*, it adds the triplet-loss.
In order to do so you have to provide it via the parameter *triplet_loss*.
The VAE loss is similar to
>>> vae_loss = max(0.0, pos_dist - neg_dist + alpha)
See the literature for details.
Additional args:
triplet_loss (string): The base-loss for the triplet-loss. Values are either *euclidean* for euclidean norm or *cosine* for cosine similarity.
"""
assert loss == None, "Not expected to provide an explicit loss for TL. Use 'triplet_loss'"
self.triplet_loss = triplet_loss
def triplet_loss_function(y_true, y_pred, alpha=0.4):
anchor = y_pred[:, 0:self.latent_dim]
positive = y_pred[:, self.latent_dim:self.latent_dim * 2]
negative = y_pred[:, self.latent_dim * 2:self.latent_dim * 3]
if triplet_loss == "euclidean":
pos_dist = euclidean_loss(positive, anchor)
neg_dist = euclidean_loss(negative, anchor)
elif triplet_loss == "cosine":
pos_dist = cosine_loss(positive, anchor)
neg_dist = cosine_loss(negative, anchor)
else:
raise Exception("Unexpected: " + triplet_loss)
basic_loss = pos_dist - neg_dist + alpha
loss = K.maximum(basic_loss, 0.0)
return loss
loss = triplet_loss_function
self.siamese.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics, **kwargs)
def fit(self,
x=None,
y=None,
batch_size=None,
minibatch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
**kwargs):
"""
This is basically the same as in vanilla Keras.
Additional args:
minibatch_size (int): The model internally does some sampling. The *minibatch_size* specifies how many candidates to use in order to create a triplet for training.
"""
assert minibatch_size != None, "ERROR! Must provide 'minibatch_size'."
assert steps_per_epoch != None, "ERROR! Must provide 'steps_per_epoch'."
assert validation_steps != None, "ERROR! Must provide 'validation_steps'."
y_dummy = np.zeros((batch_size, self.latent_dim * 3))
# Template generator.
def triplet_loss_generator(x_generator, y_generator, model, sampling):
# Get the classes.
classes = sorted(list(set(y_generator)))
# Sort by classes for easy indexing.
class_indices = {}
for c in classes:
class_indices[c] = []
for index, c in enumerate(y_generator):
class_indices[c].append(index)
# Compute the complements.
class_complements = {}
for c in classes:
class_complements[c] = [c2 for c2 in classes if c2 != c]
# Generator loop.
while True:
x_input_anchors = []
x_input_positives = []
x_input_negatives = []
# Generate a whole batch.
for _ in range(batch_size):
anchor_class = random.choice(classes)
anchor_index = random.choice(class_indices[anchor_class])
anchor_input = x_generator[anchor_index]
#print("anchor_class", anchor_class)
anchor_latent = self.base.predict(
np.expand_dims(anchor_input, axis=0))[0]
# Generate some positive candidates.
positive_candidates = []
while len(positive_candidates) < minibatch_size:
positive_class = anchor_class
positive_index = random.choice(
class_indices[positive_class])
positive_input = x_generator[positive_index]
assert positive_class == y_generator[positive_index]
#print("positive_class", positive_class)
positive_candidates.append(positive_input)
# Find the farthest positive candidate.
positive_candidates = np.array(positive_candidates)
positive_latents = self.base.predict(positive_candidates)
positive_extremum = compute_latent_extremum(
anchor_latent, positive_latents, "argmax",
self.triplet_loss)
positive_input = positive_candidates[positive_extremum]
# Generate some negative candidates.
negative_candidates = []
while len(negative_candidates) < minibatch_size:
negative_class = random.choice(
class_complements[anchor_class])
negative_index = random.choice(
class_indices[negative_class])
negative_input = x_generator[negative_index]
assert negative_class == y_generator[negative_index]
#print("negative_class", negative_class)
negative_candidates.append(negative_input)
# Find the closest negative candidate.
negative_candidates = np.array(negative_candidates)
negative_latents = self.base.predict(negative_candidates)
negative_extremum = compute_latent_extremum(
anchor_latent, negative_latents, "argmin",
self.triplet_loss)
negative_input = negative_candidates[negative_extremum]
# Done.
x_input_anchors.append(anchor_input)
x_input_positives.append(positive_input)
x_input_negatives.append(negative_input)
x_input_anchors = np.array(x_input_anchors)
x_input_positives = np.array(x_input_positives)
x_input_negatives = np.array(x_input_negatives)
x_input = [
x_input_anchors, x_input_positives, x_input_negatives
]
yield x_input, y_dummy
# Create the generators.
training_generator = triplet_loss_generator(x, y, batch_size,
self.siamese)
if validation_data != None:
validation_generator = triplet_loss_generator(
validation_data[0], validation_data[1], batch_size,
self.siamese)
else:
validation_generator = None
# Create the history.
history_keys = ["loss", "val_loss"]
history = {}
for history_key in history_keys:
history[history_key] = []
# Training the model
for epoch in range(epochs):
print("Epoch " + str(epoch + 1) + "/" + str(epochs) + "...")
# Generating data for training.
training_input, training_output = next(training_generator)
if validation_generator != None:
validation_input, validation_output = next(
validation_generator)
model_history = self.siamese.fit(
training_input,
training_output,
validation_data=(validation_input, validation_output),
epochs=1,
steps_per_epoch=steps_per_epoch,
verbose=0,
validation_steps=validation_steps)
# Update the history.
for history_key in history_keys:
history_value = model_history.history[history_key]
history[history_key].append(history_value)
print(history_key, history_value)
return history
def fit_generator(self,
generator,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0):
"""
Coming soon...
"""
print("TODO: implement fit_generator!")
raise Exception("Not implemented!")
return self.siamese.fit_generator(generator, steps_per_epoch, epochs,
verbose, callbacks, validation_data,
validation_steps, class_weight,
max_queue_size, workers,
use_multiprocessing, shuffle,
initial_epoch)
def evaluate(self,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None):
"""
Evaluates the model. Same as vanilla Keras.
"""
return self.siamese.evaluate(x,
y,
batch_size,
verbose,
sample_weight,
steps=None)
def predict(self, x, batch_size=None, verbose=0, steps=None):
"""
Does a prediction. Same as vanilla Keras.
"""
return self.siamese.predict(x, batch_size, verbose, steps)
def summary(self):
"""
Provides a summary.
"""
print("Basemodel:")
self.base.summary()
print("Siamese model:")
self.siamese.summary()
def save(self, path):
"""
Saves the TL.
This includes the whole Siamese Net plus the base-model.
This code
>>> tl.save("myae.h5")
will create the files *tl.h5*, and *tl-base.h5*.
"""
self.siamese.save(path)
self.base.save(append_to_filepath(path, "-base"))
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, {'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), samples_per_epoch=10, nb_epoch=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
def mse_powers(y_true, y_pred):
m = mse(y_true, y_pred)
return {
'mse_squared': K.pow(m, 2),
'mse_cubed': K.pow(m, 3)
}
model.compile(optimizer, loss, metrics=[mse, mse_powers],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * 4 # total loss, per layer: loss + 3 metrics
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
class PolicyValueNet():
"""policy-value network """
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
self._loss_train_op()
if model_file:
net_params = pickle.load(open(model_file, 'rb'))
self.model.set_weights(net_params)
def create_policy_value_net(self):
"""create the policy value network """
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
network = Conv2D(filters=32, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
# action policy layers
policy_net = Conv2D(filters=4, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width*self.board_height, activation="softmax", kernel_regularizer=l2(self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1, activation="tanh", kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""
input: board
output: a list of (action, probability) tuples for each available action and the score of the board state
"""
legal_positions = board.availables
current_state = board.current_state()
act_probs, value = self.policy_value(current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
return act_probs, value[0][0]
def _loss_train_op(self):
"""
Three loss terms：
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
"""
# get the train op
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
loss = self.model.evaluate(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
""" save model params to file """
net_params = self.get_policy_param()
pickle.dump(net_params, open(model_file, 'wb'), protocol=2)
