def fit(x_train, y_train, plot=True):
#Data preprocessing
if plot:
display_data(x_train, y_train)
x_train = x_train / 255.0
x_train = x_train.reshape([-1, 784])
y_train = one_hot(y_train)
model = Sequential([
Dense(512, input_dim=784),
Activation('relu'),
Dense(512),
Activation('relu'),
Dense(512),
Activation('relu'),
Dense(10),
Activation('softmax')
])
optimiser = keras.optimizers.Adam(learning_rate=0.01)
model.compile(
loss='categorical_crossentropy', #Loss function for one hot inputs
optimizer=optimiser,
metrics=['accuracy'],
)
model.fit(x_train, y_train, validation_split=0.2, batch_size=64, epochs=20)
return model
class NNQ:
def __init__(self, input_space, action_space):
#HyperParameters
self.GAMMA = 0.95
self.LEARNING_RATE = 0.002
self.MEMORY_SIZE = 1000000
self.BATCH_SIZE = 30
self.EXPLORATION_MAX = 1.0
self.EXPLORATION_MIN = 0.01
self.EXPLORATION_DECAY = 0.997
self.exploration_rate = self.EXPLORATION_MAX
self.reward = 0
self.actions = action_space
#Experience Replay
self.memory = deque(maxlen=self.MEMORY_SIZE)
#Create the NN model
self.model = Sequential()
self.model.add(
Dense(64, input_shape=(input_space, ), activation="relu"))
self.model.add(Dense(64, activation="relu"))
self.model.add(Dense(self.actions, activation="softmax"))
self.model.compile(loss="mse", optimizer=Adam(lr=self.LEARNING_RATE))
def act(self, state):
#Exploration vs Exploitation
if np.random.rand() < self.exploration_rate:
return random.randrange(self.actions)
q_values = self.model.predict(state)
return np.argmax(q_values[0])
def remember(self, state, action, reward, next_state, done):
#in every action put in the memory
self.memory.append((state, action, reward, next_state, done))
def experience_replay(self):
#When the memory is filled up take a batch and train the network
if len(self.memory) < self.MEMORY_SIZE:
return
batch = random.sample(self.memory, self.BATCH_SIZE)
for state, action, reward, next_state, terminal in batch:
q_update = reward
if not terminal:
q_update = (
reward +
self.GAMMA * np.amax(self.model.predict(next_state)[0]))
q_values = self.model.predict(state)
q_values[0][action] = q_update
self.model.fit(state, q_values, verbose=0)
if self.exploration_rate > self.EXPLORATION_MIN:
self.exploration_rate *= self.EXPLORATION_DECAY
def fitting(self, model: Sequential) -> Sequential:
model.compile(
optimizer="adam",
loss="sparse_categorical_crossentropy",
metrics=["accuracy"],
)
model.fit(self.data_source.x_train, self.data_source.y_train, epochs=10)
# loss, accuracyを出力してくれる
model.evaluate(self.data_source.x_test, self.data_source.y_test, verbose=2)
return model
def fitting(self, model: Sequential) -> Sequential:
# 予測値はロジットや対数オッズ比で出力される
predictions = model(self.data_source.x_train[:1]).numpy()
# 確率に変換
probability = tf.nn.softmax(predictions).numpy()
# 損失関数。下記の書き方をすればそれぞれの標本についてクラスごとに損失のスカラを返す
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True)
# loss確認する場合はコメント外す
# loss = loss_fn(mnist.y_train[:1], predictions).numpy()
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
model.fit(self.data_source.x_train, self.data_source.y_train, epochs=5)
model.evaluate(self.data_source.x_test,
self.data_source.y_test,
verbose=2)
return model
class Model:
__batch_size = 100
def __init__(self):
self.model = Sequential([
Dense(40, input_shape=(4, ), activation="relu"),
Dense(40, activation="relu"),
Dense(40, activation="relu"),
Dense(4, activation="tanh")
])
# TODO: xavier initialization?
self.model.compile(loss=keras.losses.mean_squared_error,
optimizer=keras.optimizers.Adam(lr=0.001))
self.memory = deque(maxlen=10000)
def experience_replay(self):
if len(self.memory) < self.__batch_size:
return
batch = random.sample(self.memory, self.__batch_size)
states = np.vstack([state for state, _, _, _, _ in batch])
next_states = np.vstack(
[next_state for _, _, _, next_state, _ in batch])
predicted_states = self.model.predict(states)
predicted_next_states = self.model.predict(next_states)
max_nex_state_values = np.max(predicted_next_states, 1)
for index, (_, action, reward, _, terminal) in enumerate(batch):
q_update = reward
if not terminal:
discount_factor = 0.95
q_update += discount_factor * max_nex_state_values[index]
learning_rate = 0.95
predicted_states[index][action] = (
(1 - learning_rate) * predicted_states[index][action] +
learning_rate * q_update)
self.model.fit(states, predicted_states, verbose=0)
n_chars = len(alphabet)
n_in_seq_length = n_numbers * ceil(log10(largest + 1)) + n_numbers - 1
n_out_seq_length = ceil(log10(n_numbers * (largest + 1)))
n_batch = 100
n_epoch = 500
model = Sequential([
LSTM(100, input_shape=(n_in_seq_length, n_chars)),
RepeatVector(n_out_seq_length),
LSTM(50, return_sequences=True),
TimeDistributed(Dense(n_chars, activation='softmax'))
])
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
for i in range(n_epoch):
x, y = generate_data(n_samples, largest, alphabet)
model.fit(x, y, epochs=1, batch_size=n_batch)
model.save('training/keras_classifier.h5')
# evaluate on some new patterns
x, y = generate_data(n_samples, largest, alphabet)
result = model.predict(x, batch_size=n_batch, verbose=0)
# calculate error
expected = [invert(x, alphabet) for x in y]
predicted = [invert(x, alphabet) for x in result]
# show some examples
for i in range(20):
print('Expected=%s, Predicted=%s' % (expected[i], predicted[i]))
#Adam optimisation is a stochastic gradient descent method
#Can handle sparse gradients on noisy problems
optimiser = keras.optimizers.Adam(learning_rate=0.01)
#Training configuration
model.compile(
loss='categorical_crossentropy', #Loss function for one hot inputs
optimizer=optimiser,
metrics=['accuracy'],
)
#Train the model
print("Fit model on training data")
history = model.fit(x_train,
y_train,
validation_split=0.2,
batch_size=64,
epochs=20) #initially 15
#Visualise model
print("Model history keys")
print(history.history.keys())
#Plot model accuracy
plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.title("Model Accuracy")
plt.ylabel('Accuracy')
plt.xlabel('Epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()