def mlp(inputs_train, outputs_train, inputs_test, outputs_test, loss):
"""
Fully connected neural network.
@param loss: 'categorical_crossentropy' or 'mean_squared_error'
"""
# Split to train-set and validation-set
split_at = len(inputs_train) - len(inputs_train) * 2 // 10
(inputs_train, inputs_validation) = \
(inputs_train[:split_at], inputs_train[split_at:])
(outputs_train, outputs_validation) = \
(outputs_train[:split_at], outputs_train[split_at:])
# Build MLP model
model = Sequential()
model.add(Dense(64, activation='relu', input_dim=inputs_train.shape[1]))
model.add(Dropout(0.5))
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(outputs_train.shape[1], activation='softmax'))
print(model.summary())
# compile
model.compile(loss=loss, optimizer='adam', metrics=['accuracy'])
# train
if loss == 'categorical_crossentropy':
early_stopping = EarlyStopping(min_delta=0.001,
patience=5,
restore_best_weights=True)
elif loss == 'mean_squared_error':
early_stopping = EarlyStopping(min_delta=0.0003,
patience=5,
restore_best_weights=True)
else:
raise ValueError(
"loss should be 'categorical_crossentropy' or 'mean_squared_error'."
)
model.fit(inputs_train,
outputs_train,
epochs=100,
batch_size=256,
validation_data=(inputs_validation, outputs_validation),
callbacks=[early_stopping])
# evaluate
outputs_test_pred = np.asarray(model.predict(inputs_test))
acc_eval = accuracy(outputs_test, outputs_test_pred)
fscore_eval = fscore(outputs_test, outputs_test_pred)
coef_eval = coef(outputs_test, outputs_test_pred)
print("Evaluation: acc - %.4f - fscore: %.4f - coef: %.4f - pvalue: %.4f" %
(acc_eval, fscore_eval, coef_eval[0], coef_eval[1]))
# return model
return model
def verify_model(model_file_name, inputs_test, outputs_test):
# load model
model_file_path = os.path.join(os.path.pardir, "models",
model_file_name + ".h5")
model = load_model(model_file_path)
print(model_file_name)
# model structure
print(model.summary())
# model evaluation
outputs_test_pred = np.asarray(model.predict(inputs_test))
acc_eval = accuracy(outputs_test, outputs_test_pred)
fscore_eval = fscore(outputs_test, outputs_test_pred)
coef_eval = coef(outputs_test, outputs_test_pred)
print("Evaluation: acc - %.4f - fscore: %.4f - coef: %.4f - pvalue: %.4f" %
(acc_eval, fscore_eval, coef_eval[0], coef_eval[1]))
def rnn(inputs_train, outputs_train, inputs_test, outputs_test, loss,
train_embedding):
"""
Recurrent neural network.
@param loss: 'classification' or 'regression'
@param train_embedding: 0 - initialize with word_embedding_matrix, trainable=False
1 - initialize with word_embedding_matrix, trainable=True
2 - initialize with random matrix, trainable=True
"""
# Load word-embedding matrix
word_embedding_matrix_file_path = os.path.join(os.path.pardir, "data",
"word-embedding_matrix")
with open(word_embedding_matrix_file_path,
'rb') as word_embedding_matrix_file:
word_embedding_matrix = np.load(word_embedding_matrix_file)
# Split to train-set and validation-set
split_at = len(inputs_train) - len(inputs_train) * 2 // 10
(inputs_train, inputs_validation) = \
(inputs_train[:split_at], inputs_train[split_at:])
(outputs_train, outputs_validation) = \
(outputs_train[:split_at], outputs_train[split_at:])
# Build RNN model
if train_embedding == 0:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
weights=[word_embedding_matrix],
input_length=inputs_train.shape[1],
trainable=False)
elif train_embedding == 1:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
weights=[word_embedding_matrix],
input_length=inputs_train.shape[1],
trainable=True)
elif train_embedding == 2:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
input_length=inputs_train.shape[1],
trainable=True)
else:
raise ValueError("train_embedding should be 0 or 1 or 2.")
model = Sequential()
model.add(embedding_layer)
model.add(SimpleRNN(128, unroll=True))
model.add(Dropout(0.5))
model.add(Dense(outputs_train.shape[1], activation='softmax'))
print(model.summary())
# compile
model.compile(loss=loss, optimizer='adam', metrics=['accuracy'])
# train
if loss == 'categorical_crossentropy':
early_stopping = EarlyStopping(min_delta=0.005,
patience=3,
restore_best_weights=True)
elif loss == 'mean_squared_error':
early_stopping = EarlyStopping(min_delta=0.0005,
patience=3,
restore_best_weights=True)
else:
raise ValueError(
"loss should be 'categorical_crossentropy' or 'mean_squared_error'."
)
model.fit(inputs_train,
outputs_train,
epochs=100,
batch_size=128,
validation_data=(inputs_validation, outputs_validation),
callbacks=[early_stopping])
# evaluate
outputs_test_pred = np.asarray(model.predict(inputs_test))
acc_eval = accuracy(outputs_test, outputs_test_pred)
fscore_eval = fscore(outputs_test, outputs_test_pred)
coef_eval = coef(outputs_test, outputs_test_pred)
print("Evaluation: acc - %.4f - fscore: %.4f - coef: %.4f - pvalue: %.4f" %
(acc_eval, fscore_eval, coef_eval[0], coef_eval[1]))
# return model
return model
def cnn(inputs_train, outputs_train, inputs_test, outputs_test, loss,
train_embedding):
"""
Convolutional neural network for sentence classification.
See "Kim, Y. (2014). Convolutional Neural Networks for Sentence Classification. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2014), 1746–1751." for details.
@param loss: 'categorical_crossentropy' or 'mean_squared_error'
@param train_embedding: 0 - initialize with word_embedding_matrix, trainable=False
1 - initialize with word_embedding_matrix, trainable=True
2 - initialize with random matrix, trainable=True
"""
# Load word-embedding matrix
word_embedding_matrix_file_path = os.path.join(os.path.pardir, "data",
"word-embedding_matrix")
with open(word_embedding_matrix_file_path,
'rb') as word_embedding_matrix_file:
word_embedding_matrix = np.load(word_embedding_matrix_file)
# Split to train-set and validation-set
split_at = len(inputs_train) - len(inputs_train) * 2 // 10
(inputs_train, inputs_validation) = \
(inputs_train[:split_at], inputs_train[split_at:])
(outputs_train, outputs_validation) = \
(outputs_train[:split_at], outputs_train[split_at:])
# Build CNN model
if train_embedding == 0:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
weights=[word_embedding_matrix],
input_length=inputs_train.shape[1],
trainable=False)
elif train_embedding == 1:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
weights=[word_embedding_matrix],
input_length=inputs_train.shape[1],
trainable=True)
elif train_embedding == 2:
embedding_layer = Embedding(word_embedding_matrix.shape[0],
word_embedding_matrix.shape[1],
input_length=inputs_train.shape[1],
trainable=True)
else:
raise ValueError("train_embedding should be 0 or 1 or 2.")
model_input = Input(shape=(inputs_train.shape[1], ))
embedding = embedding_layer(model_input)
conv_blocks = []
for kernel_size in range(2, 8):
for _ in range(2): # 2 filters for each kernel_size
conv = Conv1D(4, kernel_size, activation="relu")(embedding)
conv = MaxPooling1D(3)(conv)
conv = Dropout(0.25)(conv)
conv = Flatten()(conv)
conv_blocks.append(conv)
x = Concatenate()(conv_blocks)
x = Dense(128, activation="relu")(x)
x = Dropout(0.5)(x)
model_output = Dense(outputs_train.shape[1], activation="softmax")(x)
model = Model(model_input, model_output)
print(model.summary())
# compile
model.compile(loss=loss, optimizer='adam', metrics=['accuracy'])
# train
if loss == 'categorical_crossentropy':
early_stopping = EarlyStopping(min_delta=0.003,
patience=5,
restore_best_weights=True)
elif loss == 'mean_squared_error':
early_stopping = EarlyStopping(min_delta=0.0003,
patience=5,
restore_best_weights=True)
else:
raise ValueError(
"loss should be 'categorical_crossentropy' or 'mean_squared_error'."
)
model.fit(inputs_train,
outputs_train,
epochs=100,
batch_size=128,
validation_data=(inputs_validation, outputs_validation),
callbacks=[early_stopping])
# evaluate
outputs_test_pred = np.asarray(model.predict(inputs_test))
acc_eval = accuracy(outputs_test, outputs_test_pred)
fscore_eval = fscore(outputs_test, outputs_test_pred)
coef_eval = coef(outputs_test, outputs_test_pred)
print("Evaluation: acc - %.4f - fscore: %.4f - coef: %.4f - pvalue: %.4f" %
(acc_eval, fscore_eval, coef_eval[0], coef_eval[1]))
# return model
return model