def train():
# 指定样本文件
positive_data_file = "./rt-polaritydata/rt-polarity.pos"
negative_data_file = "./rt-polaritydata/rt-polarity.neg"
# 设置训练参数
num_steps = 50 # 定义训练次数
SaveFileName = "text_cnn_model" # 定义保存模型文件夹名称
# 设置模型参数
num_classes = 2 # 设置模型分类
l2_reg_lambda = 0.1 # 定义正则化系数
filter_sizes = "3,4,5" # 定义多通道卷积核
num_filters = 64 # 定义每通道的输出个数
# 加载数据集
data, vocab_processor, max_len = dataset(positive_data_file,
negative_data_file)
#搭建模型
text_cnn = TextCNN(seq_length=max_len,
num_classes=num_classes,
vocab_size=len(vocab_processor.vocabulary_),
embeding_size=128,
filter_sizes=list(map(int, filter_sizes.split(','))),
num_filters=num_filters)
def l2_loss(y_true, y_pred):
l2_loss = tf.constant(0.0)
for tf_var in text_cnn.trainable_weights:
if tf_var.name == "fully_connecred":
l2_loss += tf.reduce_mean(tf.nn.l2_loss(tf_var))
loss = tf.nn.softmax_cross_entropy_with_logits(logits=y_pred,
labels=y_true)
return loss + l2_reg_lambda * l2_loss
text_cnn.compile(loss=l2_loss,
optimizer=tf.keras.optimizers.Adam(lr=1e-3),
metrics=['acc'])
text_cnn.fit(data, epochs=num_steps)
text_cnn.save("textcnn.h5")
logger.info('loading data...')
try:
(x_train, y_train), (x_test,
y_test) = imdb.load_data(num_words=max_features)
except:
logger.info('np bug occur...')
(x_train, y_train), (x_test, y_test) = load_data(num_words=max_features)
logger.info('train data length: {}'.format(len(x_train)))
logger.info('test data length: {}'.format(len(x_test)))
logger.info('padding data...')
x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
logger.info('build model...')
model = TextCNN(max_features=max_features, maxlen=maxlen,
emb_dim=emb_dim).build_model()
logger.info('training...')
earlystop = EarlyStopping(patience=3, mode='max', monitor='val_acc')
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
callbacks=[earlystop],
validation_data=(x_test, y_test))
logger.info('test...')
pred = model.predict(x_test[:10])
logger.info(list(zip(pred, y_test[:10])))
# training hyperparameter
BATCH_SZIE = 128
EPOCHS = 10
# load data
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=MAX_WORD_NUM)
# padding sequence
x_train = sequence.pad_sequences(x_train, maxlen=MAX_SENT_LEN)
x_test = sequence.pad_sequences(x_test, maxlen=MAX_SENT_LEN)
# build model
model = TextCNN(max_sent_len=MAX_SENT_LEN,
max_word_num=MAX_WORD_NUM,
embedding_dims=EMBEDDING_DIMS,
class_num=CLASS_NUM,
last_activation=LAST_ACTIVATION).build_model()
model.compile('adam', 'binary_crossentropy', metrics=['accuracy'])
# train
early_stopping = EarlyStopping(monitor='val_acc', patience=3, mode='max')
model.fit(x_train,
y_train,
batch_size=BATCH_SZIE,
epochs=EPOCHS,
callbacks=[early_stopping],
validation_data=(x_test, y_test))
# save model
# model.save('textcnn_model.h5')
