def train(gan, inputs, epochs=4000, batch_size=32):
for epoch in range(epochs):
# ==========================================================================================================
# train discriminator
# ==========================================================================================================
if hasattr(gan, "embedding_size"):
seed_noises = standard_normal(size=(batch_size, gan.noise_size,
gan.embedding_size)).astype(
dtype="float32")
else:
seed_noises = standard_normal(
size=(batch_size, gan.noise_size)).astype(dtype="float32")
fake_samples = gan.generator.predict(seed_noises)
batch_index = sample(range(inputs.shape[0]), batch_size)
if hasattr(gan, "word2embedded"):
real_samples = gan.word2embedded(inputs[batch_index])
else:
real_samples = inputs[batch_index]
cat_samples = np_concate((real_samples, fake_samples), axis=0)
target = zeros(shape=(batch_size * 2, 2), dtype="int32")
target[:batch_size, 1] = 1
target[batch_size:, 0] = 1
# feed the concatenated samples and corresponding target into discriminator
fix_model(gan.discriminator, is_trainable=True)
dis_loss = gan.discriminator.train_on_batch(x=cat_samples, y=target)
gan.losses["dis_loss"].append(dis_loss[0])
print(('epoch: {}, training discriminator, '
'loss: {:.2f}, accuracy: {:.2f}').format(epoch + 1, *dis_loss))
# ======================================================================================================
# train generator
# ======================================================================================================
if hasattr(gan, "embedding_size"):
seed_noises = standard_normal(size=(batch_size, gan.noise_size,
gan.embedding_size)).astype(
dtype="float32")
else:
seed_noises = standard_normal(
size=(batch_size, gan.noise_size)).astype(dtype="float32")
target = zeros([batch_size, 2], dtype="int32")
target[:, 1] = 1
# train gan with discriminator fixed
fix_model(gan.discriminator, is_trainable=False)
gen_loss = gan.gan_model.train_on_batch(x=seed_noises, y=target)
gan.losses["gen_loss"].append(gen_loss[0])
print(("epoch: {}, training generator, "
"loss: {:.2f}, accuracy: {:.2f}").format(epoch + 1, *gen_loss))
print('-' * 60)
validation_data=(X_test, Y_test))
model.save('/Users/leegho/Documents/pape/nlstm/sentiment_lstm_model.h5')
model.evaluate(X_test, y_test, verbose=True, batch_size=32)
if __name__ == '__main__':
from sklearn.externals import joblib
# 加载数据
with open('/Users/leegho/Desktop/pos_embedding_tensor.pkl', 'rb') as fr:
embedding_tensor_pos = joblib.load(fr)
labels_pos = repeat(0, 10000)
with open('/Users/leegho/Desktop/neg_embedding_tensor.pkl', 'rb') as fr:
embedding_tensor_neg = joblib.load(fr)
labels_neg = repeat(1, 10000)
x = np_concate((embedding_tensor_pos, embedding_tensor_neg), axis=0)
x = x.astype("float32")
y = np_concate((labels_pos, labels_neg), axis=0)
y = y.astype("int32")
# 划分数据集
X_train, X_test, y_train, y_test = train_test_split(x,
y,
test_size=0.7,
stratify=y)
y_train = to_categorical(y_train)
y_test = to_categorical(y_test)
# 训练模型
train_lstm(X_train, y_train, X_test, y_test)
model_filepath = '/Users/leegho/Documents/pape/nlstm/sentiment_lstm_model.h5'
if __name__ == '__main__':
# 加载数据
x_train, x_test = load_file_and_preprocessing()
# 获得词向量
get_train_vecs(x_train, x_test)
# 获得训练集与测试集
train_vecs, y_train, test_vecs, y_test = get_data()
# 训练模型
train_bayes(train_vecs, y_train)
clf = joblib.load(
'/Users/leegho/Documents/pape/bayes_data/sentiment_bayes_model.m')
y_pred = clf.predict(test_vecs)
print(classification_report(y_test, y_pred))
model_filepath = '/Users/leegho/Documents/pape/bayes_data/sentiment_bayes_model.m'
evaluate_bayes(model_filepath, test_vecs, y_test)
# 把训练集连同测试集的预测结果结合，重新训练模型
y_pred = clf.predict(test_vecs)
xt = np_concate((train_vecs, test_vecs), axis=0)
yt = np_concate((y_train, y_pred), axis=0)
train_bayes(xt, yt)
clf = joblib.load(
'/Users/leegho/Documents/pape/bayes_data/sentiment_bayes_model.m')
y_pred = clf.predict(test_vecs)
print(classification_report(y_test, y_pred))
model_filepath = '/Users/leegho/Documents/pape/bayes_data/sentiment_bayes_model.m'
evaluate_bayes(model_filepath, test_vecs, y_test)