X2.append(iclasses[label])
labels.append(0)
categoricals = {"匹配":1, "不匹配":0}
return X1, X2, labels, categoricals
X1, X2, y, classes = convert_to_pairs(X, y, classes)
X1_train = X1[:-1000]
X2_train = X2[:-1000]
y_train = y[:-1000]
X1_test = X1[-1000:]
X2_test = X2[-1000:]
y_test = y[-1000:]
num_classes = len(classes)
tokenizer = SimpleTokenizer()
tokenizer.fit(X1 + X2)
X1_train = tokenizer.transform(X1_train)
X2_train = tokenizer.transform(X2_train)
maxlen = 48
hdims = 128
epochs = 2
X1_train = sequence.pad_sequences(
X1_train,
maxlen=maxlen,
dtype="int32",
padding="post",
truncating="post",
value=0
delta = self.epsilon * grads / (tf.norm(grads) + 1e-6) # 计算扰动
embeddings.assign_add(delta) # 添加扰动到Embedding矩阵
results = super(AdversarialTrainer,
self).train_step(data) # 执行普通的train_step
embeddings.assign_sub(delta) # 删除Embedding矩阵上的扰动
return results
X, y, classes = load_hotel_comment()
X_train, X_test, y_train, y_test = train_test_split(X,
y,
train_size=0.6,
random_state=7384672)
num_classes = len(classes)
tokenizer = SimpleTokenizer()
tokenizer.fit(X_train)
maxlen = find_best_maxlen(X_train)
# maxlen = 256
def create_dataset(X, y, maxlen=maxlen):
X = tokenizer.transform(X)
X = sequence.pad_sequences(X,
maxlen=maxlen,
dtype="int32",
padding="post",
truncating="post",
value=0.0)
y = tf.keras.utils.to_categorical(y)
from pooling import MaskGlobalMaxPooling1D
from pooling import MaskGlobalAveragePooling1D
from dataset import SimpleTokenizer, find_best_maxlen
from dataset import load_THUCNews_title_label
from dataset import load_weibo_senti_100k
from dataset import load_simplifyweibo_4_moods
from dataset import load_hotel_comment
X, y, classes = load_hotel_comment()
X_train, X_test, y_train, y_test = train_test_split(X,
y,
train_size=0.8,
random_state=7384672)
num_classes = len(classes)
tokenizer = SimpleTokenizer()
tokenizer.fit(X_train)
X_train = tokenizer.transform(X_train)
maxlen = 48
maxlen = find_best_maxlen(X_train)
X_train = sequence.pad_sequences(X_train,
maxlen=maxlen,
dtype="int32",
padding="post",
truncating="post",
value=0.0)
y_train = tf.keras.utils.to_categorical(y_train)
num_words = len(tokenizer)
from dataset import load_THUCNews_title_label
from dataset import load_weibo_senti_100k
from dataset import load_simplifyweibo_4_moods
from dataset import load_simplifyweibo_3_moods
from dataset import load_hotel_comment
# 来自Transformer的激活函数，效果略有提升
def gelu(x):
return 0.5 * x * (1.0 + tf.math.erf(x / tf.sqrt(2.0)))
X, y, classes = load_hotel_comment()
X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8, random_state=73672)
class_weight = balance_class_weight(y_train)
num_classes = len(classes)
tokenizer = SimpleTokenizer(min_freq=32)
tokenizer.fit(X_train)
X_train = tokenizer.transform(X_train)
maxlen = 48
maxlen = find_best_maxlen(X_train)
X_train = sequence.pad_sequences(
X_train,
maxlen=maxlen,
dtype="int32",
padding="post",
truncating="post",
value=0
)
y_train = tf.keras.utils.to_categorical(y_train, num_classes)
i, j = split_index(size=len(files))
files_train = files[:i]
files_val = files[i:j]
files_test = files[j:]
# train tokenizer
def Xiter(files):
for content, label in gen(files):
yield content
tokenizer = SimpleTokenizer()
tokenizer.fit(*[Xiter(files)])
class DataGenerator:
def __init__(self, files, loop):
self.files = files
self.loop = loop
def __call__(self):
for _ in range(self.loop):
random.shuffle(self.files)
for content, label in gen(self.files):
content = content[:maxlen]
content = tokenizer.transform([content])[0]
label = tf.keras.utils.to_categorical(label, num_classes)
loss = tf.math.maximum(ploss - nloss + self.margin, 0.0)
self.add_loss(tf.reduce_mean(loss))
return loss
Xa, Xp, Xn, classes = convert_to_triplet(load_lcqmc)
Xa_train = Xa[:-1000]
Xp_train = Xp[:-1000]
Xn_train = Xn[:-1000]
Xa_test = Xa[:-1000]
Xp_test = Xp[:-1000]
Xn_test = Xn[:-1000]
num_classes = len(classes)
tokenizer = SimpleTokenizer()
tokenizer.fit(Xa)
def pad(X, maxlen):
X = sequence.pad_sequences(X,
maxlen=maxlen,
dtype="int32",
padding="post",
truncating="post",
value=0)
return X
def create_dataset(Xa, Xp, Xn, maxlen):
Xa = tokenizer.transform(Xa)