同一数据的单热编码数组永远不会出现问题，但创建嵌入式数组会导致GPU内存页面出错，并且在训练模型时会出现CUDA_ERROR_ILLEGAL_ADDRESS错误。
在发生错误之前，训练有时可以进行几个时期。
完整登录errors/CUDA_ERROR_ILLEGAL_ADDRESS.txt。
处的完整python文件创建嵌入式aray的功能：
def char2vec(dataset):
"""Convert dataset into an integer array for an Embedding layer
x: Embedded array
y: one hot encoding array
:param dataset:
:return: x, y, samples, timesteps, features, char_to_int, int_to_char
"""
try:
raw_text = open(dataset, 'r').read().lower()
print('[*]', dataset)
except:
raise
chars = sorted(list(set(raw_text)))
char_to_int = dict((c, i) for i, c in enumerate(chars))
int_to_char = dict((i, c) for i, c in enumerate(chars))
nb_chars = raw_text.__len__()
features = chars.__len__()
timesteps = seq_length
# cut the text in semi-redundant sequences of seq_length
step = 3
X = []
Y = []
for i in range(0, nb_chars - seq_length, step):
X.append(raw_text[i: i + seq_length])
Y.append(raw_text[i + seq_length])
samples = X.__len__()
print('[*] Corpus Length:', nb_chars) # 163817
print('[*] Features:', features) # 61
print('[*] Samples:', samples) # 163761
print('[*] Timestep:', seq_length) # 56
# https://github.com/minimaxir/char-embeddings/blob/master/text_generator_keras.py#L48
# x = np.zeros((len(sentences), maxlen), dtype=np.int)
# y = np.zeros((len(sentences), len(chars)), dtype=np.bool)
# for i, sentence in enumerate(sentences):
# for t, char in enumerate(sentence):
# X[i, t] = char_indices[char]
# y[i, char_indices[next_chars[i]]] = 1
print('[*] Vectorization...')
x = np.zeros((samples, seq_length), dtype=np.int32)
y = np.zeros((samples, features), dtype=np.bool)
for i, sentence in enumerate(X):
for t, char in enumerate(sentence):
x[i, t] = char_to_int[char]
y[i, char_to_int[Y[i]]] = 1
return x, y, samples, timesteps, features, char_to_int, int_to_char
型号：
model = Sequential()
model.add(Embedding(output_dim=64, input_dim=features))
model.add(Dropout(0.2))
model.add(LSTM(128, return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(64))
model.add(Dropout(0.2))
model.add(Dense(features))
model.add(Activation('relu'))
model.compile(loss='categorical_crossentropy',
optimizer='adam', metrics=['accuracy'])
model.fit(x, y,
batch_size=batch_size,
epochs=epochs,
verbose=1,
callbacks=callbacks_list,
# validation_data=(x_val, y_val),
# validation_split=0.33,
shuffle=False,
initial_epoch=initial_epoch)