import os
os.environ["KERAS_BACKEND"] = "tensorflow" # @param ["tensorflow", "jax", "torch"]
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import keras_nlp
import keras
import tensorflow as tf
import sj_utils
sj_utils.use_gpu()
BATCH_SIZE = 64
MIN_STRING_LEN = 512 # Strings shorter than this will be discarded
SEQ_LEN = 128 # Length of training sequences, in tokens
Model
EMBED_DIM = 256
FEED_FORWARD_DIM = 128
NUM_HEADS = 3
NUM_LAYERS = 2
VOCAB_SIZE = 5000 # 限制模型参数
Training
EPOCHS = 5
Inference
NUM_TOKENS_TO_GENERATE = 80
keras.utils.get_file(
origin="https://dldata-public.s3.us-east-2.amazonaws.com/simplebooks.zip",
extract=True,
)
dir = os.path.expanduser("datasets/simplebooks")
Load simplebooks-92 train set and filter out short lines.
raw_train_ds = (
tf.data.TextLineDataset(dir + "/simplebooks-92-raw/train.txt")
.filter(lambda x: tf.strings.length(x) > MIN_STRING_LEN)
.batch(BATCH_SIZE)
.shuffle(buffer_size=256)
)
Load simplebooks-92 validation set and filter out short lines.
raw_val_ds = (
tf.data.TextLineDataset(dir + "/simplebooks-92-raw/valid.txt")
.filter(lambda x: tf.strings.length(x) > MIN_STRING_LEN)
.batch(BATCH_SIZE)
)
我们从训练数据集中训练分词器,以得到一个词汇大小(VOCAB_SIZE),这是一个经过调整的超参数。我们希望尽可能限
制词汇表的大小,因为稍后会看到,这对模型参数的数量有很大影响。同时,我们也不想包含太少的词汇项,否则会产生太
多的未知词汇(Out-Of-Vocabulary, OOV)子词。此外,词汇表中还预留了三个标记:"[PAD]":用于将序列填充到固
定长度(SEQ_LEN)。这个标记在vocab[0]和其他层中都以索引0出现,因为(以及其他层)将<pad>/[PAD]视为默认的填
充标记。"[UNK]":代表OOV子词,它应该与WordPieceTokenizer中的默认设置相匹配。这里oov_token="[UNK]"指定
了未知词汇的标记。 "[BOS]":代表句子的开始,但在这里,从技术上讲,它是一个标记,代表训练数据中每一行的开始。
通过这些设计,我们可以有效地控制模型的大小和泛化能力,同时确保模型能够处理训练数据中可能遇到的各种情况,包括那
些不在初始词汇表中的词汇。使用WordPieceTokenizer这样的分词器还可以帮助我们处理未知词汇,通过将它们分解为已
知的子词组合,从而提高模型的覆盖率和准确性。
训练标记器词汇
vocab = keras_nlp.tokenizers.compute_word_piece_vocabulary(
raw_train_ds,
vocabulary_size=VOCAB_SIZE,
lowercase=True,
reserved_tokens=["[PAD]", "[UNK]", "[BOS]"], # 保留token:填充,未知,起始
)
WordPieceTokenizer是BERT和其他模型使用的WordPiece算法的有效实现。
tokenizer = keras_nlp.tokenizers.WordPieceTokenizer(
vocabulary=vocab,
sequence_length=SEQ_LEN,
lowercase=True,
)
packer adds a start token
start_packer = keras_nlp.layers.StartEndPacker(
sequence_length=SEQ_LEN,
start_value=tokenizer.token_to_id("[BOS]"),
)
def preprocess(inputs):
outputs = tokenizer(inputs)
加了开始符,但是长度还是128,这样原来的最后一个token会被截断
#相当于之前的[:-1]
features = start_packer(outputs)
原始分词,长度128,相当于之前的[1:]
labels = outputs
#labels是1--n+1,features是0--n,这样刚好是预测下一个token
return features, labels
Tokenize and split into train and label sequences.
train_ds = raw_train_ds.map(preprocess, num_parallel_calls=tf.data.AUTOTUNE).prefetch(
tf.data.AUTOTUNE
)
for f,l in train_ds.take(1):
print(f[0])
print(l[0])
break
val_ds = raw_val_ds.map(preprocess, num_parallel_calls=tf.data.AUTOTUNE).prefetch(
tf.data.AUTOTUNE
)
inputs = keras.layers.Input(shape=(None,), dtype="int32")
Embedding.
embedding_layer = keras_nlp.layers.TokenAndPositionEmbedding(
vocabulary_size=VOCAB_SIZE,
sequence_length=SEQ_LEN,
embedding_dim=EMBED_DIM,
mask_zero=True,
)
x = embedding_layer(inputs)
transformer decode层,只有自注意力和前馈层,不包括跨注意力,因为不是翻译任务
#没有源序列,只有目标序列,就是根据目标序列前i个token预测第i+1个token
for _ in range(NUM_LAYERS): #有两层,所以x被解码两次
decoder_layer = keras_nlp.layers.TransformerDecoder(
num_heads=NUM_HEADS,
intermediate_dim=FEED_FORWARD_DIM,
)
只传一个目标序列输入,而不传编码器输出的输入,是因为任务的特殊性,文本生成,只需要
基于目标序列预测下个token,所以它只包括解码器部分,为了用因果掩码来让机器学会预测
下一个token
x = decoder_layer(x)
outputs = keras.layers.Dense(VOCAB_SIZE)(x)
model = keras.Model(inputs=inputs, outputs=outputs)
Perplexity(或称为困惑度)是自然语言处理(NLP)中一个常见的评估指标,特别是在语言模型(Language
Model, LM)的评估中。它衡量了模型对测试数据的预测能力,即模型对测试集中每个词(或标记)预测的不确定
性。困惑度越低,表示模型对测试数据的预测能力越好,不确定性越低。
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
perplexity = keras_nlp.metrics.Perplexity(from_logits=True, mask_token_id=0)
model.compile(optimizer="adam", loss=loss_fn, metrics=[perplexity])
让我们来看一下我们的模型概要------绝大多数参数都位于(嵌入层和)输出层!这意味着词汇表大小(VOCAB_SIZE)
对模型参数的大小有很大影响,而Transformer解码器层的数量(NUM_LAYERS)对其影响则不那么显著。token_and
_position_embedding(词嵌入和位置嵌入)与dense(密集层)共同贡献了这些参数。
model.fit(train_ds, validation_data=val_ds, epochs=EPOCHS)
model.save_weights('./checkpoints/gpt_text_generate_12.weights.h5')
def next(prompt, cache, index):
logits = model(prompt)[:, index - 1, :]
Ignore hidden states for now; only needed for contrastive search.
hidden_states = None
return logits, hidden_states, cache
sampler = keras_nlp.samplers.GreedySampler()
output_tokens = sampler(
next=next,
prompt=prompt_tokens,
index=1, # Start sampling immediately after the [BOS] token.
)
txt = tokenizer.detokenize(output_tokens)
print(f"Greedy search generated text: \n{txt}\n")
sampler = keras_nlp.samplers.BeamSampler(num_beams=10)
output_tokens = sampler(
next=next,
prompt=prompt_tokens,
index=1,
)
txt = tokenizer.detokenize(output_tokens)
print(f"Beam search generated text: \n{txt}\n")
sampler = keras_nlp.samplers.RandomSampler()
output_tokens = sampler(
next=next,
prompt=prompt_tokens,
index=1,
)
txt = tokenizer.detokenize(output_tokens)
print(f"Random search generated text: \n{txt}\n")
sampler = keras_nlp.samplers.TopKSampler(k=10)
output_tokens = sampler(
next=next,
prompt=prompt_tokens,
index=1,
)
txt = tokenizer.detokenize(output_tokens)
print(f"Top-K search generated text: \n{txt}\n")
sampler = keras_nlp.samplers.TopPSampler(p=0.5)
output_tokens = sampler(
next=next,
prompt=prompt_tokens,
index=1,
)
txt = tokenizer.detokenize(output_tokens)
print(f"Top-P search generated text: \n{txt}\n")
class TopKTextGenerator(keras.callbacks.Callback):
"""A callback to generate text from a trained model using top-k."""
def init(self, k):
self.sampler = keras_nlp.samplers.TopKSampler(k)
def on_epoch_end(self, epoch, logs=None):
output_tokens = self.sampler(
next=next,
prompt=prompt_tokens,
index=1,
)
txt = tokenizer.detokenize(output_tokens)
print(f"Top-K search generated text: \n{txt}\n")
text_generation_callback = TopKTextGenerator(k=10)
Dummy training loop to demonstrate callback.
model.fit(train_ds.take(1), verbose=2, epochs=2, callbacks=[text_generation_callback])