了解 Trax 中变压器的介绍性示例

Question

我的目标是理解 Trax 中变压器的介绍性示例，

import trax

# Create a Transformer model.

# Pre-trained model config in gs://trax-ml/models/translation/ende_wmt32k.gin

model = trax.models.Transformer(

    input_vocab_size=33300,

    d_model=512, d_ff=2048,

    n_heads=8, n_encoder_layers=6, n_decoder_layers=6,

    max_len=2048, mode='predict')

# Initialize using pre-trained weights.

model.init_from_file('gs://trax-ml/models/translation/ende_wmt32k.pkl.gz',

                     weights_only=True)

# Tokenize a sentence.

sentence = 'It is nice to learn new things today!'

tokenized = list(trax.data.tokenize(iter([sentence]),  # Operates on streams.

                                    vocab_dir='gs://trax-ml/vocabs/',

                                    vocab_file='ende_32k.subword'))[0]

# Decode from the Transformer.

tokenized = tokenized[None, :]  # Add batch dimension.

tokenized_translation = trax.supervised.decoding.autoregressive_sample(

    model, tokenized, temperature=0.0)  # Higher temperature: more diverse results.

# De-tokenize,

tokenized_translation = tokenized_translation[0][:-1]  # Remove batch and EOS.

translation = trax.data.detokenize(tokenized_translation,

                                   vocab_dir='gs://trax-ml/vocabs/',

                                   vocab_file='ende_32k.subword')

print(translation)

这个例子工作得很好。但是，当我尝试使用初始化模型翻译另一个示例时，例如

sentence = 'I would like to try another example.'

tokenized = list(trax.data.tokenize(iter([sentence]),

                                    vocab_dir='gs://trax-ml/vocabs/',

                                    vocab_file='ende_32k.subword'))[0]

tokenized = tokenized[None, :]

!我在本地机器和 Google Colab 上都得到了输出。其他示例也会发生同样的情况。

当我构建并初始化一个新模型时，一切正常。

这是一个错误吗？如果不是，这里发生了什么，我怎样才能避免/修复这种行为？

Tokenization 和 detokenization 似乎运行良好，我对此进行了调试。. 中的事情似乎出了问题/出乎意料trax.supervised.decoding.autoregressive_sample。

Answer 1

我自己发现的……需要重置模型的state. 所以下面的代码对我有用：def translate(model, sentence, vocab_dir, vocab_file):    empty_state = model.state # save empty state    tokenized_sentence = next(trax.data.tokenize(iter([sentence]), vocab_dir=vocab_dir,                                                 vocab_file=vocab_file))    tokenized_translation = trax.supervised.decoding.autoregressive_sample(        model, tokenized_sentence[None, :], temperature=0.0)[0][:-1]    translation = trax.data.detokenize(tokenized_translation, vocab_dir=vocab_dir,                                       vocab_file=vocab_file)    model.state = empty_state # reset state    return translation# Create a Transformer model.# Pre-trained model config in gs://trax-ml/models/translation/ende_wmt32k.ginmodel = trax.models.Transformer(input_vocab_size=33300, d_model=512, d_ff=2048, n_heads=8,                                n_encoder_layers=6, n_decoder_layers=6, max_len=2048,                                mode='predict')# Initialize using pre-trained weights.model.init_from_file('gs://trax-ml/models/translation/ende_wmt32k.pkl.gz',                     weights_only=True)print(translate(model, 'It is nice to learn new things today!',                vocab_dir='gs://trax-ml/vocabs/', vocab_file='ende_32k.subword'))print(translate(model, 'I would like to try another example.',                vocab_dir='gs://trax-ml/vocabs/', vocab_file='ende_32k.subword'))