在预测期间使用来自 tensorflow hub 的 Elmo 作为自定义 tf.keras 层的问题

Question

我正在尝试将来自 tensorflow hub 的 Elmo 与 tf.keras 一起使用来执行 NER。训练很好，损失在减少，测试集也给出了很好的结果。但我无法预测，因为出现以下错误：

2019-05-02 15:41:42.785946: I tensorflow/stream_executor/dso_loader.cc:152] successfully opened CUDA library libcublas.so.10.0 locally
Traceback (most recent call last):
  File "elmo_eva_brain.py", line 668, in <module>
    np.array([['hello', 'world'] + ['--PAD--'] * 18])))
  File "/home/ashwanipandey/eva_ml/experimental/eva_brain/venv/lib64/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 1113, in predict
    self, x, batch_size=batch_size, verbose=verbose, steps=steps)
  File "/home/ashwanipandey/eva_ml/experimental/eva_brain/venv/lib64/python3.6/site-packages/tensorflow/python/keras/engine/training_arrays.py", line 329, in model_iteration
    batch_outs = f(ins_batch)
  File "/home/ashwanipandey/eva_ml/experimental/eva_brain/venv/lib64/python3.6/site-packages/tensorflow/python/keras/backend.py", line 3076, in __call__
    run_metadata=self.run_metadata)
  File "/home/ashwanipandey/eva_ml/experimental/eva_brain/venv/lib64/python3.6/site-packages/tensorflow/python/client/session.py", line 1439, in __call__
    run_metadata_ptr)
  File "/home/ashwanipandey/eva_ml/experimental/eva_brain/venv/lib64/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 528, in __exit__
    c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: len(seq_lens) != input.dims(0), (256 vs. 1)
         [[{{node Embed/elmo/elmo_module_apply_tokens/bilm/ReverseSequence}}]]
         [[{{node Tag/t_output/transpose_1}}]]

256 是我训练时的批量大小。我试图预测一个句子。

我试着在网上搜索了很多，但都是徒劳的。任何帮助深表感谢。如果我重复我的向量 256 次并在预测期间将 batch_size 设置为 256，我绝对可以获得预测。但正如您所见，这是一种非常低效的解决方法。

这是自定义图层的代码

class ElmoEmbeddingLayer(keras.layers.Layer):
    def __init__(self, dimensions=1024, batch_size=512, word_size=20, **kwargs):
        self.dimensions = 1024
        self.trainable = True
        self.batch_size = _BATCH_SIZE
        self.word_size = _WORD_SIZE
        super().__init__(**kwargs)

    def build(self, input_shape):
        self.elmo = hub.Module('https://tfhub.dev/google/elmo/2', trainable=self.trainable,
                               name=f"{self.name}_module")
        super().build(input_shape)

    def call(self, x, mask=None):
        result = self.elmo(inputs={
            "tokens": K.cast(x, tf.string),
            "sequence_len": K.constant(self.batch_size*[self.word_size], dtype=tf.int32)
        },
            as_dict=True,
            signature='tokens',
        )['elmo']
        return result

    def compute_mask(self, inputs, mask=None):
        return K.not_equal(inputs, '--PAD--')

    def compute_output_shape(self, input_shape):
        return (None, self.word_size, self.dimensions)

    def get_config(self):
        config = {
            'dimensions': self.dimensions,
            'trainable': self.trainable,
            'batch_size': self.batch_size,
            'word_size': self.word_size
        }
        base_config = super().get_config()
        return dict(list(base_config.items()) + list(config.items()))

这是我的模型架构： model architecture

Answer 1

样本数量（在训练和测试集中）必须能被 batch_size 整除。否则 keras 中的最后一批将破坏架构。因此，例如，一种解决方案是使用 split_tr 之前的样本进行训练，使用 split_te 进行预测：

split_tr = (X_train.shape[0]//BATCH_SIZE)*BATCH_SIZE
split_te = (X_test.shape[0]//BATCH_SIZE)*BATCH_SIZE
model.fit(X_train_text[:split_tr], y_train[:split_tr], batch_size=BATCH_SIZE, epochs=15, validation_data=(X_test_text[:split_te], y_test[:split_te]), verbose=1)

Answer 2

我和你有同样的问题，在 RNN ELMo post-tagger 模型上工作。最后我按照解决方案批量预测并保留我想要的测试样本：

model.predict([X_test[:split_te]], batch_size=256)[0]

有关更多想法（例如复制权重），请查看 here！

在预测期间使用来自 tensorflow hub 的 Elmo 作为自定义 tf.keras 层的问题

Problem using Elmo from tensorflow hub as custom tf.keras layer during prediction

python-3.x

tensorflow

tensorflow-hub

tf.keras