Question

我正在使用LSTM + CRF进行实体识别任务。我使用tensorflow estimator API构建，下面是用于训练和评估的input_fn代码。

原始输入功能首先通过tf.train.SequenceExample进行序列化。输入功能具有动态填充，可用于动态DSTM的每个批次。 char_inputs，seg_inputs具有预期的形状[batch_size，max_length]

input_features = {
    CHAR_INPUT_TENSOR_NAME: tf.FixedLenSequenceFeature([],  tf.int64, allow_missing=True),
    SEG_INPUT_TENSOR_NAME: tf.FixedLenSequenceFeature([], tf.int64, allow_missing=True)
}

context_features = {
    "length": tf.FixedLenFeature([], dtype=tf.int64)
}

def training_input_fn(tfrecord_file):
    # read from TFRecord, serialized with SequenceExamples
    file_queue = tf.train.string_input_producer([tfrecord_file])
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(file_queue)

    sequence_features = {
       **input_features,
       "labels":  tf.FixedLenSequenceFeature([], tf.int64)
    }
    # parse examples
    context_parsed, sequence_parsed = tf.parse_single_sequence_example(
      serialized=serialized_example,
      context_features=context_features,
      sequence_features=sequence_features
    )
    t_char = sequence_parsed[CHAR_INPUT_TENSOR_NAME]
    t_seg = sequence_parsed[SEG_INPUT_TENSOR_NAME]
    t_label = sequence_parsed["labels"]
    input_tensors = [t_char, t_seg, t_label]

    # batch and dynamic padding
    char_inputs, seg_inputs, labels = tf.train.batch(
        input_tensors,
        batch_size=batch_size,
        dynamic_pad=True,
        capacity=1000 + 3 * batch_size,
        allow_smaller_final_batch=True)


    return {
        CHAR_INPUT_TENSOR_NAME: char_inputs,
        SEG_INPUT_TENSOR_NAME: seg_inputs,
        }, labels

在估算器model_fn中，计算动态长度：

def model_fn(features, labels, mode, params):
    char_inputs = features[CHAR_INPUT_TENSOR_NAME]
    used = tf.sign(tf.abs(char_inputs))
    length = tf.reduce_sum(used, reduction_indices=1)
    lengths = tf.cast(length, tf.int32)

要导出和预测，请输入我的serve_input_receiver_fn：

def serving_input_receiver_fn():
    serialized_example = tf.placeholder(dtype=tf.string, shape=None,
                                       name='input_example_tensor')
    receiver_tensors = { RECIEVER_INPUT: serialized_example }
    _, features = tf.parse_single_sequence_example(
        serialized=serialized_example,
        sequence_features=input_features
    )

    return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)

问题：

当我使用serving_input_receiver_fn构建估算器并导出模型时，在预测阶段，输入特征（char_inputs，seg_inputs）似乎具有[1]的形状。

length = tf.reduce_sum(used, reduction_indices=1)

ValueError: Invalid reduction dimension 1 for input with 1 dimensions. for 'Sum' (op: 'Sum') with input shapes: [?], [] and with computed input tensors: input[1] = <1>.

Tensorflow与Sequence示例一起使用

0 个答案: