在TensorFlow 1.15中使用BiLSTM-CRF实现CRF层

时间:2020-06-11 22:59:18

标签: tensorflow keras deep-learning lstm crf

我使用keraskeras_contrib(带有后者的CRF来实现,它带有条件随机场层(BiLSTM-CRF))实现了双向长短期记忆神经网络。原生keras functionality。该任务被命名为“实体识别”,分为6类之一。网络的输入是300维预训练的GloVe词嵌入的序列。这是我的模型摘要:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 648)               0   
_________________________________________________________________
embedding_1 (Embedding)      (None, 648, 300)          1500000   
_________________________________________________________________
bidirectional_1 (Bidirection (None, 648, 10000)        3204000   
_________________________________________________________________
crf_1 (CRF)                  (None, 648, 6)            6054      
=================================================================

现在,我想在TensorFlow 1.15中实现相同的模型。由于keras_contrib CRF模块仅适用于keras,而不适用于TensorFlow,因此我使用了this回购中为TensorFlow 1.X构建的CRF实现。该仓库包含两个不错的CRF here示例实现,但是在对我的数据进行训练时,每个都会产生不同的错误。

实施1

from tensorflow.keras.layers import Bidirectional, Embedding, LSTM, TimeDistributed
from tensorflow.keras.models import Sequential

from tf_crf_layer.layer import CRF
from tf_crf_layer.loss import crf_loss
from tf_crf_layer.metrics import crf_accuracy

MAX_WORDS = 50000
EMBEDDING_LENGTH = 300
MAX_SEQUENCE_LENGTH = 648
HIDDEN_SIZE = 512

model = Sequential()
model.add(Embedding(MAX_WORDS, EMBEDDING_LENGTH, input_length=MAX_SEQUENCE_LENGTH, mask_zero=True, weights=[embedding_matrix], trainable=False))
model.add(Bidirectional(LSTM(HIDDEN_SIZE, return_sequences=True)))
model.add(CRF(len(labels)))

model.compile('adam', loss=crf_loss, metrics=[crf_accuracy])

这是我尝试编译模型时遇到的错误:

File "/.../tf_crf_layer/metrics/crf_accuracy.py", line 48, in crf_accuracy
    crf, idx = y_pred._keras_history[:2]

AttributeError: 'Tensor' object has no attribute '_keras_history'

根据上述存储库计算crf_accuracy时会出现错误。

def crf_accuracy(y_true, y_pred):
    """
    Get default accuracy based on CRF `test_mode`.
    """
    import pdb; pdb.set_trace()
    crf, idx = y_pred._keras_history[:2]
    if crf.test_mode == 'viterbi':
        return crf_viterbi_accuracy(y_true, y_pred)
    else:
        return crf_marginal_accuracy(y_true, y_pred)

根据this线程,当张量对象不是keras层的输出时,显然会发生这种错误。为什么这里会出现此错误?

实施2

from tf_crf_layer.layer import CRF
from tf_crf_layer.loss import crf_loss, ConditionalRandomFieldLoss
from tf_crf_layer.metrics import crf_accuracy
from tf_crf_layer.metrics.sequence_span_accuracy import SequenceSpanAccuracy

model = Sequential()
model.add(Embedding(MAX_WORDS, EMBEDDING_LENGTH, input_length=MAX_SEQUENCE_LENGTH, mask_zero=True, weights=[embedding_matrix], trainable=False))
model.add(Bidirectional(LSTM(HIDDEN_SIZE, return_sequences=True)))
model.add(CRF(len(labels), name="crf_layer"))

model.summary()

crf_loss_instance = ConditionalRandomFieldLoss()  
model.compile(loss={"crf_layer": crf_loss_instance}, optimizer='adam', metrics=[SequenceSpanAccuracy()])

在这里编译模型,但是一旦训练的第一个纪元开始,这个错误就会浮出水面:

InvalidArgumentError: Expected begin and size arguments to be 1-D tensors of size 3, but got shapes [2] and [2] instead.
     [[{{node loss_4/crf_layer_loss/Slice_1}}]]

我正在使用小批量训练模型,这可以解释错误吗?我还注意到,尽管CRF层的参数数量与上面相同,但是我的CRF层的模型摘要没有一个尺寸(比较上面的摘要和下面的摘要中的CRF层规范)。为什么导致这种不匹配,如何解决?

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_5 (Embedding)      (None, 648, 300)          1500000   
_________________________________________________________________
bidirectional_5 (Bidirection (None, 648, 1000)         3204000   
_________________________________________________________________
crf_layer (CRF)              (None, 648)               6054      
=================================================================

0 个答案:

没有答案