#Keras Invalid reduction dimension 2 for input with 2 dimensions.

在跑Bert-CRF的时候发现公司的电脑带不动，就把Bert改成了BiLSTM，因为之前没有试过中文命名实体识别的代码，所以出了一个小失误…花了半天才发现（这里用了（双向）最大熵隐马尔可夫模型，作用和用法都类似CRF，但是比CRF更快更简单。）

原代码如下：

	input_id = Input(batch_shape=(None, None), )
    x = Embedding(21128, 8)(input_id)
    x = Bidirectional(LSTM(256))(x)
    x = Dropout(0.3)(x)
    x = Dense(1)(x)
    output = Dense(len(categories) * 2 + 1)(x)
    MME = MaximumEntropyMarkovModel(lr_multiplier=crf_lr_multiplier)
    crf = CRF(7, sparse_target=True)
    output = MME(output)

    model = Model(input_id, output)
    model.summary()
    model.compile(
        loss=MME.sparse_loss,
        optimizer=Adam(learning_rate),
        metrics=[MME.sparse_accuracy]
    )

这里用的是苏剑林老师的bert4keras（苏老师很厉害！），一直报错：

ValueError: Invalid reduction dimension 2 for input with 2 dimensions. 
for 'metrics/sparse_accuracy/All' (op: 'All') with input shapes: [?,7], [] and with computed input tensors: input[1] = <2>.

因为一直跑不痛就去看了别人的代码是怎么用keras构建的，结果model.summary之后才发现…CRF层要接收的参数是三维的，也就是说，LSTM层要return_sequences=True…

改了之后问题迎刃而解

	input_id = Input(batch_shape=(None, None), )
    x = Embedding(21128, 8)(input_id)
    x = Bidirectional(LSTM(256, return_sequences=False))(x)
    x = Dropout(0.3)(x)
    x = Dense(1)(x)
    output = Dense(len(categories) * 2 + 1)(x)
    MME = MaximumEntropyMarkovModel(lr_multiplier=crf_lr_multiplier)
    crf = CRF(7, sparse_target=True)
    output = MME(output)

    model = Model(input_id, output)
    model.summary()
    model.compile(
        loss=MME.sparse_loss,
        optimizer=Adam(learning_rate),
        metrics=[MME.sparse_accuracy]
    )

model.summary要这样的

Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, None)              0         
_________________________________________________________________
embedding_1 (Embedding)      (None, None, 8)           169024    
_________________________________________________________________
bidirectional_1 (Bidirection (None, None, 512)         542720    
_________________________________________________________________
dropout_1 (Dropout)          (None, None, 512)         0         
_________________________________________________________________
dense_1 (Dense)              (None, None, 1)           513       
_________________________________________________________________
dense_2 (Dense)              (None, None, 7)           14        
_________________________________________________________________
maximum_entropy_markov_model (None, None, 7)           49        
=================================================================
Total params: 712,320
Trainable params: 712,320
Non-trainable params: 0
_________________________________________________________________