keras：layer.shape与model.summary（）中的图层输出形状不同

Question

在keras中：我发现layer.shape与model.summary（）中的layer输出形状不同。可能是什么原因？

我正在尝试嵌入[https://arxiv.org/pdf/1703.03130.pdf]

的结构化自专心句子

A.shape
Out[44]: TensorShape([Dimension(None), Dimension(10), Dimension(339)])

vs。 seq_self_attention_4形状（无，339、256）如下：

    Model: "model_3"
    __________________________________________________________________________________________________
    Layer (type)                    Output Shape         Param #     Connected to                     
    ==================================================================================================
    lstm_input (InputLayer)         (None, 339)          0                                            
    __________________________________________________________________________________________________
    lstm/embedding (Embedding)      (None, 339, 100)     5000000     lstm_input[0][0]                 
    __________________________________________________________________________________________________
    bidirectional_4 (Bidirectional) (None, 339, 256)     234496      lstm/embedding[0][0]             
    __________________________________________________________________________________________________
    seq_self_attention_4 (SeqSelfAt (None, 339, 256)     1330        bidirectional_4[0][0]            
    __________________________________________________________________________________________________
    layer_mat_mul_4 (LayerMatMul)   [(None, 339, 256), ( 0           seq_self_attention_4[0][0]       
                                                                     bidirectional_4[0][0]            
    __________________________________________________________________________________________________
    flatten_4 (Flatten)             (None, 86784)        0           layer_mat_mul_4[0][0]            
    __________________________________________________________________________________________________
    lstm_predictions (Dense)        (None, 2)            173570      flatten_4[0][0]                  

class SeqSelfAttention(keras.layers.Layer):
    def __init__(self, da, r, hidden_state):
        super(SeqSelfAttention, self).__init__()
        self.da = da
        self.r = r
        self.hidden_state = hidden_state
        w_init = tf.random_normal_initializer()
        self.W_s1 = tf.Variable(initial_value=w_init(shape=(self.da, 2*self.hidden_state),
                                              dtype='float32'),
                                trainable=True)
        self.W_s2 = tf.Variable(initial_value=w_init(shape=(self.r, self.da),
                                              dtype='float32'),
                                trainable=True)
    def call(self, inputs):
        W_s1_H  = K.tanh(tf.matmul(self.W_s1, inputs, transpose_b=True))
        W_s2_H  = tf.matmul(self.W_s2, W_s1_H)
        return K.softmax(W_s2_H, axis=1)
da=5
r=10
vocab_size=50000
embedding_size=100
embedding_weights=None
hidden_state=128
input_shape=339
    message = Input(shape=(input_shape,), dtype='int32', name='lstm_input')
    mask = Lambda(lambda inputs: K.not_equal(inputs, 0))(message)
    embedding = make_embedding('lstm', vocab_size, embedding_size, embedding_weights, mask_zero=False)(message)

        return_sequences = True
        lstm_1 = Bidirectional(LSTM(units=hidden_state, return_sequences=return_sequences))(embedding, mask=mask)
        A = SeqSelfAttention(da, r, hidden_state)(lstm_1)
        final = Flatten()(m)

    preds = Dense(units=2, activation='softmax', name='lstm_predictions')(final)

    model = Model(
        inputs=[message],
        outputs=[preds],
    )
    model.compile(loss='sparse_categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

谢谢。