tf.contrib.rnn.BasicLSTMCell()中间程序解释

data = tf.reshape(x,[-1, 16, 16])
    #lstmUnits = 64
    lstmCell = tf.contrib.rnn.BasicLSTMCell(lstmUnits)
    lstmCell = tf.contrib.rnn.DropoutWrapper(cell=lstmCell, output_keep_prob=keep_prob)
    value, _ = tf.nn.dynamic_rnn(lstmCell, data, dtype=tf.float32) #value.shape: [batch_size，step，lstmUnits] 
    
    #weight = tf.Variable(tf.truncated_normal([lstmUnits, K]))
    weight = tf.Variable(tf.truncated_normal([lstmUnits, n_output])/np.sqrt(n_output))
    bias = tf.Variable(tf.constant(0.1, shape=[n_output]))
    value = tf.transpose(value, [1, 0, 2])#对value进行value = tf.transpose(value, [1, 0, 2])操作后得到的value.shapeshape为[step,batch_size,lstmUnits]
    
    #取最终的结果值
    last = tf.gather(value, int(value.get_shape()[0]) - 1) #其中value.get_shape()[0]) - 1找到value经过transpose后的最后一个分片
   
    pred = tf.nn.relu6(tf.matmul(last, weight) + bias) / 6
    y_label = tf.argmax(pred, axis =1)
    corr = tf.equal(tf.argmax(pred,1), tf.argmax(y,1)) 
    accu = tf.reduce_mean(tf.cast(corr, tf.float32))