Question

我正在研究LSTM递归神经网络，想区分我拥有的最后一个致密层的输出/预测。

考虑一个简单的网络-

data_in = Input(shape=(10, 4)) # 10 time steps and 4 features
x = LSTM(4, activation='relu')(data_in)
x = Dense(32)(x)
out = Dense(1)(x)

我想实现的目标有些棘手，用简单的代码来实现，就像这样-

if batch_no == 1:
    # for the first batch, since no history is available, assume to start wiht 0 --> prepend=0
    final_out = np.diff(out, 1, prepend=0)
    # hold on to the final prediction in batch to differntiate the output of next batch
    carry_forward = out[-1]          
else:
    # use the final prediction from previous batch to differentiate the output of the current batch
    final_out = np.diff(out, 1, prepend=carry_forward)
    # again hold on to the final prediction to use for next batch
    carry_forward = out[-1]

到目前为止，我已经尝试过了-

Lambda层看起来很有希望，但是我不确定如何在此处使用“ carry_forward”。我可以做这样的事情，但是有一个折衷方案，就是将一个特定的预测（每个批次的第一个）与无关的值区分开来

final_out = Lambda(lambda m: m - tf.roll(m, 1, axis=0))(out)

与我已经写过的if-else非常相似，但是具有张量运算-

if batch_no == 1:
    subtract = tf.identity(out)
    subtract = tf.roll(subtract, 1, axis=0)
    subtract[0] = 0                           # doesn't work, cause of eager tensor not being enabled
    final_out = out - subtract
    carry_forward = out[-1].numpy()        
else:
    subtract = tf.identity(out)
    subtract = tf.roll(subtract, 1, axis=0)
    subtract[0] = carry_forward               # doesn't work, cause of eager tensor not being enabled
    final_out = out - subtract
    carry_forward = out[-1]

注意

我无法在此处使用tf.experimental.numpy.diff，因为我正在tensorflow 2.1.0上运行

这使我想起了问题-

启用渴望的张量有副作用吗？
如何实现？

任何指针都将非常有帮助

Tensorflow Lambda层可区分批次之间的预测

0 个答案: