如何打印凯拉斯张量值？

问题描述

我想打印LSTM层的状态值。

class CustomCallback(keras.callbacks.Callback):
   def on_epoch_end(self, epoch, logs=None):
        encoder_outputs, state_h, state_c = self.model.layers[1].output
        print(state_h)
        print(state_c)

打印的内容如下

32/32 - 0s - loss: 39.6719 - accuracy: 0.2420
KerasTensor(type_spec=TensorSpec(shape=(None, 5), dtype=tf.float32, name=None), name='lstm_5/PartitionedCall:3', description="created by layer 'lstm_5'")
Epoch 5/20000
32/32 - 0s - loss: 39.6549 - accuracy: 0.2420
KerasTensor(type_spec=TensorSpec(shape=(None, 5), dtype=tf.float32, name=None), name='lstm_5/PartitionedCall:3', description="created by layer 'lstm_5'")

如何打印张量的实值？

---

解决方案

您必须向Callback函数内的LSTM层提供一些数据才能看到一些实际值：

import tensorflow as tf

class CustomCallback(tf.keras.callbacks.Callback):
   def __init__(self, data, sample_size):
        self.data = data
        self.sample_size = sample_size
   def on_epoch_end(self, epoch, logs=None):
        encoder_outputs, state_h, state_c = lstm_layer(self.data[:self.sample_size])
        tf.print('state_h --> ', state_h)
        tf.print('state_c --> ', state_c)


inputs = tf.keras.layers.Input((5, 10))
x, _, _ = tf.keras.layers.LSTM(32, return_state=True, return_sequences=True)(inputs)
x = tf.keras.layers.Flatten()(x)
x = tf.keras.layers.Dense(units=1)(x)
model = tf.keras.Model(inputs, x)
model.compile(loss=tf.keras.losses.BinaryCrossentropy())

lstm_layer = model.layers[1]
x_train = tf.random.normal((10, 5, 10))
x_test = tf.random.normal((10, 5, 10))
model.fit(x_train, tf.random.uniform((10, 1), maxval=2), epochs=2, callbacks=[CustomCallback(x_test, 1)], batch_size=2)

Epoch 1/2
5/5 [==============================] - 2s 5ms/step - loss: 12.2492
state_h -->  [[-0.157256633 -0.0619691685 0.102620631 ... 0.0852451548 -0.0657120794 -0.201934695]]
state_c -->  [[-0.316935241 -0.157902092 0.184583426 ... 0.196862131 -0.134880155 -0.467693359]]
Epoch 2/2
5/5 [==============================] - 0s 4ms/step - loss: 11.8095
state_h -->  [[-0.15817374 -0.0611076616 0.103141323 ... 0.0845508352 -0.0648964494 -0.201082334]]
state_c -->  [[-0.319411457 -0.156104326 0.186640084 ... 0.194445729 -0.13365829 -0.464410305]]
<keras.callbacks.History at 0x7f9baadc8b90>

注意，我创建了一个x_test张量，但您也可以将x_train反馈给您的回调。lstm_layer根据您的训练进度保存当前权重。您可以通过在Callback函数tf.print(lstm_layer.get_weights())中打印层权重来验证这一点。