获取内部Keras层的预测值。

Question

我有一个像这样的TensorFlow模型-

​

​

我希望知道特定输入的红色标记层(5个浮点值)的值，以检查模型在此层(注意层)如何响应。我需要这个值，这样我就可以知道我的注意力层是否正确地提取了值。

由于模型是一个端到端模型，我不知道如何提取内部层的值以供特定的输入。有人能帮忙吗？

Answer 1

您可以编写class Callback，然后传递输入并检查所需每一层的输出：

import tensorflow as tf
import numpy as np

class CustomCallback(tf.keras.callbacks.Callback):
   def __init__(self):
        self.data = np.random.rand(1,10)
   def on_epoch_end(self, epoch, logs=None):
        dns_layer = self.model.layers[6]
        outputs = dns_layer(self.data)
        tf.print(f'\n input: {self.data}')
        tf.print(f'\n output: {outputs}')


x_train = tf.random.normal((10, 32, 32))
y_train = tf.random.uniform((10, 1), maxval=10)

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.LSTM(256, input_shape=(x_train.shape[1], x_train.shape[2]), return_sequences=True))
model.add(tf.keras.layers.Dropout(0.2))
model.add(tf.keras.layers.LSTM(256))
model.add(tf.keras.layers.Dropout(0.2))
model.add(tf.keras.layers.Dense(10, activation='softmax')) 
model.add(tf.keras.layers.Dropout(0.2))
model.add(tf.keras.layers.Dense(5, activation='softmax'))
model.add(tf.keras.layers.Dropout(0.2))
model.add(tf.keras.layers.Dense(10, activation='softmax')) 
model.compile(optimizer='adam', loss = tf.keras.losses.SparseCategoricalCrossentropy(False))
model.summary()    

for layer in model.layers:
    print(layer)

model.fit(x_train, y_train , epochs=3, callbacks=[CustomCallback()], batch_size=32)

输出：

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 lstm (LSTM)                 (None, 32, 256)           295936    
                                                                 
 dropout (Dropout)           (None, 32, 256)           0         
                                                                 
 lstm_1 (LSTM)               (None, 256)               525312    
                                                                 
 dropout_1 (Dropout)         (None, 256)               0         
                                                                 
 dense (Dense)               (None, 10)                2570      
                                                                 
 dropout_2 (Dropout)         (None, 10)                0         
                                                                 
 dense_1 (Dense)             (None, 5)                 55        
                                                                 
 dropout_3 (Dropout)         (None, 5)                 0         
                                                                 
 dense_2 (Dense)             (None, 10)                60        
                                                                 
=================================================================
Total params: 823,933
Trainable params: 823,933
Non-trainable params: 0
_________________________________________________________________
<keras.layers.recurrent_v2.LSTM object at 0x7f6e2163dbd0>
<keras.layers.core.dropout.Dropout object at 0x7f6da1d2efd0>
<keras.layers.recurrent_v2.LSTM object at 0x7f6d9dfe0a50>
<keras.layers.core.dropout.Dropout object at 0x7f6d9de1ec90>
<keras.layers.core.dense.Dense object at 0x7f6d9de04dd0>
<keras.layers.core.dropout.Dropout object at 0x7f6d9dd549d0>
<keras.layers.core.dense.Dense object at 0x7f6d9dd8ec90>
<keras.layers.core.dropout.Dropout object at 0x7f6d9dedd650>
<keras.layers.core.dense.Dense object at 0x7f6d9ddc2ed0>
Epoch 1/3
1/1 [==============================] - ETA: 0s - loss: 2.4188
 input: [[0.91498145 0.98430978 0.22720893 0.76032816 0.78405846 0.72664182
  0.7772921  0.9851892  0.41715033 0.21014543]]

 output: [[0.5767021  0.04140956 0.1909151  0.06737834 0.12359484]]
1/1 [==============================] - 12s 12s/step - loss: 2.4188
Epoch 2/3
1/1 [==============================] - ETA: 0s - loss: 2.4111
 input: [[0.91498145 0.98430978 0.22720893 0.76032816 0.78405846 0.72664182
  0.7772921  0.9851892  0.41715033 0.21014543]]

 output: [[0.5780218  0.04101932 0.18909878 0.06769065 0.12416941]]
1/1 [==============================] - 0s 376ms/step - loss: 2.4111
Epoch 3/3
1/1 [==============================] - ETA: 0s - loss: 2.3978
 input: [[0.91498145 0.98430978 0.22720893 0.76032816 0.78405846 0.72664182
  0.7772921  0.9851892  0.41715033 0.21014543]]

 output: [[0.579072   0.04067017 0.1874026  0.0679936  0.12486164]]
1/1 [==============================] - 0s 458ms/step - loss: 2.3978