Autoencoder Python Keras - 错误图已断开答案

【问题标题】：Autoencoder Python Keras - Error graph disconnectedAutoencoder Python Keras - 错误图已断开
【发布时间】：2021-06-17 17:47:24
【问题描述】：

我尝试使用 Keras 库在 Python 中创建自动编码器。我想将我的模型分成两部分：编码器和解码器。但是当我尝试创建解码器时出现错误。

我的代码是：

autoencoder = tf.keras.models.Sequential()
encoding_dim = 250  
nb_classe = 4

# Inputs
input_imgA = keras.Input(shape=(784,))
input_imgB = keras.Input(shape=(784,))
input_imgC = keras.Input(shape=(784,))

# Encoder part
encodedA = keras.layers.Dense(encoding_dim, activation='relu')(input_imgA)
encodedB = keras.layers.Dense(encoding_dim, activation='relu')(input_imgB)

encoded = keras.layers.Multiply()([encodedA, encodedB])
encoded = keras.layers.Dense(nb_classe, activation='softmax')(encoded)

# Decoder part
decodedA = keras.layers.Dense(encoding_dim, activation='relu')(encoded)
decodedB = keras.layers.Dense(encoding_dim, activation='relu')(input_imgC)

decoded = keras.layers.Multiply()([decodedA, decodedB])
decoded = keras.layers.Dense(784, activation='relu')(decoded)

# Autoencoder
autoencoder = keras.Model(inputs = [input_imgA, input_imgB, input_imgC], outputs = decoded)
encoder = keras.Model(inputs = [input_imgA, input_imgB], outputs = encoded)

encoded_input = keras.Input(shape=(nb_classe,))
decoder = keras.Model(inputs = [encoded_input, input_imgC], outputs = decoded)

当我定义自动编码器和编码器时，没有问题。但我无法定义解码器。我希望解码器将编码器输出和图像作为输入。但我有一个错误，例如：

Graph disconnected: cannot obtain value for tensor Tensor("input_8:0", shape=(None, 4), dtype=float32) at layer "dense_6". The following previous layers were accessed without issue: []

我不明白我的错误。请问你能帮帮我吗？

谢谢。

【问题讨论】：

标签： python tensorflow keras-layer autoencoder

【解决方案1】：

如果您需要访问模型中的中间层，通常最好通过子类化 tf.keras.Model 并返回您需要的任何内容来构建您自己的单个模型。

示例：

相关包

import tensorflow as tf

from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import Multiply

构建模型

class Autoencoder(tf.keras.Model):
    """``Autoencoder``."""
    def __init__(self, orig_units=784, num_units=250, num_classes=4):
        super().__init__()
        self._orig_units = orig_units
        self._num_units = num_units
        self._num_classes = num_classes
        # encoder
        self.enc_a = Dense(self._num_units, activation="relu")
        self.enc_b = Dense(self._num_units, activation="relu")
        self.probs = Dense(self._num_classes, activation="softmax")
        self.enc = Multiply()
        # decoder
        self.dec_a = Dense(self._num_units, activation="relu")
        self.dec_b = Dense(self._num_units, activation="relu")
        self.logits = Dense(self._orig_units, activation="relu")
        self.dec = Multiply()

    def call(self, xa, xb, xc):
        # encoder
        exa = self.enc_a(xa)
        exb = self.enc_b(xb)
        encoded = self.enc([exa, exb])
        encoded = self.probs(encoded)
        # decoder
        dxa = self.dec_a(encoded)
        dxb = self.dec_b(xc)
        decoded = self.dec([dxa, dxb])
        decoded = self.logits(decoded)
        return encoded, decoded

实例化并测试所需的输出

# model
ae = Autoencoder()

# data
Xa = tf.random.normal([5, 784])
Xb = tf.random.normal([5, 784])
Xc = tf.random.normal([5, 784])

# correct output?
enc_out, dec_out = ae(Xa, Xb, Xc)

print(enc_out)
# tf.Tensor(
#    [[0.15498394 0.00756733 0.5070575  0.33039123]
#     [0.12203674 0.05137487 0.2758123  0.5507761 ]
#     [0.11994947 0.5594128  0.08857376 0.23206405]
#     [0.6519191  0.06745841 0.2350422  0.04558026]
#     [0.16059074 0.11788747 0.46389887 0.257623  ]],
# shape=(5, 4), dtype=float32)

print(dec_out)
# tf.Tensor(
#    [[0.00234556 0.         0.00978717 ... 0.         0.         0.00275952]
#    [0.03659576 0.00313164 0.00698952 ... 0.         0.01964222 0.        ]
#    [0.00446499 0.01008888 0.         ... 0.00357639 0.00307285 0.00277464]
#    [0.00713126 0.00656428 0.03171315 ... 0.0217022  0.         0.        ]
#    [0.         0.01419089 0.         ... 0.         0.         0.00467858]],
# shape=(5, 784), dtype=float32)

【讨论】：