将TensorFlow Keras模型作为H5文件时输入错误

Question

我正在尝试制作一个自动编码器，以便获得图像的矢量格式（x：640，y：480）。但是，当我尝试调用 encoder.save（“ encoder.h5”）时，我会收到以下错误：

TypeError: ('Not JSON Serializable:', <tf.Variable 'batch_normalization/gamma:0' shape=(32,) dtype=float32, numpy=
array([1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
      dtype=float32)>)

我很确定该模型在拟合函数工作时可以正常工作，并且我能够打电话 encoder.predict（）完成拟合并拿出编码的向量（长度为60） 我的代码是：

import os
import tensorflow as tf
from tensorflow.python.keras.layers import Input, UpSampling2D, Add, Conv2D, MaxPooling2D, LeakyReLU
import cv2
from tensorflow.python.keras import Model
from tensorflow.python.keras import layers, losses
from tensorflow.python.keras.models import Model
import h5py

def get_encoder(shape=(640, 480, 3)):
    def res_block(x, n_features):
        _x = x
        x = tf.keras.layers.BatchNormalization()(x)
        x = LeakyReLU()(x)

        x = Conv2D(n_features, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
        x = Add()([_x, x])
        return x

    inp = Input(shape=shape)

    # 640 x 480
    x = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), padding='same')(inp)
    x = tf.keras.layers.BatchNormalization()(x)
    x = LeakyReLU()(x)

    x = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    x = tf.keras.layers.BatchNormalization()(x)
    x = LeakyReLU()(x)

    x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)

    # 320 x 240
    x = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    for _ in range(2):
        x = res_block(x, 32)
    x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)

    # 160 x 120
    x = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    for _ in range(2):
        x = res_block(x, 32)
    x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)

    # 80 x 60
    x = Conv2D(1, kernel_size=(1, 1), strides=(1, 1), padding='same')(x)
    return Model(inp, x)


def get_decoder(shape=(240, 180, 3)):
    inp = Input(shape=shape)

    # 60 x 80
    x = UpSampling2D((2, 2))(inp)
    x = Conv2D(64, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    x = tf.keras.layers.BatchNormalization()(x)
    x = LeakyReLU()(x)

    # 120 x 160
    x = UpSampling2D((2, 2))(x)
    x = Conv2D(64, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    x = tf.keras.layers.BatchNormalization()(x)
    x = LeakyReLU()(x)

    # 240 x 320
    x = UpSampling2D((2, 2))(x)
    x = Conv2D(8, kernel_size=(3, 3), strides=(1, 1), padding='same')(x)
    x = tf.keras.layers.BatchNormalization()(x)
    x = LeakyReLU()(x)

    # 480 x 640
    x = Conv2D(1, kernel_size=(1, 1), strides=(1, 1), padding='same')(x)
    return Model(inp, x)


encoder = get_encoder((480, 640, 3))
decoder = get_decoder((60, 80, 1))
inp = Input((480, 640, 3))
e = encoder(inp)
d = decoder(e)
autoencoder = Model(inp, d)

autoencoder.compile(optimizer='adam', loss=losses.MeanSquaredError())

batch_size = 8
SHAPE = (480, 640)
IMAGES = "pathToImages"
image_generator = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1.0 / 255)
train_gen = image_generator.flow_from_directory(
    os.path.join(IMAGES, "train"),
    class_mode="input", target_size=SHAPE, batch_size=batch_size,
)
val_gen = image_generator.flow_from_directory(
    os.path.join(IMAGES, "test"),
    class_mode="input", target_size=SHAPE, batch_size=batch_size,
)

for i in range(10):
    autoencoder.fit(train_gen, validation_data=val_gen, epochs=1, steps_per_epoch=10, validation_steps= 5, batch_size=batch_size)
    encoder.save('encoder.h5')
    decoder.save('decoder.h5')

我基于此

关于我如何修复它的任何想法，还是节省模型权重的替代方法？

编辑：使用 encoder.save_weights（'encoder.h5'） works

Answer 1

OP对萨德拉的回答说了这一点：

使用 encoder.save_weights（'encoder.h5'） works