A minimal working example for Tensorflow issue #32239

tf-github-32239.py

from tensorflow.keras import models
from tensorflow.keras import layers
from tensorflow.keras import optimizers

from tensorflow.keras.preprocessing import image
from tensorflow.keras.preprocessing.image import ImageDataGenerator

from tensorflow.keras.applications import VGG16
from tensorflow.keras.applications import vgg16

from tensorflow.keras.callbacks import ModelCheckpoint

import os


# Create Keras model
image_size = 150
input_layer = layers.Input(shape=(image_size, image_size, 3), name="model_input")
base_model = VGG16(weights="imagenet", include_top=False, input_tensor=input_layer)
model_head = base_model.output
model_head = layers.Flatten(name="model_head_flatten")(model_head)
model_head = layers.Dense(256, activation="relu")(model_head)
model_head = layers.Dense(2, activation="softmax")(model_head)
model = models.Model(inputs=input_layer, outputs=model_head)

# Create image date generators
# You need one image data folder with three sub-folders "train", "validation", "test"
image_dir = "/home/mfb/Development/tf-github/data"
datagen = ImageDataGenerator(preprocessing_function=vgg16.preprocess_input)
training_img_generator = datagen.flow_from_directory(os.path.join(image_dir, 'train'),
                         target_size=(image_size, image_size), batch_size=20, class_mode="categorical")
validation_img_generator = datagen.flow_from_directory(os.path.join(image_dir, 'validation'),
                           target_size=(image_size, image_size), batch_size=20, class_mode="categorical")
test_img_generator = datagen.flow_from_directory(os.path.join(image_dir, 'test'),
                     target_size=(image_size, image_size), batch_size=20, class_mode="categorical")

# Compile Keras model
model.compile(loss="categorical_crossentropy", optimizer=optimizers.Adam(), metrics=["accuracy"])

# Train Keras model
auto_save_path = "/home/mfb/Development/tf-github/models"
checkpoint = ModelCheckpoint(auto_save_path, monitor="val_acc", verbose=0, save_best_only=True)
model.fit_generator(training_img_generator,
                    steps_per_epoch=50, epochs=25, validation_steps=50,
                    validation_data=validation_img_generator,
                    callbacks=[checkpoint], verbose=1)