Keras ImageDatagenerator Load Images from one Subfolder

Question

I want to code an image anomaly detection with autoencoder.

Therefore I splitted my images into one subfolder for trainining images, one for validation images and one for anomaly images.

history = model.fit(
    train_generator,
    steps_per_epoch=500 // batch_size,
    epochs=1000,
    validation_data=validation_generator,
    validation_steps=75 // batch_size,
    shuffle=True)

For model.fit I need an seperated train_generator and validation_generator

For getting train_generator I am using this one.

train_generator = datagen.flow_from_directory(
    'C:/Python/Anomaly_dection_V1/cell_images/Uninfected_Train/',
    target_size=(SIZE, SIZE),
    color_mode='rgb',
    batch_size=batch_size,
    class_mode='input',
)

So I take the images directly from the subfolder, where just images and no other folders are located.

Sadly I get the following message: Found 0 images belonging to 0 classes.

Can anyone help me out?

If I take one folder above instead of the subfolder, it can find images from three classes. But for the model.fit I need just the train data and the validatio data. I didnt find any solution just to use the subfolders/classes for that.

Is flow_from_directory not working to get images from just one subfolder?

Answer 1

You can access the specific single folder images using image_dataset_from_directory by specifying the labels = None as below:

This way you can also split the dataset of that folder into train and validation portion using validation_split=0.2.

data_dir  = "/content/drive/MyDrive/data/PetImages/Cat"

# in your case: data_dir = "C:/Python/Anomaly_dection_V1/cell_images/Uninfected_Train"

Training data :

train_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="training",
  seed=123,
  labels = None,
  image_size=(img_height, img_width),
  batch_size=batch_size)

Output:

Found 12500 files belonging to 1 classes.
Using 10000 files for training.

Validation dataset:

val_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="validation",
  seed=123,
  labels = None,
  label_mode=None,
  image_size=(img_height, img_width),
  batch_size=batch_size)

Output:

Found 12500 files belonging to 1 classes.
Using 2500 files for validation.