How to attach or get filenames from MapDataset from image_dataset_from

I am training convolutional autoencoder and I have this code for loading data (images):

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    'path/to/images',
    image_size=image_size
)
normalization_layer = layers.experimental.preprocessing.Rescaling(1./255)

def adjust_inputs(images, labels):
    return normalization_layer(images), normalization_layer(images)

normalized_train_ds = train_ds.map(adjust_inputs)

As I don't need class labels but images itself as Y, I am mapping function adjust_inputs to dataset. But now when I try to access attribute filenames, I get error: AttributeError: 'MapDataset' object has no attribute 'filenames'. That is logical, because MapDataset is not Dataset.

How would I attach or get filenames of loaded images that are in my Dataset?

I am really surprised that there is not an easier interface for this, this looks like pretty common thing.

CodePudding user response：

I did it the following way.

After training of my model, I just reloaded all images, this time with option shuffle=False and ran them through my model to extract features. Since shuffle is off, ordering of images and file paths is same. So image at index 0, with corresponding features at index 0 has its file path at index 0.

CodePudding user response：

Just in case you want to add the filepaths as part of your dataset:

import tensorflow as tf
import pathlib
import matplotlib.pyplot as plt

dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file('flower_photos', origin=dataset_url, untar=True)
data_dir = pathlib.Path(data_dir)

batch_size = 32
train_ds = tf.keras.utils.image_dataset_from_directory(data_dir, shuffle=False, batch_size=batch_size)

normalization_layer = tf.keras.layers.Rescaling(1./255)
def change_inputs(images, labels, paths):
  x = normalization_layer(images)
  return x, x, tf.constant(paths)

normalized_ds = train_ds.map(lambda images, labels: change_inputs(images, labels, paths=train_ds.file_paths))

image, image, path = next(iter(normalized_ds.take(1)))

image = images[0]
path = paths[0]
print(path)
plt.imshow(image.numpy())

Found 3670 files belonging to 5 classes.
tf.Tensor(b'/root/.keras/datasets/flower_photos/daisy/100080576_f52e8ee070_n.jpg', shape=(), dtype=string)
<matplotlib.image.AxesImage at 0x7f9b113d1a10>

You will have to just make sure to you use the same batch size for the paths.