CODEWITHSUNDEEP
pythontensorflowtensorflow-datasets
Rafael Sanjuan
Rafael Sanjuan

Reputation: 11

Tensorflow load image dataset with image labels

I'm trying to find an easy way to get my data into a TensorFlow dataset without having to load it before and process it as a NumPy array.

In this case, I'm working on a segmentation model and my data is structured so one directory has the training data images and a different directory has the 'masks', essentially also images.

With tf.keras.preprocessing.image_dataset_from_directory, I can load an image dataset and it will either set the labels as the name of the directory or let me set them myself through a function argument but it won't let me set an argument so it takes labels from a different directory. I'm reading through the documentation but I don't see an easy way to load these types of datasets where the labels are images on their own.

Upvotes: 1

Views: 2205

Answers (1)

Aventau
Aventau

Reputation: 190

This is one possible way to do it in Keras:

train_generator_images = image_data_generator_train.flow_from_dataframe(
    dataframe=train,
    directory='..//VOCdevkit/VOC2009/JPEGImages',
    x_col='filename',
    class_mode=None,
    color_mode="rgb",
    target_size=(image_size[1],image_size[0]),
    batch_size=batchSize,
    seed=seed)

train_generator_mask = mask_data_generator_train.flow_from_dataframe(
    dataframe=train,
    directory='..//VOCdevkit/VOC2009/SegmentationClass',
    x_col='segmentation',
    class_mode=None,
    color_mode="grayscale",
    target_size=(image_size[1],image_size[0]),
    batch_size=batchSize,
    seed=seed)

train_generator = zip(train_generator_images, train_generator_mask)

Important here is to set the same seed so the images and labels match. I copied it from an older project of mine so it might not be up to date.

Upvotes: 1

Related Questions