批量导入图片，解决非分类问题

Question

I'm trying to do a superresolution network, but I am having trouble importing my own data. I have two types of images: resized images (smaller), original images. The first one is going to be used as an input of the network and the second ones will be used for training the network.

The problem is that I need to load my images in batches because my computer doesn't have enough GPU memory for constructing the whole dataset at once. I thought that using the following code could work:

train_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="training",
  seed=123,
  image_size=(img_height, img_width),
  batch_size=batch_size)

the problem is that I only know how to make it work for classification problems because, as far as I am concerned, it's designed for only having training and validation dataset.

For doing the superresolution I need four datasets:

normal-size-train

small-size-train

normal-size-test

small-size-test

NOTE: My program works when I create a tensor for resized images and another one for original images, but now I want to implement a larger dataset.

Answer 1

I think it's better implement a data generator for this task. here an example. you can/must add image reshaping if the images in the dateset does not have the same shape.

def image_generator(path, batch_size=16):
    list_path = glob.glob(path)
    index = 0
    list_of_low_dim_images = []
    list_of_high_dim_images = []
    size = len(list_path)
    while True:
        index +=1
        for path in list_path:
            path2 = path.replace("small", "normal")
            small_img = tf.io.read_file(path)
            small_img = decode_img(small_img)
            normal_img = tf.io.read_file(path2)
            normal_img = decode_img(normal_img)
            list_of_low_dim_images.append(small_img)
            list_of_high_dim_images.append(normal_img)
            if index == batch_size:
                inedx = 0
                yield list_of_low_dim_images,list_of_high_dim_images