How to show more images than the batch size value?

user_15853851 0
Viewed 186

I have the following code:

train_ds = tf.keras.utils.image_dataset_from_directory(
  '/media/Tesi',
  validation_split=0.2,
  subset="training",
  seed=123,
  image_size=(360, 360),
  batch_size=18)

class_names = train_ds.class_names

val_ds = tf.keras.utils.image_dataset_from_directory(
  '/media/Tesi',
  validation_split=0.2,
  subset="validation",
  seed=123,
  image_size=(360, 360),
  batch_size=18)

num_classes = len(class_names)

Then I create a model and make some probabilities. When I show the images in val_ds, my code is:

plt.figure(figsize=(20, 20))
for images, _ in val_ds.take(1):
    for i in range(18):
        ax = plt.subplot(6, 6, i + 1)
        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[np.argmax(predictions[i])])
        plt.axis("off")

In this way I show always the first 18 images of val_ds. How can I show the images for example from index 18 to 36? thanks

tensorflow-datasets keras tensorflow matplotlib python
1 Answers

You can use tf.data.Dataset.skip and tf.data.Dataset.take:

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 = 18

val_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="validation",
  seed=123,
  image_size=(360, 360),
  batch_size=batch_size, shuffle=False)

for images, _ in val_ds.skip(1).take(1):
  for i in range(18):
    ax = plt.subplot(6, 6, i + 1)
    plt.imshow(images[i].numpy().astype("uint8"))
    plt.axis("off")

In this example, the first 18 images are skipped (1 batch) and afterwards you take the next 18 images (also 1 batch). You just need to make sure that shuffle=False, in order to make sure that you do not get the same images when calling take(1).

user_96578610
Related