Why does TensorFlow convert my decoded image to float32 instead of uint8/16?

Question

I am using the following code to import a bunch of .png images and decode them using TensorFlow:

from __future__ import absolute_import, division, print_function
import tensorflow as tf
import numpy as np
import os

tf.enable_eager_execution()

NUM_TRAINING_SAMPLES = 333
NUM_CLASSES = 3
BATCH_SIZE = 5
NUM_EPOCHS = 6
INPUT_SIZE = (256, 256, 3)

random_indices = np.random.choice(range(13000), NUM_TRAINING_SAMPLES)
directory = "/home/local/CYCLOMEDIA001/ebos/Downloads/SYNTHIA_RAND_CVPR16"
directory_images = "/home/Downloads/SYNTHIA_RAND_CVPR16/RGB"
directory_labels = "/home/Downloads/SYNTHIA_RAND_CVPR16/GT"
train_images = np.array(os.listdir(directory_images))
train_labels = np.array(os.listdir(directory_images))
train_images = train_images[random_indices]
train_labels = train_labels[random_indices]
train_images = [tf.read_file(os.path.join(directory_images, img)) for img in train_images]
train_labels = [tf.read_file(os.path.join(directory_labels, img)) for img in train_labels]
train_images = [tf.io.decode_image(img, channels=3) for img in train_images]
train_labels = [tf.io.decode_image(img, channels=3) for img in train_labels]
train_images = tf.image.resize_images(train_images, INPUT_SIZE[:2])
train_labels = tf.image.resize_images(train_labels, INPUT_SIZE[:2])

train_dataset = tf.data.Dataset.from_tensor_slices((train_images, train_labels))
train_dataset = train_dataset.batch(3)
print(train_dataset.output_types)

This returns:

(tf.float32, tf.float32)

However, according to the documentation it should return a tensor of uint8's or uint16's. Why and where does the conversion take place?

I checked all intermediate steps with print statements, which doesn't tell me much as most intermediate lists are of class 'tensorflow.python.framework.ops.EagerTensor'.

Answer 1

Types are changing to float due to tf.image.resize_images.

Convert them back to uint as follows:

train_images = tf.cast(train_images, dtype=tf.uint8)
train_labels = tf.cast(train_labels, dtype=tf.uint8)

Output:

(tf.uint8, tf.uint8)

Versions of my code:

tensorflow version: 1.14.1-dev20190330
numpy version: 1.16.2