如何在第一个epoch直接使用map函数生成的变换，而不是每个epoch都执行map函数？

Question

I want to apply some transformation to dataset by using map function. However, I found the map function was executed in every epoch. Is it possible that the map function is only executed in first epoch and the following epochs directly use the transformation generated in first epoch?

Answer 1

If you set the seed with Tensorflow and apply transformations with tf.image , the random transformations will be consistent between epochs.

import tensorflow as tf
from skimage import data
tf.random.set_seed(42)
import matplotlib.pyplot as plt

def transform(image):
    image = tf.image.random_hue(image, 0.5, 1.)
    image = tf.image.random_flip_left_right(image)
    image = tf.image.random_flip_up_down(image)
    return image

X = tf.stack([data.chelsea() for i in range(4)])
ds = tf.data.Dataset.from_tensor_slices(X).map(transform)

inputs = [[], []]

for epoch in range(2):
    for sample in ds:
        inputs[epoch].append(sample)

inputs_paired = [i for s in inputs for i in s]

fig = plt.figure(figsize=(16, 8))
plt.subplots_adjust(wspace=.1, hspace=.1)
for i in range(8):
    ax = plt.subplot(2, 4, i + 1)
    ax.set_xticks([])
    ax.set_yticks([])
    ax.imshow(inputs_paired[i])
plt.show()

Top is the first epoch, bottom is the second epoch.