Keras experimental RandomFlip and RandomRotation do not work with map

Question

This code generates an error that I don't understand. Can someone explain me please?

import tensorflow as tf

def augment(img):
    data_augmentation = tf.keras.Sequential([
              tf.keras.layers.experimental.preprocessing.RandomFlip('horizontal'),
              tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
             ])
    img = tf.expand_dims(img, 0)
    return data_augmentation(img)

# generate 10 images 8x8 RGB
data = np.random.randint(0,255,size=(10, 8, 8, 3))
dataset = tf.data.Dataset.from_tensor_slices(data)

# and augment... -> bug
dataset = dataset.map(augment)

# note that the follwing works
for im in dataset:
   augment(im)

and a get

ValueError: Tensor-typed variable initializers must either be wrapped in an init_scope or callable (e.g., `tf.Variable(lambda : tf.truncated_normal([10, 40]))`) when building functions. Please file a feature request if this restriction inconveniences you.

I tried on Google Colab and have Tensorflow 2.4.1 on my computer. Note that with resize or rescale it works (as it is in this example https://www.tensorflow.org/tutorials/images/data_augmentation but they didn't tried with RandomRotate even if they use it in a loop).

Answer 1

Here is the answer...

import numpy as np
import tensorflow as tf

data_augmentation = tf.keras.Sequential([
              tf.keras.layers.experimental.preprocessing.RandomFlip('horizontal'),
              tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
             ])

# generate 10 images 8x8 RGB
data = np.random.randint(0,255,size=(10, 8, 8, 3))
dataset = tf.data.Dataset.from_tensor_slices(data).batch(5)

# and augment... -> bug
dataset = dataset.map(lambda x: data_augmentation(x))

Strange, if we use a lambda function it works, if we define a function which only calls data_augmentation it fails...