Tensorflow 数据增强中的随机数 map function

Question

I want to use the crop_central function with a random float between 0.50-1.00 for data augmentation. However, when using numpy.random.uniform(0.50, 1.00) and plotting the images the crop is constant. I debugged this by using 4 images and plotting 8 rows, the images are identical.

In general the question might be formulated as follows: How to use random numbers in the Dataset map functions?

def data_augment(image, label=None, seed=2020):
    # I want a random number here for every individual image
    image = tf.image.central_crop(image, np.random.uniform(0.50, 1.00)) # random crop central
    image = tf.image.resize(image, INPUT_SHAPE) # the original image size

    return image

train_dataset = (
    tf.data.Dataset
        .from_tensor_slices((train_paths, train_labels))
        .map(decode_image, num_parallel_calls=AUTO)
        .map(data_augment, num_parallel_calls=AUTO)
        .repeat()
        .batch(4)
        .prefetch(AUTO)
    )

# Code to view the images
for idx, (imgs, _) in enumerate(train_dataset):
    show_imgs(imgs, 'image', imgs_per_row=4)
    if idx is 8:
        del imgs
        gc.collect()
        break

Answer 1

Earlier, I misread the question. Here is the answer you were looking for.

I was able to recreate your issue using the below code -

Code to reproduce the issue - The output for the crop images were all identical.

%tensorflow_version 2.x
import tensorflow as tf
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array, array_to_img
from matplotlib import pyplot as plt
import numpy as np
AUTOTUNE = tf.data.experimental.AUTOTUNE

# Set the sub plot parameters
f, axarr = plt.subplots(5,4,figsize=(15, 15))

# Load just 4 images of Cifar10
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
images = x_train[:4]

for i in range(4):
  axarr[0,i].title.set_text('Original Image')
  axarr[0,i].imshow(x_train[i])

def data_augment(images):
    image = tf.image.central_crop(images, np.random.uniform(0.50, 1.00)) # random crop central
    image = tf.image.resize(image, (32,32)) # the original image size
    return image

dataset = tf.data.Dataset.from_tensor_slices((images)).map(lambda x: data_augment(x)).repeat(4) 

print(dataset)

ix = 0
i = 1
count = 0

for f in dataset:
  crop_img = array_to_img(f)
  axarr[i,ix].title.set_text('Crop Image')
  axarr[i,ix].imshow(crop_img)
  ix=ix+1
  count = count + 1
  if count == 4:
    i = i + 1
    count = 0
    ix = 0

Output - 1st Row is the Original Image. Remaining Rows are Crop Images.

Well it was very challenging and have provided below the two solutions -

Solution 1: Using np.random.uniform and tf.py_function .

1. Used np.random.uniform(0.50, 1.00) .
2. Used tf.py_function to decorate the function call - tf.py_function(data_augment, [x], [tf.float32]) .

Code to Solve the issue - The crop output images are now different and not identical.

%tensorflow_version 2.x
import tensorflow as tf
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array, array_to_img
from matplotlib import pyplot as plt
import numpy as np
AUTOTUNE = tf.data.experimental.AUTOTUNE

# Set the sub plot parameters
f, axarr = plt.subplots(5,4,figsize=(15, 15))

# Load just 4 images of Cifar10
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
images = x_train[:4]

for i in range(4):
  axarr[0,i].title.set_text('Original Image')
  axarr[0,i].imshow(x_train[i])

def data_augment(images):
    image = tf.image.central_crop(images, np.random.uniform(0.50, 1.00)) # random crop central
    image = tf.image.resize(image, (32,32)) # the original image size
    return image

dataset = tf.data.Dataset.from_tensor_slices((images)).map(lambda x: tf.py_function(data_augment, [x], [tf.float32])).repeat(4)

ix = 0
i = 1
count = 0

for f in dataset:
  for l in f:
    crop_img = array_to_img(l)
    axarr[i,ix].title.set_text('Crop Image')
    axarr[i,ix].imshow(crop_img)
    ix=ix+1
    count = count + 1
    if count == 4:
      i = i + 1
      count = 0
      ix = 0

Output - 1st Row is the original image. Remaining rows are Crop Images.

Solution 2: Using tf.random.uniform and tf.py_function .

1. Used tf.random.uniform(shape=(), minval=0.50, maxval=1).numpy() .
2. Just by using the above option, the code doesn't work as it throws the error AttributeError: 'Tensor' object has no attribute 'numpy' . To fix this issue, you need to decorate your function with tf.py_function(data_augment, [x], [tf.float32]) .

Code to Solve the issue - The crop output images are now different and not identical.

%tensorflow_version 2.x
import tensorflow as tf
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array, array_to_img
from matplotlib import pyplot as plt
import numpy as np
AUTOTUNE = tf.data.experimental.AUTOTUNE

# Set the sub plot parameters
f, axarr = plt.subplots(5,4,figsize=(15, 15))

# Load just 4 images of Cifar10
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
images = x_train[:4]

for i in range(4):
  axarr[0,i].title.set_text('Original Image')
  axarr[0,i].imshow(x_train[i])

def data_augment(images):
    image = tf.image.central_crop(images, tf.random.uniform(shape=(), minval=0.50, maxval=1).numpy()) # random crop central
    image = tf.image.resize(image, (32,32)) # the original image size
    return image

dataset = tf.data.Dataset.from_tensor_slices((images)).map(lambda x: tf.py_function(data_augment, [x], [tf.float32])).repeat(4)

ix = 0
i = 1
count = 0

for f in dataset:
  for l in f:
    crop_img = array_to_img(l)
    axarr[i,ix].title.set_text('Crop Image')
    axarr[i,ix].imshow(crop_img)
    ix=ix+1
    count = count + 1
    if count == 4:
      i = i + 1
      count = 0
      ix = 0

Output - 1st Row is the Original Image. Remaining Rows are Crop Images.

Hope this answers your question. Happy Learning.