使用Tensorflow随机旋转和裁剪顺序抛出错误

Question

我正在尝试在 Keras 页面上编辑此代码：https://keras.io/examples/vision/deeplabv3_plus/。我想做的一件事是为数据集添加一个扩充函数。这是我写的：

def image_augmentation(img):
    img = tf.image.random_flip_left_right(img)
    img = tf.image.random_flip_up_down(img)
    img = tf.image.random_brightness(img, 0.2)
    img = tf.image.random_crop(value = img, size=(2, 2))
    img = tf.keras.preprocessing.image.random_rotation(img, 90, row_axis=0, col_axis=1,channel_axis=2)

但是，最后两行引发了我遇到问题的错误。

这一行 img = tf.image.random_crop(value = img, size=(2, 2)) 抛出这个错误：

ValueError: Dimensions must be equal, but are 3 and 2 for '{{node random_crop/GreaterEqual}} = GreaterEqual[T=DT_INT32](random_crop/Shape, random_crop/size)' with input shapes: [3], [2].

这一行 img = tf.keras.preprocessing.image.random_rotation(img, 90, row_axis=0, col_axis=1,channel_axis=2) 抛出这个错误：

AttributeError: in user code:

    File "<ipython-input-3-8db8a894e0d0>", line 41, in data_loader  *
        mask = image_process(mask_list, mask=True)
    File "<ipython-input-3-8db8a894e0d0>", line 28, in image_process  *
        img = image_augmentation(img)
    File "<ipython-input-7-037af66af3d3>", line 18, in image_augmentation  *
        img = tf.keras.preprocessing.image.random_rotation(img, 90, row_axis=0, col_axis=1,channel_axis=2)
    File "/usr/local/lib/python3.7/dist-packages/keras_preprocessing/image/affine_transformations.py", line 56, in random_rotation  *
        x = apply_affine_transform(x, theta=theta, channel_axis=channel_axis,
    File "/usr/local/lib/python3.7/dist-packages/keras_preprocessing/image/affine_transformations.py", line 323, in apply_affine_transform  *
        x = np.rollaxis(x, channel_axis, 0)
    File "<__array_function__ internals>", line 6, in rollaxis  **
        
    File "/usr/local/lib/python3.7/dist-packages/numpy/core/numeric.py", line 1290, in rollaxis
        n = a.ndim

    AttributeError: 'Tensor' object has no attribute 'ndim'

Answer 1

对于行中的第一个错误，

img = tf.image.random_crop(value = img, size=(2, 2))

在方法的documentation中tf.image.random_crop明确提到，

If a dimension should not be cropped, pass the full size of that dimension. For example, RGB images can be cropped with size = [crop_height, crop_width, 3].

所以，为了修复错误，

img = tf.image.random_crop(value = img, size=(2, 2, 3))

对于行中的第二个错误，

img = tf.keras.preprocessing.image.random_rotation(img, 90, row_axis=0, col_axis=1,channel_axis=2)

此方法要求 img 是 NumPy 数组（张量）而不是 Tensor 对象。此外，方法 returns 以 NumPy 数组的形式旋转图像。

如果函数 image_augmentation 的期望输出是 ndarray 那么您只需要进行以下更改，

img = tf.keras.preprocessing.image.random_rotation(img.numpy(), 90, row_axis=0, col_axis=1,channel_axis=2)

否则，如果所需的输出类型是 Tensor，我们可以使用来自 TensorFlow Addons 包的 tfa.image.rotate 方法，

import math
import random 

upper = 90 * (math.pi/180.0) # degrees -> radian
lower = 0 * (math.pi/180.0)

def rand_degree():
    return random.uniform( lower , upper )

def image_augmentation(img):
    ...
    img = tfa.image.rotate( img , rand_degree() )
    # img is a Tensor
    return img

使用Tensorflow随机旋转和裁剪顺序抛出错误

Using Tensorflow random rotation and cropping sequentially throws errors

python

computer-vision

keras

tensorflow

semantic-segmentation