Question

我训练了CNN以预测图像中的数字并保存了它。如果我在lib内部的x_test数据上进行测试，则可以预测正常。但是，当我加载自定义图像时-它的预测效果不佳，我也不知道为什么。

from tensorflow import keras
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
import PIL

def normalize_image (img) :
    """
    Normalize image
    :param img: input image with range 0-255 int
    :return: normalized image (with range 0 - 1 float)
    """

    # taken from https://stackoverflow.com/questions/46689428/convert-np-array-of-type-float64-to-type-uint8-scaling-values
    info = np.iinfo(img.dtype)
    norm_img = img.astype(np.float64) / info.max

    # tensorflow mnist use black background - so change bg from white to black
    for i in range(len(norm_img)):
        for j in range(len(norm_img[i])):
            norm_img[i][j] = abs(norm_img[i][j] - 1)

    return norm_img


# load my model
new_model = keras.models.load_model('numbers.model')

# image with number 4
img = PIL.Image.open('C:/Users/Horseman.mini/Desktop/4.png').convert('L').resize((28, 28), PIL.Image.ANTIALIAS)
img = np.array(img)
img = normalize_image(img)

# create my custom tests list
my_tests = np.array([img])

# show the image which value I want to predict
plt.imshow(img, cmap="gray")
plt.show()
predictions = new_model.predict(my_tests)
print("Prediction value:", np.argmax(predictions[0])) # value that should be predicted

My 4.png image as is

My 4.png image plotted after resizing and normalizing

output

Prediction value: 1

预测是错误的，所以我认为dtypes不同：

for comparing x_test data and my custom data lists

mnist = tf.keras.datasets.mnist  # mnist is a dataset of 28x28 images of handwritten digits and their labels
(x_train, y_train),(x_test, y_test) = mnist.load_data()  # unpacks images to x_train/x_test and labels to y_train/y_test

x_train = tf.keras.utils.normalize(x_train, axis=1)  # scales data between 0 and 1
x_test = tf.keras.utils.normalize(x_test, axis=1)  # scales data between 0 and 1

print(x_test[0])
print(my_tests[0])

输出很大，我将发布一行以代表数字和数据类型

# x_test[0]
 [0.         0.         0.         0.         0.         0.
  0.         0.         0.         0.         0.33861822 0.450819
  0.32890224 0.02719964 0.         0.         0.         0.
  0.         0.         0.         0.         0.         0.
  0.         0.         0.         0.        ] 

# my_tests[0]
 [0.         0.         0.         0.         0.         0.
  0.         0.         0.00392157 0.         0.09803922 0.01960784
  0.         0.         0.         0.09019608 0.19607843 0.
  0.         0.         0.         0.         0.         0.
  0.         0.         0.         0.        ]

# x_test[0].dtype
float64

# my_tests[0].dtype
float64

无法理解为什么它不起作用。数据类型相同，图像数组从0到1统一，图像正确，经过训练且自定义图像为28x28像素（自定义大小在调整大小后为自定义），但是出了点问题，我不完全是。如果有人可以帮助我，我将不胜感激！

Link to my model on github (numbers.model.rar)

MNIST模型无法正确预测值

0 个答案: