我的Keras卷积模型预测了从不同路径导入的相同图像，但是预测结果不同

如何解决我的Keras卷积模型预测了从不同路径导入的相同图像，但是预测结果不同

我创建了一个CNN模型，用于使用 mnist时尚数据集来预测时尚。训练完模型后，我尝试预测了从Keras加载的测试图像之一和从我的PC导入到我的Google Colab笔记本中相同但又相同的另一图像，事实证明，预测结果并不相同。我该如何解决这个问题？

这是我导入数据集的方式：

import tensorflow as tf
from tensorflow import keras
fashion_mnist = keras.datasets.fashion_mnist
(x_train,y_train),(x_test,y_test) = fashion_mnist.load_data()

数据操作：

from keras.utils import to_categorical
yTest = to_categorical(y_test)
yTrain = to_categorical(y_train)
xTrain = x_train.reshape((60000,28,1))
xTest = x_test.reshape(10000,1)

模型设置：

from keras.layers import Dense,Flatten,Conv2D,Dropout,MaxPool2D,BatchNormalization
from keras.callbacks import ModelCheckpoint

model = keras.Sequential()

#Adding the convolutional layer
model.add(Conv2D(50,kernel_size=3,activation='relu',padding = 'same',input_shape = (28,1)))
model.add(MaxPool2D(pool_size = (2,2),strides = 1,padding = 'valid'))
model.add(Dropout(0.5))
model.add(Conv2D(40,kernel_size = 3,activation = 'relu',padding = 'same'))
model.add(MaxPool2D(pool_size = (2,padding = 'valid'))
model.add(Dropout(0.5))
model.add(Conv2D(30,strides = 2,padding = 'valid'))
model.add(Dropout(0.5))
model.add(Conv2D(10,padding = 'same'))
model.add(Dropout(0.5))

#Connecting the CNN layers to the ANN
model.add(Flatten())
model.add(Dense(60,activation='relu'))
model.add(Dense(40,activation = 'relu'))
model.add(Dense(10,activation = 'softmax'))
model.load_weights('mnist_fashion.h5')

# Compiling the model
opt = tf.keras.optimizers.Adam(learning_rate=0.0001)
model.compile(optimizer=opt,loss = 'categorical_crossentropy',metrics = ['accuracy']

培训模型：

model = keras.Sequential()

#Adding the convolutional layer
model.add(Conv2D(50,activation = 'softmax'))

模型的性能：

            precision    recall  f1-score   support

       0       0.89      0.88      0.88      1000
       1       0.99      0.99      0.99      1000
       2       0.88      0.89      0.89      1000
       3       0.93      0.93      0.93      1000
       4       0.87      0.89      0.88      1000
       5       0.99      0.98      0.99      1000
       6       0.79      0.78      0.78      1000
       7       0.97      0.98      0.97      1000
       8       0.99      0.98      0.99      1000
       9       0.97      0.97      0.97      1000



   accuracy                           0.93     10000
   macro avg       0.93      0.93      0.93     10000
   weighted avg    0.93      0.93      0.93     10000

来自数据集预测的图片

 #From the dataset
    import numpy as np
    image = xTrain[0].reshape(1,1)
    prd = model.predict(image)
    new_prd = np.argmax(prd,axis  = 1)
    print(f"Prediction = {new_prd}")
    print(f"Full Prediction = {prd}")
    print(f"Label = {y_train[0]}")

数据集结果

Prediction = [9]
Full Prediction = [[1.6268513e-07 2.3548612e-08 1.5456487e-07 8.6898848e-07 1.9692785e-09
  4.4544859e-04 6.6932116e-06 1.4004705e-02 4.1784686e-05 9.8550016e-01]]
Label = 9

导入的图片预测

imported_img = plt.imread("mnist fashion sample.png")
yolo = imported_img.reshape(1,1)
super_prd = model.predict(yolo)
prediction = np.argmax(super_prd,axis = 1)
print(f"Prediction = {prediction}")
print(f"Full Prediction = {super_prd}")
print(f"Label = {y_train[0]}")

导入的图片预测结果

Prediction = [8]
Full Prediction = [[2.49403762e-04 1.69450897e-04 4.47237398e-04 3.05729372e-05
  1.10463676e-04 4.34053177e-03 5.16198808e-04 8.16224664e-02
  8.73587310e-01 3.89263593e-02]]
Label = 9

解决方法

我解决了问题！

我做错了，因为我在训练之前没有对图片进行归一化。这可能会导致错误，因为数据像素范围可能太复杂，以至于relu激活函数无法计算或预测。

谢谢！