在MNIST上的LeNet CNN的准确率非常低

Question

我一直在看其他教程，它们能够在仅仅10个时代之后获得90%的准确率。所以我猜我的实现中有一些错误，因为我的准确率很低，在10个时代之后，它不到1%，而且几乎没有增加。我正在使用MNIST数据集，非常感谢您的帮助

from PIL import Image
from tensorflow.keras import datasets, layers, models
from keras.layers import Dense, Dropout, Flatten
import matplotlib.pyplot as plt
import numpy as np
import keras

(X_train, y_train), (X_test, y_test) = tf.keras.datasets.mnist.load_data()
X_train = X_train/255
X_test = X_test/255
X_train_processed = np.reshape(X_train,[X_train.shape[0],X_train.shape[1],X_train.shape[2],1])
X_test_processed = np.reshape(X_test,[X_test.shape[0],X_test.shape[1],X_test.shape[2],1])
X_train_processed = np.pad(X_train_processed, ((0,0),(2,2),(2,2),(0,0)), 'constant')
X_test_processed = np.pad(X_test_processed, ((0,0),(2,2),(2,2),(0,0)), 'constant')

X_train_processed = tf.image.resize(
    images = X_train_processed,
    size = np.array([32,32])

)
X_test_processed = tf.image.resize(
    images = X_test_processed,
    size = np.array([32,32])

)
Y_train_processed = tf.one_hot(y_train,10)
Y_test_processed = tf.one_hot(y_test,10)

Lnet = tf.keras.Sequential()
#First Layer
Lnet.add(
    tf.keras.layers.Conv2D(
        filters = 6, 
        kernel_size = (5,5), 
        strides = (1,1), 
        padding = 'valid', 
        activation = 'relu', 
        #kernel_initializer = keras.initializers.glorot_normal(seed=0)
    )
)
Lnet.add(
    tf.keras.layers.AveragePooling2D(
        pool_size = (2,2),
        strides = (2,2),
        padding = 'valid'
    )
)
#Second Layer
Lnet.add(
    tf.keras.layers.Conv2D(
        filters = 16, 
        kernel_size = (5,5), 
        strides = (1,1), 
        padding = 'valid', 
        activation = 'relu'#, 
        #kernel_initializer = keras.initializers.glorot_normal(seed=0)
    )
)
Lnet.add(
    tf.keras.layers.AveragePooling2D(
        pool_size = (2,2),
        strides = (2,2),
        padding = 'valid'
    )
)
Lnet.add(tf.keras.layers.Flatten())
Lnet.add(
    tf.keras.layers.Dense(
        units = 120,
        activation = 'relu'
    )
)
Lnet.add(tf.keras.layers.Flatten())
Lnet.add(
    tf.keras.layers.Dense(
        units = 84,
        activation = 'relu'
    )
)
Lnet.add(
    tf.keras.layers.Dense(
        units = 10,
        activation = 'softmax'
    )
)

Lnet.compile(
    loss = keras.losses.categorical_crossentropy,
    optimizer = 'Adam',
    metrics = ['Accuracy']
)

Lnet.fit(
    x = X_train_processed,
    y = Y_train_processed,
    batch_size = 128,
    epochs = 10,
)

score = Lnet.evaluate(
    x = X_test_processed,
    y = Y_test_processed
)
print(score[1])

乌普图特：

Epoch 1/10
469/469 [==============================] - 8s 18ms/step - loss: 0.3533 - accuracy: 0.0000e+00
Epoch 2/10
469/469 [==============================] - 8s 18ms/step - loss: 0.1013 - accuracy: 5.1667e-05
Epoch 3/10
469/469 [==============================] - 8s 18ms/step - loss: 0.0730 - accuracy: 2.3167e-04
Epoch 4/10
469/469 [==============================] - 10s 21ms/step - loss: 0.0582 - accuracy: 4.8833e-04
Epoch 5/10
469/469 [==============================] - 9s 19ms/step - loss: 0.0478 - accuracy: 9.3333e-04
Epoch 6/10
469/469 [==============================] - 11s 23ms/step - loss: 0.0405 - accuracy: 0.0019
Epoch 7/10
469/469 [==============================] - 12s 25ms/step - loss: 0.0371 - accuracy: 0.0026
Epoch 8/10
469/469 [==============================] - 11s 23ms/step - loss: 0.0301 - accuracy: 0.0057
Epoch 9/10
469/469 [==============================] - 12s 25ms/step - loss: 0.0280 - accuracy: 0.0065
Epoch 10/10
469/469 [==============================] - 11s 24ms/step - loss: 0.0260 - accuracy: 0.0085
313/313 [==============================] - 1s 3ms/step - loss: 0.0323 - accuracy: 0.0080
0.008030000142753124

Answer 1

我将两个导入和度量从Accuracy更改为“准确性”。请检查下面的更改。通过这些小的修改，精度达到98.7%（见下文）。一个建议是不要混合来自两个不同包（Keras和Tensorflow）的函数

from PIL import Image
from tensorflow.keras import datasets, layers, models
from tensorflow.keras.layers import Dense, Dropout, Flatten
import matplotlib.pyplot as plt
import numpy as np
import keras

(X_train, y_train), (X_test, y_test) = tf.keras.datasets.mnist.load_data()
X_train = X_train/255
X_test = X_test/255
X_train_processed = np.reshape(X_train,[X_train.shape[0],X_train.shape[1],X_train.shape[2],1])
X_test_processed = np.reshape(X_test,[X_test.shape[0],X_test.shape[1],X_test.shape[2],1])
X_train_processed = np.pad(X_train_processed, ((0,0),(2,2),(2,2),(0,0)), 'constant')
X_test_processed = np.pad(X_test_processed, ((0,0),(2,2),(2,2),(0,0)), 'constant')

X_train_processed = tf.image.resize(
    images = X_train_processed,
    size = np.array([32,32])

)
X_test_processed = tf.image.resize(
    images = X_test_processed,
    size = np.array([32,32])

)
Y_train_processed = tf.one_hot(y_train,10)
Y_test_processed = tf.one_hot(y_test,10)

Lnet = tf.keras.Sequential()
#First Layer
Lnet.add(
    tf.keras.layers.Conv2D(
        filters = 6, 
        kernel_size = (5,5), 
        strides = (1,1), 
        padding = 'valid', 
        activation = 'relu', 
        #kernel_initializer = keras.initializers.glorot_normal(seed=0)
    )
)
Lnet.add(
    tf.keras.layers.AveragePooling2D(
        pool_size = (2,2),
        strides = (2,2),
        padding = 'valid'
    )
)
#Second Layer
Lnet.add(
    tf.keras.layers.Conv2D(
        filters = 16, 
        kernel_size = (5,5), 
        strides = (1,1), 
        padding = 'valid', 
        activation = 'relu'#, 
        #kernel_initializer = keras.initializers.glorot_normal(seed=0)
    )
)
Lnet.add(
    tf.keras.layers.AveragePooling2D(
        pool_size = (2,2),
        strides = (2,2),
        padding = 'valid'
    )
)
Lnet.add(tf.keras.layers.Flatten())
Lnet.add(
    tf.keras.layers.Dense(
        units = 120,
        activation = 'relu'
    )
)
Lnet.add(tf.keras.layers.Flatten())
Lnet.add(
    tf.keras.layers.Dense(
        units = 84,
        activation = 'relu'
    )
)
Lnet.add(
    tf.keras.layers.Dense(
        units = 10,
        activation = 'softmax'
    )
)

Lnet.compile(
    loss = tf.keras.losses.categorical_crossentropy,
    optimizer = 'Adam',
    metrics = ['accuracy']
)

Lnet.fit(
    x = X_train_processed,
    y = Y_train_processed,
    batch_size = 128,
    epochs = 10,
)

score = Lnet.evaluate(
    x = X_test_processed,
    y = Y_test_processed
)
print(score[1])

Epoch 9/10
469/469 [==============================] - 30s 64ms/step - loss: 0.0260 - accuracy: 0.9916
Epoch 10/10
469/469 [==============================] - 30s 64ms/step - loss: 0.0239 - accuracy: 0.9922
313/313 [==============================] - 3s 10ms/step - loss: 0.0394 - accuracy: 0.9876
0.9876000285148621

完整的代码是here