训练 CNN 后准确率低

Question

我尝试训练一个使用 Keras 对手写数字进行分类的 CNN 模型，但我在训练中得到了低准确率（低于 10%）和一个很大的错误。 我尝试了一个简单的神经网络，没有任何问题。

这是我的代码。

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

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

#Explore data
print(y_train[12])
print(np.shape(x_train))
print(np.shape(x_test))
#we have 60000 imae for the training and 10000 for testing

# Scaling data
x_train = x_train/255
y_train = y_train/255
#reshape the data
x_train = x_train.reshape(60000,28,28,1)
x_test = x_test.reshape(10000,28,28,1)
y_train = y_train.reshape(60000,1)
y_test = y_test.reshape(10000,1)

#Create a model
model = keras.Sequential([
keras.layers.Conv2D(64,(3,3),(1,1),padding = "same",input_shape=(28,28,1)),
keras.layers.MaxPooling2D(pool_size = (2,2),padding = "valid"),
keras.layers.Conv2D(32,(3,3),(1,1),padding = "same"),
keras.layers.MaxPooling2D(pool_size = (2,2),padding = "valid"),
keras.layers.Flatten(),
keras.layers.Dense(128,activation = "relu"),
keras.layers.Dense(10,activation = "softmax")])

model.compile(optimizer = "adam",
loss = "sparse_categorical_crossentropy",
metrics  = ['accuracy'])

model.fit(x_train,y_train,epochs=10)
test_loss,test_acc = model.evaluate(x_test,y_test)
print("\ntest accuracy:",test_acc)

有人可以建议我如何改进我的模型吗？

Answer 1

你的问题在这里：

x_train = x_train/255
y_train = y_train/255 # makes no sense

您应该重新缩写x_test，而不是y_train。

x_train = x_train/255
x_test = x_test/255

可能只是你的拼写错误。更改这些行，您将有95％+精度。

Answer 2

您的模型有一些缩放问题并尝试使用 tf 2.0

x_train /= 255
x_test /= 255

您不需要缩放测试的所有数据 正如你所做的那样：

x_train = x_train/255
y_train = y_train/255

之后，我们可以将标签转换为 one-hot 编码

from tensorflow.keras.utils import to_categorical

y_train = to_categorical(y_train, 10)

y_test = to_categorical(y_test, 10)

这有助于：

loss='categorical_crossentropy', 

Sequential API 允许我们将层堆叠在一起。唯一的缺点是我们在使用这些模型时不能有多个输出或输入。不过，我们可以创建一个 Sequential 对象并使用 add() 函数将层添加到我们的模型中。 尝试使用更多 API 来使您的模型更加流畅和准确，因为 Tf 2.0 上存在使用 add 函数 因为我们可以给 Conv2D 4 时间来平滑：

seq_model.add(Conv2D(filters=32, kernel_size=(5,5), activation='relu', 
input_shape=x_train.shape[1:]))
seq_model.add(Conv2D(filters=32, kernel_size=(5,5), activation='relu'))
seq_model.add(Conv2D(filters=64, kernel_size=(3, 3), activation='relu'))
seq_model.add(Conv2D(filters=64, kernel_size=(3, 3), activation='relu'))

在代码中你可以使用 dropout ：

seq_model.add(Dropout(rate=0.25))

完整模型：

%tensorflow_version 2.x
from tensorflow.keras.datasets import mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()    
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
x_train = x_train.reshape(x_train.shape[0], 28, 28, 1)
x_test = x_test.reshape(x_test.shape[0], 28, 28, 1)
from tensorflow.keras.utils import to_categorical
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D, Dense, Flatten, Dropout

seq_model = Sequential()
seq_model.add(Conv2D(filters=32, kernel_size=(5,5), activation='relu', 
input_shape=x_train.shape[1:]))
seq_model.add(Conv2D(filters=32, kernel_size=(5,5), activation='relu'))
seq_model.add(MaxPool2D(pool_size=(2, 2)))
seq_model.add(Dropout(rate=0.25))
seq_model.add(Conv2D(filters=64, kernel_size=(3, 3), activation='relu'))
seq_model.add(Conv2D(filters=64, kernel_size=(3, 3), activation='relu'))
seq_model.add(MaxPool2D(pool_size=(2, 2)))
seq_model.add(Dropout(rate=0.25))
seq_model.add(Flatten())
seq_model.add(Dense(256, activation='relu'))
seq_model.add(Dropout(rate=0.5))
seq_model.add(Dense(10, activation='softmax'))


seq_model.compile(
    loss='categorical_crossentropy', 
    optimizer='adam', 
    metrics=['accuracy']
)

epochsz = 3 # number of epch 
batch_sizez = 32 # the batch size ,can be 64 , 128 so other
seq_model.fit(x_train,y_train, batch_size=batch_sizez, epochs=epochsz)

结果：

Train on 60000 samples
Epoch 1/3

60000/60000 [==============================] - 186s 3ms/sample - loss: 0.1379 - accuracy: 0.9588

Epoch 2/3
60000/60000 [==============================] - 187s 3ms/sample - loss: 0.0677 - accuracy: 0.9804
Epoch 3/3

60000/60000 [==============================] - 187s 3ms/sample - loss: 0.0540 - accuracy: 0.9840