如何提高我的 cnn 模型的准确性？答案

【问题标题】：How to increase the accuracy of my cnn model?如何提高我的 cnn 模型的准确性？
【发布时间】：2019-11-27 07:55:14
【问题描述】：

我有 4 种不同类型的图像，例如自行车、汽车、飞机和乐器，我尝试使用这个数据集进行图像分类。然后当我训练模型时，我得到了这个准确度：0.62 我应该怎么做才能提高准确性？

from tensorflow import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPool2D
from keras.optimizers import RMSprop,Adam

#build the model
#set up the layers
model = Sequential()

model.add(Conv2D(filters = 8, kernel_size = (5,5),padding = 'Same', 
             activation ='relu', input_shape = (IMG_SIZE,IMG_SIZE,1)))
model.add(MaxPool2D(pool_size=(2,2)))
model.add(Dropout(0.25))

model.add(Conv2D(filters = 16, kernel_size = (3,3),padding = 'Same', 
             activation ='relu'))
model.add(MaxPool2D(pool_size=(2,2), strides=(2,2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation = "relu"))
model.add(Dropout(0.5))
model.add(Dense(10, activation = "softmax"))

# Define the optimizer
#Adam optimizer: Change the learning rate
optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999)

# Compile the model
model.compile(optimizer = optimizer , loss = "categorical_crossentropy", 
metrics=["accuracy"])

#train the model
model.fit(train_images, train_labels, epochs=10,batch_size=250) 

#evaluate accuracy
val_loss, val_acc = model.evaluate(val_images, val_labels)
print('validation accuracy:', val_acc)
print('validation loss:' , val_loss)

结果如下：

Epoch 1/10
18620/18620 [==============================] - 987s 53ms/step - loss: 2.0487 - acc: 0.39380/18620 [=======>......................] - ETA: 11:57 - loss: 4.0915 - acc: 0.278410500/18620 [===============>..............] - ETA: 7:07 - loss: 2.7013 - acc: 0.325015500/18620 [=======================>......] - ETA: 2:45 - loss: 2.2196 - acc: 0.3754
Epoch 2/10
18620/18620 [==============================] - 985s 53ms/step - loss: 1.1145 - acc: 0.5409TA: 14:05 - loss: 1.1721 - acc: 0.4987 7750/18620 [===========>..................] - ETA: 9:31 - loss: 1.1378 - acc: 0.5288 - ETA: 2:44 - loss: 1.1183 - acc: 0.5392
Epoch 3/10
18620/18620 [==============================] - 978s 53ms/step - loss: 1.0331 - acc: 0.5830TA: 14:17 - loss: 1.0323 - acc: 0.5845
Epoch 4/10
18620/18620 [==============================] - 975s 52ms/step - loss: 1.0032 - acc: 0.5942TA: 9:37 - loss: 1.0127 - acc: 0.5875 9750/18620 [==============>...............] - ETA: 7:41 - loss: 1.0119 - acc: 0.5892 - ETA: 5:19 - loss: 1.0122 - acc: 0.5902
Epoch 5/10
18620/18620 [==============================] - 973s 52ms/step - loss: 0.9680 - acc: 0.6137TA: 11:27 - loss: 0.9670 - acc: 0.6137 7000/18620 [==========>...................] - ETA: 9:58 - loss: 0.9718 - acc: 0.6066 15000/18620 [=======================>......] - ETA: 3:08 - loss: 0.9694 - acc: 0.6115
Epoch 6/10
18620/18620 [==============================] - 979s 53ms/step - loss: 0.9308 - acc: 0.62960/18620 [=============>................] - ETA: 8:36 - loss: 0.9311 - acc: 0.633110500/18620 [===============>..............] - ETA: 7:05 - loss: 0.9310 - acc: 0.6304
Epoch 7/10
18620/18620 [==============================] - 976s 52ms/step - loss: 0.9052 - acc: 0.63860/18620 [==========>...................] - ETA: 10:02 - loss: 0.9112 - acc: 0.6347 - ETA: 9:11 - loss: 0.9055 - acc: 0.6368 - ETA: 5:19 - loss: 0.9105 - acc: 0.6362
Epoch 8/10
18620/18620 [==============================] - 1008s 54ms/step - loss: 0.8755 - acc: 0.6507/18620 [==================>...........] - ETA: 5:52 - loss: 0.8746 - acc: 0.6513
Epoch 9/10
18620/18620 [==============================] - 994s 53ms/step - loss: 0.8479 - acc: 0.66140/18620 [===>..........................] - ETA: 14:12 - loss: 0.8474 - acc: 0.6560 3500/18620 [====>.........................] - ETA: 13:27 - loss: 0.8437 - acc: 0.6566 - ETA: 11:54 - loss: 0.8318 - acc: 0.6672 - ETA: 9:30 - loss: 0.8273 - acc: 0.6681 9500/18620 [==============>...............] - ETA: 8:08 - loss: 0.8390 - acc: 0.6653 - ETA: 6:09 - loss: 0.8399 - acc: 0.6660 - ETA: 1:53 - loss: 0.8473 - acc: 0.6628
Epoch 10/10
18620/18620 [==============================] - 997s 54ms/step - loss: 0.8108 - acc: 0.67490/18620 [=======>......................] - ETA: 11:54 - loss: 0.8146 - acc: 0.6650 - ETA: 10:45 - loss: 0.8196 - acc: 0.6652
4656/4656 [==============================] - 40s 9ms/stepETA: 1s
validation accuracy: 0.6265034364261168
validation loss: 0.964772748373628

【问题讨论】：

我认为您需要在每个时期获取验证损失的值，以了解数据是否过拟合。

标签： tensorflow keras classification conv-neural-network image-recognition

【解决方案1】：

首先你可以增加 CNN 过滤器（8 或 16 都不够），然后在 MaxPool 层之后使用 BatchNormalization 层。如果它不起作用，请更改优化器（如 SGD 等）

【讨论】：

【解决方案2】：

您可能需要根据

调整过滤器的确切值

数据集的复杂性。
您的神经网络的深度，但我建议从较早的 [32, 64, 128] 范围内的过滤器开始，在更深的层中增加到 [256, 512, 1024]。

如果您的输入图像大于 128×128，您可以选择使用大于 3 的内核大小。

学习更大的空间过滤器。
帮助减小卷大小。

如果您的图像小于 128×128，您可能需要考虑严格使用 1×1 和 3×3 滤镜。

还可以使用 Adam 优化器及其原始值。

【讨论】：