深度学习分类CIFAR-10验证精度低

Question

经过5次测试，我的验证分数约为0.07，我无法理解我做错了什么。我正在努力学习生成的图像分类器。

In1

from tensorflow.keras.datasets import cifar10

In2

(features_train, label_train), (features_test, label_test) = cifar10.load_data()

In3

features_train.shape

In4

features_test.shape

In5

batch_size = 16
img_height = 32
img_width = 32

In6

from tensorflow.keras.preprocessing.image import ImageDataGenerator

In7

train_img_gen = ImageDataGenerator(rescale=1./255,
                                  width_shift_range = 0.1,
                                  height_shift_range = 0.1,
                                  horizontal_flip = True)

In8

val_img_gen = ImageDataGenerator(rescale = 1./255)

In9

train_data_gen = train_img_gen.flow(features_train, label_train, batch_size = batch_size)

In10

val_data_gen = train_img_gen.flow(features_test, label_test, batch_size = batch_size)

In11

train_data_gen

In12

import numpy as np
import tensorflow as tf
from tensorflow.keras import layers

In13

np.random.seed(8)
tf.random.set_seed(8)

In14

model = tf.keras.Sequential([layers.Conv2D(64, 3, activation = 'relu', input_shape = (img_height, img_width, 3)),
                            layers.MaxPooling2D(),
                            layers.Conv2D(128, 3, activation='relu'),
                            layers.MaxPooling2D(),
                            layers.Flatten(),
                            layers.Dense(128, activation='relu'),
                            layers.Dense(10, activation='softmax')])

In15

optimizer = tf.keras.optimizers.Adam(0.001)

In16

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

In17

model.summary()

In18

model.fit(train_data_gen, 
          steps_per_epoch=len(features_train)//batch_size, 
          epochs=5, 
          validation_data=val_data_gen,
         validation_steps=len(features_test)//batch_size)

Answer 1

标准的“精度”度量不应该工作，因为模型的输出是一个热编码，但是您的标签是整数编码的。因此，在编译模型时尝试使用sparse_categorical_accuracy，如下所示：

model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['sparse_categorical_accuracy'])

而且，为了达到更好的表现，训练更多的时代会更好。例如，100个时代：

model.fit(train_data_gen, 
          steps_per_epoch=len(features_train)//batch_size, 
          epochs=100, 
          validation_data=val_data_gen,
         validation_steps=len(features_test)//batch_size)

另一件事是，Conv2D层的过滤器数量似乎太大，不适合这样的直进图像分类任务。如果你使用更多的卷积层，也会比使用更多的滤波器更好。同样，这可能会导致模型过度适应，所以最后尝试使用Dropout层来解决这个问题。如下所示：

model = tf.keras.Sequential([
    layers.Conv2D(32, 3, activation = 'relu', input_shape = (img_height, img_width, 3)),
    layers.Conv2D(32, 3, activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(64, 3, activation='relu'),
    layers.Conv2D(64, 3, activation='relu'),
    layers.MaxPooling2D(),
    layers.Flatten(),
    layers.Dropout(0.25),
    layers.Dense(128, activation='relu'),
    layers.Dense(10, activation='softmax')])