第T10周：数据增强

• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍖 原作者：K同学啊

要求：

1. 学会在代码中使用数据增强手段来提高acc
2. 请探索更多的数据增强手段并记录

在本教程中，你将学会如何进行数据增强，并通过数据增强用少量数据达到非常非常棒的识别准确率。我将展示两种数据增强方式，以及如何自定义数据增强方式并将其放到我们代码当中，两种数据增强方式如下：

• 将数据增强模块嵌入model中
• 在Dataset数据集中进行数据增强

我的环境

• 操作系统：CentOS7
• 显卡：RTX3090 两张
• 显卡驱动：550.78
• CUDA版本: 12.4
• 语言环境：Python3.9.19
• 编译器：Jupyter Lab
• 深度学习环境：
  • TensorFlow-2.17.0 (GPU版本)

一、前期准备工作

1. 设置GPU

import matplotlib.pyplot as plt
import numpy as np
#隐藏警告
import warnings
warnings.filterwarnings('ignore')

from tensorflow.keras import layers
import tensorflow as tf
gpus = tf.config.list_physical_devices("GPU")

if gpus:
    tf.config.experimental.set_memory_growth(gpus[0], True)  #设置GPU显存用量按需使用
    tf.config.set_visible_devices([gpus[0]],"GPU")

# 打印显卡信息，确认GPU可用
print(gpus)

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU'), PhysicalDevice(name='/physical_device:GPU:1', device_type='GPU')]

2. 加载数据

关于 tf.keras.preprocessing.image_dataset_from_directory 的介绍，我这里就不赘述了，不明白的同学直接看这里：blog.csdn.net/qq_38251616…

data_dir   = "./data/"
img_height = 224
img_width  = 224
batch_size = 32

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="training",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 600 files belonging to 2 classes.
Using 420 files for training.

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="validation",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 600 files belonging to 2 classes.
Using 180 files for validation.

由于原始数据集不包含测试集，因此需要创建一个。使用 tf.data.experimental.cardinality 确定验证集中有多少批次的数据，然后将其中的 20% 移至测试集。

val_batches = tf.data.experimental.cardinality(val_ds)
test_ds     = val_ds.take(val_batches // 5)
val_ds      = val_ds.skip(val_batches // 5)

print('Number of validation batches: %d' % tf.data.experimental.cardinality(val_ds))
print('Number of test batches: %d' % tf.data.experimental.cardinality(test_ds))

Number of validation batches: 4
Number of test batches: 1

class_names = train_ds.class_names
print(class_names)

['cat', 'dog']

AUTOTUNE = tf.data.AUTOTUNE

def preprocess_image(image,label):
    return (image/255.0,label)

# 归一化处理
train_ds = train_ds.map(preprocess_image, num_parallel_calls=AUTOTUNE)
val_ds   = val_ds.map(preprocess_image, num_parallel_calls=AUTOTUNE)
test_ds  = test_ds.map(preprocess_image, num_parallel_calls=AUTOTUNE)

train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds   = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

plt.figure(figsize=(15, 10))  # 图形的宽为15高为10

for images, labels in train_ds.take(1):
    for i in range(8):
        
        ax = plt.subplot(5, 8, i + 1) 
        plt.imshow(images[i])
        plt.title(class_names[labels[i]])
        
        plt.axis("off")

二、数据增强

我们可以使用 tf.keras.layers.experimental.preprocessing.RandomFlip 与 tf.keras.layers.experimental.preprocessing.RandomRotation 进行数据增强

• tf.keras.layers.experimental.preprocessing.RandomFlip：水平和垂直随机翻转每个图像。

• tf.keras.layers.experimental.preprocessing.RandomRotation：随机旋转每个图像

data_augmentation = tf.keras.Sequential([
  tf.keras.layers.experimental.preprocessing.RandomFlip("horizontal_and_vertical"),
  tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
])

# Add the image to a batch.
image = tf.expand_dims(images[i], 0)

plt.figure(figsize=(8, 8))
for i in range(9):
    augmented_image = data_augmentation(image)
    ax = plt.subplot(3, 3, i + 1)
    plt.imshow(augmented_image[0])
    plt.axis("off")

• 更多的数据增强方式可以参考：www.tensorflow.org/api_docs/py…

三、增强方式

方法一：将其嵌入model中

这样做的好处是：

• 数据增强这块的工作可以得到GPU的加速（如果你使用了GPU训练的话） 注意：只有在模型训练时（Model.fit）才会进行增强，在模型评估(Model.evaluate)以及预测(Model.predict)时并不会进行增强操作。

model = tf.keras.Sequential([
  data_augmentation,
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
])

方法二：在Dataset数据集中进行数据增强

batch_size = 32
AUTOTUNE = tf.data.AUTOTUNE

def prepare(ds):
    ds = ds.map(lambda x, y: (data_augmentation(x, training=True), y), num_parallel_calls=AUTOTUNE)
    return ds

四、模型训练

在准备对模型进行训练之前，还需要再对其进行一些设置。以下内容是在模型的编译步骤中添加的：

• 损失函数（loss）：用于衡量模型在训练期间的准确率。
• 优化器（optimizer）：决定模型如何根据其看到的数据和自身的损失函数进行更新。
• 评价函数（metrics）：用于监控训练和测试步骤。以下示例使用了准确率，即被正确分类的图像的比率。

model = tf.keras.Sequential([
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(32, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(64, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Flatten(),
  layers.Dense(128, activation='relu'),
  layers.Dense(len(class_names))
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

## 开始训练
epochs=20
history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

Epoch 1/20
14/14 [==============================] - ETA: 0s - loss: 1.6685 - accuracy: 0.5476

14/14 [==============================] - 10s 103ms/step - loss: 1.6685 - accuracy: 0.5476 - val_loss: 0.6737 - val_accuracy: 0.5000
Epoch 2/20

14/14 [==============================] - 0s 22ms/step - loss: 0.5432 - accuracy: 0.8024 - val_loss: 0.3742 - val_accuracy: 0.8581
Epoch 3/20
14/14 [==============================] - 0s 18ms/step - loss: 0.2339 - accuracy: 0.9048 - val_loss: 0.4585 - val_accuracy: 0.8108
Epoch 4/20
14/14 [==============================] - 0s 18ms/step - loss: 0.1523 - accuracy: 0.9310 - val_loss: 0.3016 - val_accuracy: 0.8784
Epoch 5/20
14/14 [==============================] - 0s 21ms/step - loss: 0.0726 - accuracy: 0.9762 - val_loss: 0.2156 - val_accuracy: 0.9257
Epoch 6/20
14/14 [==============================] - 0s 18ms/step - loss: 0.0328 - accuracy: 0.9952 - val_loss: 0.2049 - val_accuracy: 0.9257
Epoch 7/20
14/14 [==============================] - 0s 18ms/step - loss: 0.0125 - accuracy: 1.0000 - val_loss: 0.3468 - val_accuracy: 0.9257
Epoch 8/20
14/14 [==============================] - 0s 18ms/step - loss: 0.0366 - accuracy: 0.9881 - val_loss: 0.4456 - val_accuracy: 0.8514
Epoch 9/20
14/14 [==============================] - 0s 20ms/step - loss: 0.0464 - accuracy: 0.9857 - val_loss: 0.1834 - val_accuracy: 0.9324
Epoch 10/20
14/14 [==============================] - 0s 19ms/step - loss: 0.0112 - accuracy: 0.9976 - val_loss: 0.2679 - val_accuracy: 0.9122
Epoch 11/20
14/14 [==============================] - 0s 18ms/step - loss: 0.0407 - accuracy: 0.9857 - val_loss: 0.3633 - val_accuracy: 0.9257
Epoch 12/20
14/14 [==============================] - 0s 22ms/step - loss: 0.0355 - accuracy: 0.9881 - val_loss: 0.3285 - val_accuracy: 0.9054
Epoch 13/20
14/14 [==============================] - 0s 20ms/step - loss: 0.0453 - accuracy: 0.9833 - val_loss: 0.2559 - val_accuracy: 0.9595
Epoch 14/20
14/14 [==============================] - 0s 22ms/step - loss: 0.0068 - accuracy: 1.0000 - val_loss: 0.8832 - val_accuracy: 0.8378
Epoch 15/20
14/14 [==============================] - 0s 25ms/step - loss: 0.0281 - accuracy: 0.9952 - val_loss: 0.7596 - val_accuracy: 0.8919
Epoch 16/20
14/14 [==============================] - 0s 24ms/step - loss: 0.1525 - accuracy: 0.9429 - val_loss: 0.3287 - val_accuracy: 0.9054
Epoch 17/20
14/14 [==============================] - 0s 27ms/step - loss: 0.0285 - accuracy: 0.9881 - val_loss: 0.3073 - val_accuracy: 0.9324
Epoch 18/20
14/14 [==============================] - 0s 28ms/step - loss: 0.0302 - accuracy: 0.9905 - val_loss: 0.3036 - val_accuracy: 0.9392
Epoch 19/20
14/14 [==============================] - 0s 18ms/step - loss: 5.4468e-04 - accuracy: 1.0000 - val_loss: 0.2788 - val_accuracy: 0.9392
Epoch 20/20
14/14 [==============================] - 0s 17ms/step - loss: 0.0012 - accuracy: 1.0000 - val_loss: 0.2933 - val_accuracy: 0.9392

loss, acc = model.evaluate(test_ds)
print("Accuracy", acc)

1/1 [==============================] - 0s 196ms/step - loss: 0.3975 - accuracy: 0.9062
Accuracy 0.90625

五、自定义增强函数

import random
# 这是大家可以自由发挥的一个地方
def aug_img(image):
    seed = (random.randint(0,9), 0)
    # 随机改变图像对比度
    stateless_random_brightness = tf.image.stateless_random_contrast(image, lower=0.1, upper=1.0, seed=seed)
    return stateless_random_brightness

image = tf.expand_dims(images[3]*255, 0)
print("Min and max pixel values:", image.numpy().min(), image.numpy().max())

Min and max pixel values: 14.000048 253.28577

plt.figure(figsize=(8, 8))
for i in range(9):
    augmented_image = aug_img(image)
    ax = plt.subplot(3, 3, i + 1)
    plt.imshow(augmented_image[0].numpy().astype("uint8"))

    plt.axis("off")

• 不进行数据增强

gpus = tf.config.list_physical_devices("GPU")
tf.config.experimental.set_memory_growth(gpus[0], True)  
tf.config.set_visible_devices([gpus[0]],"GPU")

data_dir   = "./data/"
img_height = 224
img_width  = 224
batch_size = 32

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="training",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="validation",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

val_batches = tf.data.experimental.cardinality(val_ds)
test_ds     = val_ds.take(val_batches // 5)
val_ds      = val_ds.skip(val_batches // 5)

class_names = train_ds.class_names
print(class_names)

AUTOTUNE = tf.data.AUTOTUNE

## 数据预处理时进行数据增强
train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds   = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

model = tf.keras.Sequential([
  #tf.keras.layers.experimental.preprocessing.RandomFlip("horizontal_and_vertical"),
  #tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(32, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(64, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Flatten(),
  layers.Dense(128, activation='relu'),
  layers.Dense(len(class_names))
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

epochs=60
history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

loss, acc = model.evaluate(test_ds)
print("Accuracy", acc)

Found 600 files belonging to 2 classes.
Using 420 files for training.
Found 600 files belonging to 2 classes.
Using 180 files for validation.
['cat', 'dog']
Epoch 1/60
14/14 [==============================] - 2s 32ms/step - loss: 111.9658 - accuracy: 0.5310 - val_loss: 7.4655 - val_accuracy: 0.5878
Epoch 2/60
14/14 [==============================] - 0s 16ms/step - loss: 1.3849 - accuracy: 0.8333 - val_loss: 0.7918 - val_accuracy: 0.8446
Epoch 3/60
14/14 [==============================] - 0s 16ms/step - loss: 0.1112 - accuracy: 0.9595 - val_loss: 0.3117 - val_accuracy: 0.9459
Epoch 4/60
14/14 [==============================] - 0s 19ms/step - loss: 0.0172 - accuracy: 0.9976 - val_loss: 0.3034 - val_accuracy: 0.9324
Epoch 5/60
14/14 [==============================] - 0s 20ms/step - loss: 0.0106 - accuracy: 0.9952 - val_loss: 0.3100 - val_accuracy: 0.9324
Epoch 6/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0035 - accuracy: 1.0000 - val_loss: 0.3177 - val_accuracy: 0.9459
Epoch 7/60
14/14 [==============================] - 0s 24ms/step - loss: 0.0011 - accuracy: 1.0000 - val_loss: 0.3153 - val_accuracy: 0.9527
Epoch 8/60
14/14 [==============================] - 0s 19ms/step - loss: 4.8140e-04 - accuracy: 1.0000 - val_loss: 0.3062 - val_accuracy: 0.9392
Epoch 9/60
14/14 [==============================] - 0s 21ms/step - loss: 2.0318e-04 - accuracy: 1.0000 - val_loss: 0.3092 - val_accuracy: 0.9324
Epoch 10/60
14/14 [==============================] - 0s 22ms/step - loss: 1.0784e-04 - accuracy: 1.0000 - val_loss: 0.3131 - val_accuracy: 0.9392
Epoch 11/60
14/14 [==============================] - 0s 18ms/step - loss: 5.5197e-05 - accuracy: 1.0000 - val_loss: 0.3271 - val_accuracy: 0.9324
Epoch 12/60
14/14 [==============================] - 0s 17ms/step - loss: 2.7965e-05 - accuracy: 1.0000 - val_loss: 0.3446 - val_accuracy: 0.9324
Epoch 13/60
14/14 [==============================] - 0s 17ms/step - loss: 1.6096e-05 - accuracy: 1.0000 - val_loss: 0.3575 - val_accuracy: 0.9324
Epoch 14/60
14/14 [==============================] - 0s 18ms/step - loss: 1.0265e-05 - accuracy: 1.0000 - val_loss: 0.3668 - val_accuracy: 0.9392
Epoch 15/60
14/14 [==============================] - 0s 16ms/step - loss: 7.2758e-06 - accuracy: 1.0000 - val_loss: 0.3723 - val_accuracy: 0.9392
Epoch 16/60
14/14 [==============================] - 0s 19ms/step - loss: 5.4888e-06 - accuracy: 1.0000 - val_loss: 0.3763 - val_accuracy: 0.9392
Epoch 17/60
14/14 [==============================] - 0s 29ms/step - loss: 4.3794e-06 - accuracy: 1.0000 - val_loss: 0.3800 - val_accuracy: 0.9392
Epoch 18/60
14/14 [==============================] - 0s 28ms/step - loss: 3.6315e-06 - accuracy: 1.0000 - val_loss: 0.3838 - val_accuracy: 0.9392
Epoch 19/60
14/14 [==============================] - 0s 22ms/step - loss: 3.0701e-06 - accuracy: 1.0000 - val_loss: 0.3871 - val_accuracy: 0.9392
Epoch 20/60
14/14 [==============================] - 0s 21ms/step - loss: 2.6617e-06 - accuracy: 1.0000 - val_loss: 0.3899 - val_accuracy: 0.9392
Epoch 21/60
14/14 [==============================] - 0s 24ms/step - loss: 2.3401e-06 - accuracy: 1.0000 - val_loss: 0.3924 - val_accuracy: 0.9392
Epoch 22/60
14/14 [==============================] - 0s 23ms/step - loss: 2.0787e-06 - accuracy: 1.0000 - val_loss: 0.3951 - val_accuracy: 0.9392
Epoch 23/60
14/14 [==============================] - 0s 21ms/step - loss: 1.8628e-06 - accuracy: 1.0000 - val_loss: 0.3967 - val_accuracy: 0.9392
Epoch 24/60
14/14 [==============================] - 0s 17ms/step - loss: 1.6760e-06 - accuracy: 1.0000 - val_loss: 0.3987 - val_accuracy: 0.9392
Epoch 25/60
14/14 [==============================] - 0s 17ms/step - loss: 1.5176e-06 - accuracy: 1.0000 - val_loss: 0.4008 - val_accuracy: 0.9392
Epoch 26/60
14/14 [==============================] - 0s 20ms/step - loss: 1.3814e-06 - accuracy: 1.0000 - val_loss: 0.4023 - val_accuracy: 0.9392
Epoch 27/60
14/14 [==============================] - 0s 20ms/step - loss: 1.2576e-06 - accuracy: 1.0000 - val_loss: 0.4039 - val_accuracy: 0.9392
Epoch 28/60
14/14 [==============================] - 0s 19ms/step - loss: 1.1518e-06 - accuracy: 1.0000 - val_loss: 0.4052 - val_accuracy: 0.9392
Epoch 29/60
14/14 [==============================] - 0s 19ms/step - loss: 1.0604e-06 - accuracy: 1.0000 - val_loss: 0.4065 - val_accuracy: 0.9392
Epoch 30/60
14/14 [==============================] - 0s 21ms/step - loss: 9.8091e-07 - accuracy: 1.0000 - val_loss: 0.4077 - val_accuracy: 0.9392
Epoch 31/60
14/14 [==============================] - 0s 18ms/step - loss: 9.0797e-07 - accuracy: 1.0000 - val_loss: 0.4090 - val_accuracy: 0.9392
Epoch 32/60
14/14 [==============================] - 0s 17ms/step - loss: 8.4297e-07 - accuracy: 1.0000 - val_loss: 0.4100 - val_accuracy: 0.9392
Epoch 33/60
14/14 [==============================] - 0s 17ms/step - loss: 7.8309e-07 - accuracy: 1.0000 - val_loss: 0.4112 - val_accuracy: 0.9392
Epoch 34/60
14/14 [==============================] - 0s 16ms/step - loss: 7.3143e-07 - accuracy: 1.0000 - val_loss: 0.4125 - val_accuracy: 0.9392
Epoch 35/60
14/14 [==============================] - 0s 22ms/step - loss: 6.8233e-07 - accuracy: 1.0000 - val_loss: 0.4135 - val_accuracy: 0.9392
Epoch 36/60
14/14 [==============================] - 0s 17ms/step - loss: 6.3862e-07 - accuracy: 1.0000 - val_loss: 0.4143 - val_accuracy: 0.9392
Epoch 37/60
14/14 [==============================] - 0s 17ms/step - loss: 6.0030e-07 - accuracy: 1.0000 - val_loss: 0.4151 - val_accuracy: 0.9392
Epoch 38/60
14/14 [==============================] - 0s 19ms/step - loss: 5.6596e-07 - accuracy: 1.0000 - val_loss: 0.4160 - val_accuracy: 0.9392
Epoch 39/60
14/14 [==============================] - 0s 18ms/step - loss: 5.3076e-07 - accuracy: 1.0000 - val_loss: 0.4161 - val_accuracy: 0.9392
Epoch 40/60
14/14 [==============================] - 0s 19ms/step - loss: 4.9784e-07 - accuracy: 1.0000 - val_loss: 0.4165 - val_accuracy: 0.9392
Epoch 41/60
14/14 [==============================] - 0s 16ms/step - loss: 4.6917e-07 - accuracy: 1.0000 - val_loss: 0.4170 - val_accuracy: 0.9392
Epoch 42/60
14/14 [==============================] - 0s 16ms/step - loss: 4.4476e-07 - accuracy: 1.0000 - val_loss: 0.4174 - val_accuracy: 0.9392
Epoch 43/60
14/14 [==============================] - 0s 18ms/step - loss: 4.1808e-07 - accuracy: 1.0000 - val_loss: 0.4174 - val_accuracy: 0.9392
Epoch 44/60
14/14 [==============================] - 0s 16ms/step - loss: 3.9765e-07 - accuracy: 1.0000 - val_loss: 0.4178 - val_accuracy: 0.9392
Epoch 45/60
14/14 [==============================] - 0s 15ms/step - loss: 3.7977e-07 - accuracy: 1.0000 - val_loss: 0.4182 - val_accuracy: 0.9392
Epoch 46/60
14/14 [==============================] - 0s 15ms/step - loss: 3.6075e-07 - accuracy: 1.0000 - val_loss: 0.4183 - val_accuracy: 0.9392
Epoch 47/60
14/14 [==============================] - 0s 17ms/step - loss: 3.4514e-07 - accuracy: 1.0000 - val_loss: 0.4184 - val_accuracy: 0.9392
Epoch 48/60
14/14 [==============================] - 0s 17ms/step - loss: 3.2924e-07 - accuracy: 1.0000 - val_loss: 0.4186 - val_accuracy: 0.9392
Epoch 49/60
14/14 [==============================] - 0s 16ms/step - loss: 3.1505e-07 - accuracy: 1.0000 - val_loss: 0.4187 - val_accuracy: 0.9392
Epoch 50/60
14/14 [==============================] - 0s 20ms/step - loss: 3.0058e-07 - accuracy: 1.0000 - val_loss: 0.4189 - val_accuracy: 0.9392
Epoch 51/60
14/14 [==============================] - 0s 25ms/step - loss: 2.8951e-07 - accuracy: 1.0000 - val_loss: 0.4187 - val_accuracy: 0.9392
Epoch 52/60
14/14 [==============================] - 0s 27ms/step - loss: 2.7730e-07 - accuracy: 1.0000 - val_loss: 0.4188 - val_accuracy: 0.9392
Epoch 53/60
14/14 [==============================] - 0s 22ms/step - loss: 2.6680e-07 - accuracy: 1.0000 - val_loss: 0.4189 - val_accuracy: 0.9392
Epoch 54/60
14/14 [==============================] - 0s 20ms/step - loss: 2.5743e-07 - accuracy: 1.0000 - val_loss: 0.4190 - val_accuracy: 0.9392
Epoch 55/60
14/14 [==============================] - 0s 21ms/step - loss: 2.4807e-07 - accuracy: 1.0000 - val_loss: 0.4191 - val_accuracy: 0.9392
Epoch 56/60
14/14 [==============================] - 0s 19ms/step - loss: 2.4012e-07 - accuracy: 1.0000 - val_loss: 0.4192 - val_accuracy: 0.9392
Epoch 57/60
14/14 [==============================] - 0s 16ms/step - loss: 2.3132e-07 - accuracy: 1.0000 - val_loss: 0.4194 - val_accuracy: 0.9392
Epoch 58/60
14/14 [==============================] - 0s 19ms/step - loss: 2.2196e-07 - accuracy: 1.0000 - val_loss: 0.4197 - val_accuracy: 0.9392
Epoch 59/60
14/14 [==============================] - 0s 23ms/step - loss: 2.1372e-07 - accuracy: 1.0000 - val_loss: 0.4198 - val_accuracy: 0.9392
Epoch 60/60
14/14 [==============================] - 0s 21ms/step - loss: 2.0833e-07 - accuracy: 1.0000 - val_loss: 0.4199 - val_accuracy: 0.9392
1/1 [==============================] - 0s 194ms/step - loss: 0.7822 - accuracy: 0.8438
Accuracy 0.84375

• 那么如何将自定义增强函数应用到我们数据上呢？请参考上文的 preprocess_image 函数，将 aug_img 函数嵌入到 preprocess_image 函数中，在数据预处理时完成数据增强就OK啦。

gpus = tf.config.list_physical_devices("GPU")
tf.config.experimental.set_memory_growth(gpus[0], True)  
tf.config.set_visible_devices([gpus[0]],"GPU")

data_dir   = "./data/"
img_height = 224
img_width  = 224
batch_size = 32

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="training",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.3,
    subset="validation",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

val_batches = tf.data.experimental.cardinality(val_ds)
test_ds     = val_ds.take(val_batches // 5)
val_ds      = val_ds.skip(val_batches // 5)

class_names = train_ds.class_names
print(class_names)

AUTOTUNE = tf.data.AUTOTUNE

import random
def preprocess_image2(image, label):
    seed = (random.randint(0,9), 0)
    # 随机改变图像对比度
    image2 = tf.image.stateless_random_contrast(image, lower=0.1, upper=1.0, seed=seed)
    return (image2/255.0,label)

## 数据预处理时进行数据增强
train_ds = train_ds.map(preprocess_image2, num_parallel_calls=AUTOTUNE)
val_ds   = val_ds.map(preprocess_image2, num_parallel_calls=AUTOTUNE)
test_ds  = test_ds.map(preprocess_image2, num_parallel_calls=AUTOTUNE)

train_ds = train_ds.cache().prefetch(buffer_size=AUTOTUNE)
val_ds   = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

model = tf.keras.Sequential([
  #tf.keras.layers.experimental.preprocessing.RandomFlip("horizontal_and_vertical"),
  #tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(32, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(64, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Flatten(),
  layers.Dense(128, activation='relu'),
  layers.Dense(len(class_names))
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

epochs=60
history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

loss, acc = model.evaluate(test_ds)
print("Accuracy", acc)

Found 600 files belonging to 2 classes.
Using 420 files for training.
Found 600 files belonging to 2 classes.
Using 180 files for validation.
['cat', 'dog']
Epoch 1/60
14/14 [==============================] - 2s 44ms/step - loss: 1.1446 - accuracy: 0.5310 - val_loss: 0.6992 - val_accuracy: 0.4257
Epoch 2/60
14/14 [==============================] - 0s 17ms/step - loss: 0.6912 - accuracy: 0.5262 - val_loss: 0.7037 - val_accuracy: 0.4257
Epoch 3/60
14/14 [==============================] - 0s 15ms/step - loss: 0.6873 - accuracy: 0.5262 - val_loss: 0.6813 - val_accuracy: 0.5473
Epoch 4/60
14/14 [==============================] - 0s 18ms/step - loss: 0.6843 - accuracy: 0.5476 - val_loss: 0.6579 - val_accuracy: 0.6689
Epoch 5/60
14/14 [==============================] - 0s 19ms/step - loss: 0.6627 - accuracy: 0.6548 - val_loss: 0.6087 - val_accuracy: 0.7095
Epoch 6/60
14/14 [==============================] - 0s 17ms/step - loss: 0.6365 - accuracy: 0.6310 - val_loss: 0.6015 - val_accuracy: 0.6622
Epoch 7/60
14/14 [==============================] - 0s 16ms/step - loss: 0.5804 - accuracy: 0.7048 - val_loss: 0.5760 - val_accuracy: 0.6824
Epoch 8/60
14/14 [==============================] - 0s 17ms/step - loss: 0.5789 - accuracy: 0.6952 - val_loss: 0.4624 - val_accuracy: 0.7838
Epoch 9/60
14/14 [==============================] - 0s 17ms/step - loss: 0.5092 - accuracy: 0.7381 - val_loss: 0.7293 - val_accuracy: 0.6689
Epoch 10/60
14/14 [==============================] - 0s 18ms/step - loss: 0.4975 - accuracy: 0.7333 - val_loss: 0.4033 - val_accuracy: 0.7973
Epoch 11/60
14/14 [==============================] - 0s 17ms/step - loss: 0.4157 - accuracy: 0.8024 - val_loss: 0.5234 - val_accuracy: 0.8041
Epoch 12/60
14/14 [==============================] - 0s 19ms/step - loss: 0.3645 - accuracy: 0.8643 - val_loss: 0.4355 - val_accuracy: 0.8311
Epoch 13/60
14/14 [==============================] - 0s 20ms/step - loss: 0.3208 - accuracy: 0.8643 - val_loss: 0.6993 - val_accuracy: 0.7838
Epoch 14/60
14/14 [==============================] - 0s 17ms/step - loss: 0.3187 - accuracy: 0.8619 - val_loss: 0.5579 - val_accuracy: 0.8041
Epoch 15/60
14/14 [==============================] - 0s 17ms/step - loss: 0.3037 - accuracy: 0.8738 - val_loss: 0.4870 - val_accuracy: 0.8243
Epoch 16/60
14/14 [==============================] - 0s 18ms/step - loss: 0.2444 - accuracy: 0.9000 - val_loss: 0.4687 - val_accuracy: 0.8514
Epoch 17/60
14/14 [==============================] - 0s 18ms/step - loss: 0.2139 - accuracy: 0.9119 - val_loss: 0.4536 - val_accuracy: 0.8649
Epoch 18/60
14/14 [==============================] - 0s 19ms/step - loss: 0.2019 - accuracy: 0.9167 - val_loss: 0.4112 - val_accuracy: 0.8784
Epoch 19/60
14/14 [==============================] - 0s 18ms/step - loss: 0.1865 - accuracy: 0.9167 - val_loss: 0.4651 - val_accuracy: 0.8851
Epoch 20/60
14/14 [==============================] - 0s 19ms/step - loss: 0.1785 - accuracy: 0.9405 - val_loss: 0.4609 - val_accuracy: 0.8784
Epoch 21/60
14/14 [==============================] - 0s 17ms/step - loss: 0.1371 - accuracy: 0.9524 - val_loss: 0.5768 - val_accuracy: 0.8649
Epoch 22/60
14/14 [==============================] - 0s 16ms/step - loss: 0.1865 - accuracy: 0.9262 - val_loss: 0.5060 - val_accuracy: 0.8784
Epoch 23/60
14/14 [==============================] - 0s 18ms/step - loss: 0.1977 - accuracy: 0.9214 - val_loss: 0.4349 - val_accuracy: 0.8986
Epoch 24/60
14/14 [==============================] - 0s 22ms/step - loss: 0.1337 - accuracy: 0.9548 - val_loss: 0.3162 - val_accuracy: 0.9122
Epoch 25/60
14/14 [==============================] - 0s 18ms/step - loss: 0.1153 - accuracy: 0.9595 - val_loss: 0.6187 - val_accuracy: 0.8784
Epoch 26/60
14/14 [==============================] - 0s 16ms/step - loss: 0.1268 - accuracy: 0.9571 - val_loss: 0.4181 - val_accuracy: 0.8986
Epoch 27/60
14/14 [==============================] - 0s 16ms/step - loss: 0.1157 - accuracy: 0.9571 - val_loss: 0.4264 - val_accuracy: 0.9122
Epoch 28/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0863 - accuracy: 0.9762 - val_loss: 0.4944 - val_accuracy: 0.8919
Epoch 29/60
14/14 [==============================] - 0s 22ms/step - loss: 0.1081 - accuracy: 0.9595 - val_loss: 0.5845 - val_accuracy: 0.8784
Epoch 30/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0815 - accuracy: 0.9786 - val_loss: 0.5020 - val_accuracy: 0.8919
Epoch 31/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0705 - accuracy: 0.9762 - val_loss: 0.4553 - val_accuracy: 0.9054
Epoch 32/60
14/14 [==============================] - 0s 19ms/step - loss: 0.0653 - accuracy: 0.9833 - val_loss: 0.6548 - val_accuracy: 0.8851
Epoch 33/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0862 - accuracy: 0.9667 - val_loss: 0.6245 - val_accuracy: 0.8851
Epoch 34/60
14/14 [==============================] - 0s 17ms/step - loss: 0.1219 - accuracy: 0.9714 - val_loss: 0.3035 - val_accuracy: 0.9459
Epoch 35/60
14/14 [==============================] - 0s 16ms/step - loss: 0.1272 - accuracy: 0.9476 - val_loss: 0.6282 - val_accuracy: 0.8649
Epoch 36/60
14/14 [==============================] - 0s 20ms/step - loss: 0.1104 - accuracy: 0.9524 - val_loss: 0.3613 - val_accuracy: 0.9257
Epoch 37/60
14/14 [==============================] - 0s 20ms/step - loss: 0.0942 - accuracy: 0.9643 - val_loss: 0.7265 - val_accuracy: 0.8649
Epoch 38/60
14/14 [==============================] - 0s 17ms/step - loss: 0.0938 - accuracy: 0.9738 - val_loss: 0.5351 - val_accuracy: 0.9122
Epoch 39/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0351 - accuracy: 0.9881 - val_loss: 0.4451 - val_accuracy: 0.9257
Epoch 40/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0211 - accuracy: 0.9952 - val_loss: 0.7861 - val_accuracy: 0.8784
Epoch 41/60
14/14 [==============================] - 0s 19ms/step - loss: 0.0434 - accuracy: 0.9833 - val_loss: 0.6123 - val_accuracy: 0.9122
Epoch 42/60
14/14 [==============================] - 0s 17ms/step - loss: 0.0545 - accuracy: 0.9857 - val_loss: 0.9089 - val_accuracy: 0.8716
Epoch 43/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0742 - accuracy: 0.9738 - val_loss: 0.6080 - val_accuracy: 0.8784
Epoch 44/60
14/14 [==============================] - 0s 25ms/step - loss: 0.0232 - accuracy: 0.9952 - val_loss: 0.4983 - val_accuracy: 0.9054
Epoch 45/60
14/14 [==============================] - 0s 25ms/step - loss: 0.0149 - accuracy: 1.0000 - val_loss: 0.6019 - val_accuracy: 0.9054
Epoch 46/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0227 - accuracy: 0.9952 - val_loss: 0.9803 - val_accuracy: 0.8784
Epoch 47/60
14/14 [==============================] - 0s 20ms/step - loss: 0.0698 - accuracy: 0.9714 - val_loss: 0.4859 - val_accuracy: 0.9054
Epoch 48/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0480 - accuracy: 0.9857 - val_loss: 0.4052 - val_accuracy: 0.9459
Epoch 49/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0350 - accuracy: 0.9905 - val_loss: 0.4674 - val_accuracy: 0.9122
Epoch 50/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0193 - accuracy: 0.9929 - val_loss: 0.3538 - val_accuracy: 0.9392
Epoch 51/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0137 - accuracy: 0.9976 - val_loss: 0.3533 - val_accuracy: 0.9527
Epoch 52/60
14/14 [==============================] - 0s 20ms/step - loss: 0.0105 - accuracy: 1.0000 - val_loss: 0.3799 - val_accuracy: 0.9392
Epoch 53/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0117 - accuracy: 0.9976 - val_loss: 0.4457 - val_accuracy: 0.9392
Epoch 54/60
14/14 [==============================] - 0s 19ms/step - loss: 0.0069 - accuracy: 1.0000 - val_loss: 0.5769 - val_accuracy: 0.9189
Epoch 55/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0033 - accuracy: 1.0000 - val_loss: 0.4502 - val_accuracy: 0.9392
Epoch 56/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0030 - accuracy: 1.0000 - val_loss: 0.4593 - val_accuracy: 0.9257
Epoch 57/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0019 - accuracy: 1.0000 - val_loss: 0.4775 - val_accuracy: 0.9324
Epoch 58/60
14/14 [==============================] - 0s 21ms/step - loss: 0.0018 - accuracy: 1.0000 - val_loss: 0.5305 - val_accuracy: 0.9257
Epoch 59/60
14/14 [==============================] - 0s 18ms/step - loss: 0.0014 - accuracy: 1.0000 - val_loss: 0.5102 - val_accuracy: 0.9324
Epoch 60/60
14/14 [==============================] - 0s 16ms/step - loss: 0.0014 - accuracy: 1.0000 - val_loss: 0.5365 - val_accuracy: 0.9257
1/1 [==============================] - 0s 152ms/step - loss: 0.1707 - accuracy: 0.9375
Accuracy 0.9375

六、总结

• 在数据预处理时完成了数据增强
• 使用数据增强后，可以适当增加 Epoch 数，提高模型测试准确性。