使用 Tensorflow 在 CNN 中进行微调 - 2.0

Question

我目前正在研究太阳能电池板的缺陷分类问题。这是一个多类分类问题。目前它的3类。我已经完成了编码部分，但我的准确性非常低。如何提高我的准确性？

总训练图像 - 900 测试/验证 - 300 类 - 3

我的代码如下 -

import tensorflow as tf
import keras_preprocessing
from keras_preprocessing import image
from keras_preprocessing.image import ImageDataGenerator

TRAINING_DIR = "/content/drive/My Drive/solar_images/solar_images/train/"
training_datagen = ImageDataGenerator(
      rescale = 1./255,
        rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')

VALIDATION_DIR = "/content/drive/My Drive/solar_images/solar_images/test/"
validation_datagen = ImageDataGenerator(rescale = 1./255)

train_generator = training_datagen.flow_from_directory(
    TRAINING_DIR,
    target_size=(150,150),
    class_mode='categorical',
  batch_size=64
)

validation_generator = validation_datagen.flow_from_directory(
    VALIDATION_DIR,
    target_size=(150,150),
    class_mode='categorical',
  batch_size=64
)

model = tf.keras.models.Sequential([
    # Note the input shape is the desired size of the image 150x150 with 3 bytes color
    # This is the first convolution
    tf.keras.layers.Conv2D(64, (3,3), activation='relu', input_shape=(150, 150, 3)),
    tf.keras.layers.MaxPooling2D(2, 2),
    # The second convolution
    tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),
    # The third convolution
    tf.keras.layers.Conv2D(128, (3,3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),
    # The fourth convolution
    tf.keras.layers.Conv2D(128, (3,3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),
    # Flatten the results to feed into a DNN
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dropout(0.5),
    # 512 neuron hidden layer
    tf.keras.layers.Dense(512, activation='relu'),
    tf.keras.layers.Dense(3, activation='softmax')
])


model.summary()

model.compile(loss = 'categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
batch_size=64

history = model.fit(train_generator, 
                    epochs=20, 
                    steps_per_epoch=int(894/batch_size), 
                    validation_data = validation_generator, 
                    verbose = 1, 
                    validation_steps=int(289/batch_size))

model.save("solar_images_weight.h5")

我的准确度是 -

Epoch 1/20
13/13 [==============================] - 1107s 85s/step - loss: 1.2893 - accuracy: 0.3470 - val_loss: 1.0926 - val_accuracy: 0.3594
Epoch 2/20
13/13 [==============================] - 1239s 95s/step - loss: 1.1037 - accuracy: 0.3566 - val_loss: 1.0954 - val_accuracy: 0.3125
Epoch 3/20
13/13 [==============================] - 1203s 93s/step - loss: 1.0964 - accuracy: 0.3904 - val_loss: 1.0841 - val_accuracy: 0.5625
Epoch 4/20
13/13 [==============================] - 1182s 91s/step - loss: 1.0980 - accuracy: 0.3750 - val_loss: 1.0894 - val_accuracy: 0.3633
Epoch 5/20
13/13 [==============================] - 1218s 94s/step - loss: 1.1086 - accuracy: 0.3386 - val_loss: 1.0874 - val_accuracy: 0.3125
Epoch 6/20
13/13 [==============================] - 1214s 93s/step - loss: 1.0953 - accuracy: 0.3257 - val_loss: 1.0763 - val_accuracy: 0.6094
Epoch 7/20
13/13 [==============================] - 1136s 87s/step - loss: 1.0851 - accuracy: 0.3831 - val_loss: 1.0754 - val_accuracy: 0.3164
Epoch 8/20
13/13 [==============================] - 1170s 90s/step - loss: 1.1005 - accuracy: 0.3940 - val_loss: 1.0545 - val_accuracy: 0.5039
Epoch 9/20
13/13 [==============================] - 1138s 88s/step - loss: 1.1294 - accuracy: 0.4337 - val_loss: 1.0130 - val_accuracy: 0.5703
Epoch 10/20
13/13 [==============================] - 1131s 87s/step - loss: 1.0250 - accuracy: 0.4531 - val_loss: 0.8911 - val_accuracy: 0.6055
Epoch 11/20
13/13 [==============================] - 1162s 89s/step - loss: 1.0243 - accuracy: 0.4735 - val_loss: 0.9160 - val_accuracy: 0.4727
Epoch 12/20
13/13 [==============================] - 1153s 89s/step - loss: 0.9978 - accuracy: 0.4783 - val_loss: 0.7754 - val_accuracy: 0.6406
Epoch 13/20
13/13 [==============================] - 1187s 91s/step - loss: 1.0080 - accuracy: 0.4687 - val_loss: 0.7701 - val_accuracy: 0.6602
Epoch 14/20
13/13 [==============================] - 1204s 93s/step - loss: 0.9851 - accuracy: 0.5048 - val_loss: 0.7450 - val_accuracy: 0.6367
Epoch 15/20
13/13 [==============================] - 1181s 91s/step - loss: 0.9699 - accuracy: 0.4892 - val_loss: 0.7409 - val_accuracy: 0.6289
Epoch 16/20
13/13 [==============================] - 1187s 91s/step - loss: 0.8884 - accuracy: 0.5241 - val_loss: 0.7169 - val_accuracy: 0.6133
Epoch 17/20
13/13 [==============================] - 1197s 92s/step - loss: 0.9372 - accuracy: 0.5084 - val_loss: 0.7464 - val_accuracy: 0.5859
Epoch 18/20
13/13 [==============================] - 1224s 94s/step - loss: 0.9230 - accuracy: 0.5229 - val_loss: 0.9198 - val_accuracy: 0.5156
Epoch 19/20
13/13 [==============================] - 1270s 98s/step - loss: 0.9161 - accuracy: 0.5192 - val_loss: 0.6785 - val_accuracy: 0.6289
Epoch 20/20
13/13 [==============================] - 1173s 90s/step - loss: 0.8728 - accuracy: 0.5193 - val_loss: 0.6674 - val_accuracy: 0.5781

训练和验证准确度图如下 -

Answer 1

您可以使用迁移学习。使用预训练模型（例如 mobilenet 或 inception）在数据集上进行训练。这将显着提高您的准确性。