InceptionResNetV2模型传递学习后验证精度低

Question

我需要一个Tensorflow模型来将图像分类为4种不同的类别，我正在对预先训练的InceptionResNetV2模型进行传递学习(权重=‘Imagenet’)。在model.fit()期间，我的准确率为97.4%，损失为0.3，而我的验证精度保持在84%，损失为0.4。我是否过度适应，如何提高我的验证准确性？

base_model = InceptionResNetV2(input_shape=(299,299,3),weights='imagenet', include_top=False)


x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(2048,activation='relu')(x)
x = Dropout(0.03)(x)
x = LeakyReLU(alpha=0.05)(x)
x = Dense(1024,activation='relu')(x)
x = LeakyReLU(alpha=0.05)(x)
x = Dense(128,activation='relu')(x)

predictions = Dense(4, activation='softmax')(x)

Answer 1

尝试将数据增强添加到代码中；这将解决过度匹配的问题。如下代码所示：

data_augmentation = tf.keras.Sequential([
    tf.keras.layers.experimental.preprocessing.RandomRotation(0.01),
    tf.keras.layers.experimental.preprocessing.RandomContrast(0.25),
    tf.keras.layers.experimental.preprocessing.RandomTranslation(height_factor = (-0.1, 0.1), width_factor = (-0.1, 0.1))])

inp = tf.keras.Input(shape=IMG_SIZE)
x = data_augmentation(inp)
x = base_model(x, training=False)