python - 为什么 Keras 中的 resnet50 模型不收敛？

Question

我目前正在尝试对缺陷和非缺陷图像中的集成电路进行分类。我已经尝试过 VGG16 和 InceptionV3，并且都获得了非常好的结果(95% 的验证准确度和较低的 val 损失)。现在我想尝试 resnet50 但我的模型没有收敛。它的准确率也为 95%，但验证损失不断增加，而 val acc 则停留在 50%。

这是我到目前为止的脚本:

from keras.applications.resnet50 import ResNet50
from keras.optimizers import Adam
from keras.preprocessing import image
from keras.models import Model
from keras.layers import Dense, GlobalAveragePooling2D, Dropout
from keras import backend as K
from keras_preprocessing.image import ImageDataGenerator
import tensorflow as tf

class ResNet:
    def __init__(self):
        self.img_width, self.img_height = 224, 224  # Dimensions of cropped image
        self.classes_num = 2  # Number of classifications

        # Training configurations
        self.epochs = 32
        self.batch_size = 16  # Play with this to determine number of images to train on per epoch
        self.lr = 0.0001

    def build_model(self, train_path):
        train_data_path = train_path
        train_datagen = ImageDataGenerator(rescale=1. / 255, validation_split=0.25)

        train_generator = train_datagen.flow_from_directory(
            train_data_path,
            target_size=(self.img_height, self.img_width),
            color_mode="rgb",
            batch_size=self.batch_size,
            class_mode='categorical',
            subset='training')

        validation_generator = train_datagen.flow_from_directory(
            train_data_path,
            target_size=(self.img_height, self.img_width),
            color_mode="rgb",
            batch_size=self.batch_size,
            class_mode='categorical',
            subset='validation')

        # create the base pre-trained model
        base_model = ResNet50(weights='imagenet', include_top=False, input_shape=    (self.img_height, self.img_width, 3))

        # add a global spatial average pooling layer
        x = base_model.output
        x = GlobalAveragePooling2D()(x)
        # let's add a fully-connected layer
        x = Dense(1024, activation='relu')(x)
        #x = Dropout(0.3)(x)
        # and a logistic layer -- let's say we have 200 classes
        predictions = Dense(2, activation='softmax')(x)

        # this is the model we will train
        model = Model(inputs=base_model.input, outputs=predictions)

        # first: train only the top layers (which were randomly initialized)
        # i.e. freeze all convolutional InceptionV3 layers
        for layer in base_model.layers:
            layer.trainable = True

        # compile the model (should be done *after* setting layers to non-trainable)
        opt = Adam(self.lr)  # , decay=self.INIT_LR / self.NUM_EPOCHS)
        model.compile(opt, loss='binary_crossentropy', metrics=["accuracy"])

        # train the model on the new data for a few epochs
        from keras.callbacks import ModelCheckpoint, EarlyStopping
        import matplotlib.pyplot as plt

        checkpoint = ModelCheckpoint('resnetModel.h5', monitor='val_accuracy', verbose=1, save_best_only=True,
                                 save_weights_only=False, mode='auto', period=1)

        early = EarlyStopping(monitor='val_accuracy', min_delta=0, patience=16, verbose=1, mode='auto')
        hist = model.fit_generator(steps_per_epoch=self.batch_size, generator=train_generator,
                               validation_data=validation_generator, validation_steps=self.batch_size, epochs=self.epochs,
                               callbacks=[checkpoint, early])

        plt.plot(hist.history['accuracy'])
        plt.plot(hist.history['val_accuracy'])
        plt.plot(hist.history['loss'])
        plt.plot(hist.history['val_loss'])
        plt.title("model accuracy")
        plt.ylabel("Accuracy")
        plt.xlabel("Epoch")
        plt.legend(["Accuracy", "Validation Accuracy", "loss", "Validation Loss"])
        plt.show()

        plt.figure(1)

import tensorflow as tf

if __name__ == '__main__':
    x = ResNet()
    config = tf.compat.v1.ConfigProto()
    config.gpu_options.allow_growth = True
    sess = tf.compat.v1.Session(config=config)
    x.build_model("C:/Users/but/Desktop/dataScratch/Train")

这就是模型的训练

resnet 失败但 vgg 和 inception 正常工作的原因可能是什么？ 我的脚本有错误吗？

Answer 1

至少对于代码来说，我没有发现任何可能影响训练过程的错误。

# and a logistic layer -- let's say we have 200 classes
predictions = Dense(2, activation='softmax')(x)

这些行有点可疑。不过评论好像有错别字，应该没问题。

# first: train only the top layers (which were randomly initialized)
# i.e. freeze all convolutional InceptionV3 layers
for layer in base_model.layers:
    layer.trainable = True

这些也很可疑。如果你想卡住ResNet-50的层，你需要做的是

...
base_model = ResNet50(weights='imagenet', include_top=False, input_shape=(self.img_height, self.img_width, 3))
for layer in base_model.layers:
    layer.trainable = False
...

但事实证明，layer.trainable = True实际上是你的意图，所以这也没关系。

首先，如果您使用与训练 VGG16 和 Inception V3 相同的代码，则该代码不太可能是问题所在。

为什么不检查以下易受影响的原因？

• 模型可能太小/太大，导致欠拟合/过拟合。 (参数数量)
• 模型可能需要更多时间才能收敛。 (训练更多epoch)
• ResNet 可能不适合此分类。
• 您使用的预训练权重可能不适合此分类。
• 学习率可能太小/太大。
• 等等...