使用VGG19训练kaggle数据state-farm-distracted-driver-detection

    !mv data/train/c0/img_1*.jpg data/valid/c0
    !mv data/train/c1/img_1*.jpg data/valid/c1
    !mv data/train/c2/img_1*.jpg data/valid/c2
    !mv data/train/c3/img_1*.jpg data/valid/c3
    !mv data/train/c4/img_1*.jpg data/valid/c4
    !mv data/train/c5/img_1*.jpg data/valid/c5
    !mv data/train/c6/img_1*.jpg data/valid/c6
    !mv data/train/c7/img_1*.jpg data/valid/c7
    !mv data/train/c8/img_1*.jpg data/valid/c8
    !mv data/train/c9/img_1*.jpg data/valid/c9

    from keras.applications import *
    from keras.layers import Input, Lambda, Dense, Dropout
    from keras.preprocessing.image import ImageDataGenerator
    from keras.models import Model
    from keras.optimizers import SGD
    from keras.callbacks import ModelCheckpoint, EarlyStopping
    import h5py
    height = 224
    width = 224
    x = Input((width, height, 3))
    base_model = VGG19(input_tensor=x, input_shape=(width, height, 3), include_top=False, pooling="avg", weights="imagenet")
    imgGen = ImageDataGenerator()
    train_gen = imgGen.flow_from_directory("data/train", target_size=(width, height), shuffle=True, batch_size=32)
    valid_gen = imgGen.flow_from_directory("data/valid", target_size=(width, height), shuffle=True, batch_size=32)
    new_output = Dropout(0.5)(base_model.output)
    new_output = Dense(10, activation="softmax")(new_output)
    new_model = Model(inputs=base_model.input, outputs=new_output)
    for layer in new_model.layers[:-9]:
      layer.trainable = False
    sgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)
    new_model.compile(optimizer=sgd, loss="categorical_crossentropy", metrics=["accuracy"])
    #new_model.summary()
    new_model.load_weights("drive/vgg19.hdf5")
    #这里的checkpoint可以在每个epoch后检测val_loss如果比之前的小，那么把权重存下来
    m_checkpoint = ModelCheckpoint("drive/vgg19.hdf5", monitor="val_loss", mode="min", save_best_only=True, save_weights_only=True, verbose=1, period=1)
    #这里的early stopping可以在5个epoch个没有改进就停止训练，防止过拟合
    m_earlyStopping = EarlyStopping(monitor="val_loss", min_delta=0.0003, patience=5, verbose=1, mode="min")
    new_model.fit_generator(train_gen, 19544 // 32, 15, validation_data=valid_gen, validation_steps=2880 // 32, callbacks=[m_checkpoint, m_earlyStopping])