加载后更改 MobileNet 丢失

Question

我正在处理迁移学习问题。当我仅从 Mobilenet 创建新模型时，我设置了一个 dropout。

base_model = MobileNet(weights='imagenet', include_top=False, input_shape=(200,200,3), dropout=.15)

x = base_model.output

x = GlobalAveragePooling2D()(x)

x = Dense(10, activation='softmax')(x)

我在训练时使用model_checkpoint_callback. 当我训练时，我会发现过度拟合发生的地方，并调整冻结层的数量和学习率。当我再次保存加载的模型时，我是否也可以调整 dropout？

我看到了这个答案，但是 Mobilenet 中没有实际的 dropout 层，所以这个

for layer in model.layers:

    if hasattr(layer, 'rate'):

        print(layer.name)

        layer.rate = 0.5

什么都不做。

Answer 1

过去，您必须克隆模型才能让新的 dropout 接受。我最近没试过。# This code allows you to change the dropout# Load model from .jsonmodel.load_weights(filenameToModelWeights) # Load weightsmodel.layers[-2].rate = 0.04  # layer[-2] is my dropout layer, rate is dropout attributemodel = keras.models.clone(model) # If I do not clone, the new rate is never used. Weights are re-init now.model.load_weights(filenameToModelWeights) # Load weightsmodel.predict(x)归功于http://www.gergltd.com/home/2018/03/changing-dropout-on-the-fly-during-training-time-test-time-in-keras/如果模型一开始就没有 dropout 层，就像 Keras 的预训练移动网络一样，您必须使用方法添加它们。这是您可以做到的一种方法。用于添加单层def insert_single_layer_in_keras(model, layer_name, new_layer):    layers = [l for l in model.layers]    x = layers[0].output    for i in range(1, len(layers)):        x = layers[i](x)        # add layer afterward        if layers[i].name == layer_name:            x = new_layer(x)    new_model = Model(inputs=layers[0].input, outputs=x)    return new_model用于系统地添加层def insert_layers_in_model(model, layer_common_name, new_layer):    import re    layers = [l for l in model.layers]    x = layers[0].output    layer_config = new_layer.get_config()    base_name = layer_config['name']    layer_class = type(dropout_layer)    for i in range(1, len(layers)):        x = layers[i](x)        match = re.match(".+" + layer_common_name + "+", layers[i].name)        # add layer afterward        if match:            layer_config['name'] = base_name + "_" + str(i)  # no duplicate names, could be done different            layer_copy = layer_class.from_config(layer_config)            x = layer_copy(x)    new_model = Model(inputs=layers[0].input, outputs=x)    return new_model像这样跑import tensorflow as tffrom tensorflow.keras.applications.mobilenet import MobileNetfrom tensorflow.keras.layers import Dropoutfrom tensorflow.keras.models import Modelbase_model = MobileNet(weights='imagenet', include_top=False, input_shape=(192, 192, 3), dropout=.15)dropout_layer = Dropout(0.5)# add single layer after last dropoutmobile_net_with_dropout = insert_single_layer_in_model(base_model, "conv_pw_13_bn", dropout_layer)# systematically add layers after any batchnorm layermobile_net_with_multi_dropout = insert_layers_in_model(base_model, "bn", dropout_layer)顺便说一句，您绝对应该进行实验，但您不太可能希望在 batchnorm 之上对像 mobilenet 这样的小型网络进行额外的正则化。