英文:
Transfer learning without pretrained weights
问题
我想**使用Keras上可用的深度学习架构**（[resnet50][1]），**添加一些层，并且在我的数据上训练整个确保模型没有任何预训练权重，同时使所有层都可训练。**
我是转移学习的新手，正在遵循这个[Keras链接][2]以及其他一些资源。
我附上了我在下面代码块中使用的代码。
**我应该怎么做才能确保所有层都是从头开始在我的数据上训练，没有任何预训练权重？**
pre_trained_model = ResNet50(input_shape = (150, 150, 3),
include_top = False, )
使预训练模型中的所有层都可训练（解冻）。
for layer in pre_trained_model.layers:
layer.trainable = True
last_layer = pre_trained_model.get_layer('conv5_block3_3_bn')
last_output = last_layer.output
x = layers.Flatten()(last_output)
x = layers.Dense(1024, activation='relu')(x)
x = layers.Dropout(0.2)(x)
x = layers.Dense(3, activation='relu')(x)
model = Model(pre_trained_model.input, x)
model.compile(optimizer = tf.keras.optimizers.Adam(
learning_rate=0.01),
loss = "mean_squared_error",
metrics = ["mse"])
[1]: https://keras.io/api/applications/resnet/#resnet50-function
[2]: https://keras.io/guides/transfer_learning/
英文:
I would like to use a deep learning architecture available on Keras(resnet50), add a few layers, and train the entire ensure model on my data without any pretrained weights while making all layers trainable.
I am new to transfer learning and am following this Keras link and a few other resources.
I am attaching the code that I am using in the below code block.
What should I do differently to make sure that all layers are trained from scratch on my data without any pretrained weights?
pre_trained_model = ResNet50(input_shape = (150, 150, 3),
include_top = False, )
# Make all the layers in the pre-trained model trainable (unfrozen.)
for layer in pre_trained_model.layers:
layer.trainable = True
last_layer = pre_trained_model.get_layer('conv5_block3_3_bn')
last_output = last_layer.output
x = layers.Flatten()(last_output)
x = layers.Dense(1024, activation='relu')(x)
x = layers.Dropout(0.2)(x)
x = layers.Dense(3, activation='relu')(x)
model = Model( pre_trained_model.input, x)
model.compile(optimizer = tf.keras.optimizers.Adam(
learning_rate=0.01),
loss = "mean_squared_error",
metrics = ["mse"])
答案1
得分: 2
默认情况下，keras.applications 模块中的 Keras 模型使用在 ImageNet 数据集上预训练的权重。只需将 weights 参数设置为 None，即可获得一个使用随机权重初始化的模型。
pre_trained_model = ResNet50(input_shape=(150, 150, 3),
include_top=False, weights=None)
你可以在文档中详细了解更多信息：Keras Applications / ResNet 和 ResNetV2。
英文:
By default, Keras models available in the keras.applications module come with weights pretrained on the imagenet dataset. Simply pass None to the weights argument to get a model initialized with random weights.
pre_trained_model = ResNet50(input_shape = (150, 150, 3),
include_top = False, weights=None)
You can read more in the documentation: Keras Applications / ResNet and ResNetV2 :
> Arguments
>
> - weights: one of None (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded.
通过集体智慧和协作来改善编程学习和解决问题的方式。致力于成为全球开发者共同参与的知识库，让每个人都能够通过互相帮助和分享经验来进步。
评论