通过部分加载权重进行迁移学习
[英]Transfer learning by loading weights partially
问题描述
Hi I've have one pretrained network model1 as follows:
嗨，我有一个预训练的网络模型 1，如下所示：
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 3000, 200, 1) 0
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 3000, 200, 1) 26 input_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 1500, 200, 1) 0 conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 1500, 200, 1) 26 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 750, 200, 1) 0 conv2d_2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 750, 200, 1) 26 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 375, 200, 1) 0 conv2d_3[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 375, 200, 1) 26 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D) (None, 187, 199, 1) 0 conv2d_4[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 128, 128, 1) 4321 max_pooling2d_4[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 128, 128, 16) 160 conv2d_5[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 128, 128, 16) 2320 conv2d_6[0][0]
__________________________________________________________________________________________________
max_pooling2d_5 (MaxPooling2D) (None, 64, 64, 16) 0 conv2d_7[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 64, 64, 32) 4640 max_pooling2d_5[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 64, 64, 32) 9248 conv2d_8[0][0]
__________________________________________________________________________________________________
max_pooling2d_6 (MaxPooling2D) (None, 32, 32, 32) 0 conv2d_9[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 32, 32, 64) 18496 max_pooling2d_6[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 32, 32, 64) 36928 conv2d_10[0][0]
__________________________________________________________________________________________________
max_pooling2d_7 (MaxPooling2D) (None, 16, 16, 64) 0 conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 16, 16, 128) 73856 max_pooling2d_7[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 16, 16, 128) 147584 conv2d_12[0][0]
__________________________________________________________________________________________________
max_pooling2d_8 (MaxPooling2D) (None, 8, 8, 128) 0 conv2d_13[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 8, 8, 256) 295168 max_pooling2d_8[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 8, 8, 256) 590080 conv2d_14[0][0]
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D) (None, 16, 16, 256) 0 conv2d_15[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 16, 16, 384) 0 up_sampling2d_1[0][0]
conv2d_13[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 16, 16, 128) 442496 concatenate_1[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 16, 16, 128) 147584 conv2d_16[0][0]
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D) (None, 32, 32, 128) 0 conv2d_17[0][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 32, 32, 192) 0 up_sampling2d_2[0][0]
conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 32, 32, 64) 110656 concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 32, 32, 64) 36928 conv2d_18[0][0]
__________________________________________________________________________________________________
up_sampling2d_3 (UpSampling2D) (None, 64, 64, 64) 0 conv2d_19[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 64, 64, 96) 0 up_sampling2d_3[0][0]
conv2d_9[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 64, 64, 32) 27680 concatenate_3[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 64, 64, 32) 9248 conv2d_20[0][0]
__________________________________________________________________________________________________
up_sampling2d_4 (UpSampling2D) (None, 128, 128, 32) 0 conv2d_21[0][0]
__________________________________________________________________________________________________
concatenate_4 (Concatenate) (None, 128, 128, 48) 0 up_sampling2d_4[0][0]
conv2d_7[0][0]
__________________________________________________________________________________________________
conv2d_22 (Conv2D) (None, 128, 128, 16) 6928 concatenate_4[0][0]
__________________________________________________________________________________________________
conv2d_23 (Conv2D) (None, 128, 128, 16) 2320 conv2d_22[0][0]
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 261, 261, 8) 6280 conv2d_23[0][0]
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 267, 267, 4) 1572 conv2d_transpose_1[0][0]
__________________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTrans (None, 300, 300, 2) 9250 conv2d_transpose_2[0][0]
__________________________________________________________________________________________________
conv2d_24 (Conv2D) (None, 300, 300, 1) 3 conv2d_transpose_3[0][0]
==================================================================================================
Total params: 1,983,850
Trainable params: 1,983,850
Non-trainable params: 0
I want to take weights in model1 from conv2d_6 to end and load them into the model2 (mentioned below) from conv2d_3 to end.我想从 conv2d_6 获取 model1 中的权重到 end 并将它们加载到从 conv2d_3 到 end 的 model2（如下所述）中。
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 502, 200, 3) 0
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 502, 200, 3) 30 input_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 251, 200, 3) 0 conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 251, 200, 3) 48 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 125, 100, 3) 0 conv2d_2[0][0]
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 126, 109, 3) 183 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 127, 118, 3) 183 conv2d_transpose_1[0][0]
__________________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTrans (None, 128, 128, 3) 201 conv2d_transpose_2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 128, 128, 16) 448 conv2d_transpose_3[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 128, 128, 16) 2320 conv2d_3[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 64, 64, 16) 0 conv2d_4[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 64, 64, 32) 4640 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 64, 64, 32) 9248 conv2d_5[0][0]
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D) (None, 32, 32, 32) 0 conv2d_6[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 32, 32, 64) 18496 max_pooling2d_4[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 32, 32, 64) 36928 conv2d_7[0][0]
__________________________________________________________________________________________________
max_pooling2d_5 (MaxPooling2D) (None, 16, 16, 64) 0 conv2d_8[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 16, 16, 128) 73856 max_pooling2d_5[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 16, 16, 128) 147584 conv2d_9[0][0]
__________________________________________________________________________________________________
max_pooling2d_6 (MaxPooling2D) (None, 8, 8, 128) 0 conv2d_10[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 8, 8, 256) 295168 max_pooling2d_6[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 8, 8, 256) 590080 conv2d_11[0][0]
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D) (None, 16, 16, 256) 0 conv2d_12[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 16, 16, 384) 0 up_sampling2d_1[0][0]
conv2d_10[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 16, 16, 128) 442496 concatenate_1[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 16, 16, 128) 147584 conv2d_13[0][0]
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D) (None, 32, 32, 128) 0 conv2d_14[0][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 32, 32, 192) 0 up_sampling2d_2[0][0]
conv2d_8[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 32, 32, 64) 110656 concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 32, 32, 64) 36928 conv2d_15[0][0]
__________________________________________________________________________________________________
up_sampling2d_3 (UpSampling2D) (None, 64, 64, 64) 0 conv2d_16[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 64, 64, 96) 0 up_sampling2d_3[0][0]
conv2d_6[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 64, 64, 32) 27680 concatenate_3[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 64, 64, 32) 9248 conv2d_17[0][0]
__________________________________________________________________________________________________
up_sampling2d_4 (UpSampling2D) (None, 128, 128, 32) 0 conv2d_18[0][0]
__________________________________________________________________________________________________
concatenate_4 (Concatenate) (None, 128, 128, 48) 0 up_sampling2d_4[0][0]
conv2d_4[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 128, 128, 16) 6928 concatenate_4[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 128, 128, 16) 2320 conv2d_19[0][0]
__________________________________________________________________________________________________
conv2d_transpose_4 (Conv2DTrans (None, 261, 261, 8) 6280 conv2d_20[0][0]
__________________________________________________________________________________________________
conv2d_transpose_5 (Conv2DTrans (None, 267, 267, 4) 1572 conv2d_transpose_4[0][0]
__________________________________________________________________________________________________
conv2d_transpose_6 (Conv2DTrans (None, 300, 300, 2) 9250 conv2d_transpose_5[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 300, 300, 1) 3 conv2d_transpose_6[0][0]
==================================================================================================
Total params: 1,980,358
Trainable params: 1,980,358
Non-trainable params: 0
I have taken weights_list = model1.get_weights() from pre-trained model and the length of this list is 54. I just can't understand the indexing of the weights in layers.我从预训练的 model 中获取了weights_list = model1.get_weights() ，这个列表的长度是 54。我只是无法理解层中权重的索引。 Model1 has 43 layers and indexing is confusing me. Model1 有 43 层，索引让我很困惑。 Is there any general way to understanding indexing for other models as well?是否有任何通用方法来理解其他模型的索引？ In future models I will be selecting specific layer weights inbetween the models.在未来的模型中，我将在模型之间选择特定的层权重。
1 个解决方案
解决方案1
0 2020-05-13 11:53:33
You need to extract the weights from the specific layer.您需要从特定图层中提取权重。 Using model.get_weights() by itself returns the weights for the entire model.单独使用 model.get_weights() 会返回整个 model 的权重。
The following should work以下应该工作
model1_weights = model1.get_layer('conv2d_6').get_weights()
model2.get_layer('conv2d_3').set_weights(model1_weights)
Edit: To extend this to all the desired layers the simplest thing to do will be to create a list of layer names (mode1_layers) for model1 that you wish to transfer and a list of layer names for model2 (mode2_layers) that you want to transfer to.编辑：要将其扩展到所有所需的层，最简单的方法是为要传输的模型 1 创建层名称列表 (mode1_layers)，并为要传输的模型 2 (mode2_layers) 创建层名称列表至。 Then transfer this in a for loop as below.然后将其转移到 for 循环中，如下所示。
for i in range(len(model1_layers)):
layer_weights = model1.get_layer(model1_layers[i]).get_weights()
model2.get_layer(model2_layers[i]).set_weights(layer_weights)
声明:本站的技术帖子网页，遵循CC BY-SA 4.0协议，如果您需要转载，请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.
相关问题
如何在没有 ImageNet 权重的情况下进行迁移学习？ - How to do Transfer Learning without ImageNet weights?
转学习中的问题 - 无法使用keras传递权重 - Problem in transfer learning - couldn't transfer weights using keras
TensorFLow 迁移学习加载 TFRecordDataset - TensorFLow Transfer Learning loading TFRecordDataset
迁移学习后如何查看预训练模型的权重？ - How to view weights of a pretrained model after transfer learning?
Package 训练神经网络的 model 权重，使其可用于迁移学习 - Package the model weights of a trained neural network to make it usable for transfer learning
尝试为迁移学习任务加载我自己的权重时出现 ValueError - Im getting ValueError when trying to load my own weights for a transfer learning task
如何使用现有CNN模型中的预训练权重在Keras中进行迁移学习？ - How can I use pre-trained weights from an existing CNN model for transfer learning in Keras?
Keras：如何部分加载权重？ - Keras: how to load weights partially?
如何使用来自一个预先训练的MLP的最后一个隐藏层权重作为Keras的新MLP（转移学习）的输入？ - How to use the last hidden layer weights from one pre-trained MLP as input to a new MLP (transfer learning) with Keras?
张量流迁移学习 - Tensor flow transfer learning
相关标签