Using pretrained models in convolutional autoencoders

vision
1

Is it possible to use a pretrained model like resnet/vgg in a convolutional autoencoder?If so then how can we use it.I know that for standard classification purposes,we ignore the last classification layer,should the same apply in convolutional autoencoders?

2

@sparshgarg23 For an autoencoder setup, the last layer should represent the learned embedding. So you can just remove the activation layer and then make it flow to your decoder setup

Some pointers though

  • Generally the decoder setup is exactly same as the encoder setup

  • You might end up training a huge decoder since your encoder is vgg/resnet. You will have to come up with a transpose of the pretrained model and use that as the decoder, allowing only certain layers of the encoder and decoder to get updated

Following is an article that will help you come up with the model architecture