Open mmeierer opened 4 years ago
Here is another thing to try:
One concrete example of this issue are Deformable Convolutions, which seem to improve the performance of most Convolutional Neural Networks (CNNs). An open source implementation appeared about 2 years ago. Nevertheless, the implementation was cumbersome to integrate into PyTorch/TensorFlow and the algorithm didn’t gain widespread use. Only just recently has PyTorch added support for it, and as of yet I am not aware of there being an official TensorFlow version.
Source: https://tryolabs.com/blog/2020/04/02/swift-googles-bet-on-differentiable-programming/
And other way to improve DL models (here BERT), which might also be applicable to GAN:
https://threadreaderapp.com/thread/1247998807494684672.html
--> Train long and focus on regularization/data augmentation.
This blog post might give you some additional hints how to improve parts of the model: https://towardsdatascience.com/how-to-get-beautiful-results-with-neural-style-transfer-75d0c05d6489
Accompanying GitHub can be found here: https://github.com/EugenHotaj/nn-hallucinations