A Neural Style Transfer based on VGG19 model

⭐ Star this project on GitHub — it helps!

Neural Style Transfer is a task of transferring style of one image to another. It does it by using features of some pretrained model. In this case as such Base Model the VGG19 pretrained on ImageNet was used. Firstly we create our own model from certain layers of the VGG19 network. And then by adding gradients from the network to the input image we obtain our result image with transferred style.

• Compiling model
• Training
• Results
• License
• Links

As mentioned above, first of all we should compile our model from pretrained one. In this particular case the VGG19 was used. We should define between which of the layers the Content loss and Style loss are going to be calculated. As model's input is going to be the copy of content_image we do not need so much nodes to calculate Content loss as we need for Style loss(In this case 1 node was used for Content loss and 5 nodes for Style loss.

• The model compiler is under model/__init__.py.

Parameters of training:

• Base model: VGG19
• Content loss layer: conv4
• Style loss layers: (conv1, conv2, conv3, conv4, conv5)
• LBFGS optimizer
• Number of epochs: 10

See demo for more details of training process.

• The model trainer is under trainer.py.

This project is licensed under MIT.

• Neural Style Transfer (arXiv article)