Skip to content

Latest commit

 

History

History

NNTrainer

README.md

Neural Network Post Processing

1. data preparation

I use coco2017 val for training datasets, download at coco2017, unzip into "datasets" folder, pictures should be in "datasets/coco2017/val/"

add your style image to "datasets/style", remember that the image name is <your-style-name>

2. modify the network

You can decrease layers and filters it if it runs too slow, increase if it underfit.

in "src/nets.py", method SimpleTransformNet()

3. train the network

run python train_fst.py --style <your-style-name>

4. export to unity

run python exporter.py --dataset_name <your-style-name>

5. parse model in unity

  • In Unity Project Window selected your recent imported model, should be Assets/Script/RawModel/<your-dataset-name>.json, right click it, from the pop-up menu select "ParseFromRawModel"
  • Add <your-dataset-name> to script enum "NNStyle"

Now you should be able to select the model from NNPostProcessingEffect inspector window

Requirements

python>=3.8 tensorflow>=2.3.0 scipy>1.5.4 imageio>=2.9.0 scikit-image>=0.17.2