Neural-Style-Transfer

Fast Image Stylization using Instance Normalization with Pytorch

Start Training

python src/style_transfer.py train 
      --dataset train 
      --style-image style/mosaic.jpg 
      --save-model-dir save 
      --model-name mosaic 
      --cuda 1

Flag description :

--dataset folder containing images for training
--style-image style of image you want to use
--save-model-dir name of the folder where the model will be stored
--model-name name of the model to be saved with .model extensions
--cuda set it to 1 for running in GPU and 0 for CPU

There are several other flags that you can use :

--epochs number of training epoch, default is 2
--batch-size number of batch size for training, default is 4
--pretrained-model pre-trained model path with .model extensions, default is None
--checkpoint-model-dir path to folder where checkpoints of trained models will be saved, default is None
--image-size size of training image, default is 256 x 256
--style-size size of style-image, default is the original size of style-image
--seed random seed for training, default 42
--content-weight weight for content-loss, default is 1e5
--style-weight weight for style-loss, default is 1e10
--lr learning rate, default is 1e-3
--log-interval number of images after which the training loss is logged, default is 500
--checkpoint-interval number of batches after which a checkpoint of the trained model will be created, default is 2000

Evaluate

python src/style_transfer.py eval 
      --content-image image.jpg 
      --output-image image_mosaic.jpg
      --model save/mosaic.model 
      --cuda 1

Flag description :

--content-image path to content image you want to stylize
--output-image path for saving the output image
--model saved model to be used for styling the image
--cuda set it to 1 for running in GPU and 0 for CPU

Demo

The demo notebook are available in Google Colab

Result

Reference

Fast Neural style