Fast Style Transfer for Arbitrary Styles

The original work for artistic style transfer with neural networks proposed a slow optimization algorithm that works on any arbitrary painting. Subsequent work developed a method for fast artistic style transfer that may operate in real time, but was limited to one or a limited set of styles.

This project open-sources a machine learning system for performing fast artistic style transfer that may work on arbitrary painting styles. In addition, because this system provides a learned representation, one may arbitrarily combine painting styles as well as dial in the strength of a painting style, termed "identity interpolation" (see below). To learn more, please take a look at the corresponding publication.

Exploring the structure of a real-time, arbitrary neural artistic stylization network. Golnaz Ghiasi, Honglak Lee, Manjunath Kudlur, Vincent Dumoulin, Jonathon Shlens, Proceedings of the British Machine Vision Conference (BMVC), 2017.

Stylizing an Image using a pre-trained model

• Set up your Magenta environment.

• Download our pre-trained model: Pretrained on PNB and DTD training images

In order to stylize an image according to an arbitrary painting, run the following command.

# To use images in style_images and content_images directories.
$ cd /path/to/arbitrary_image_stylization
$ arbitrary_image_stylization_with_weights \
  --checkpoint=/path/to/arbitrary_style_transfer/model.ckpt \
  --output_dir=/path/to/output_dir \
  --style_images_paths=images/style_images/*.jpg \
  --content_images_paths=images/content_images/*.jpg \
  --image_size=256 \
  --content_square_crop=False \
  --style_image_size=256 \
  --style_square_crop=False \
  --logtostderr

Example results

In order to stylize an image using the "identity interpolation" technique (see Figure 8 in paper), run the following command where $INTERPOLATION_WEIGHTS represents the desired weights for interpolation.

# To use images in style_images and content_images directories.
$ cd /path/to/arbitrary_image_stylization
# Note that 0.0 corresponds to an identity interpolation where as 1.0
# corresponds to a fully stylized photograph.
$ INTERPOLATION_WEIGHTS='[0.0,0.2,0.4,0.6,0.8,1.0]'
$ arbitrary_image_stylization_with_weights \
  --checkpoint=/path/to/arbitrary_style_transfer/model.ckpt \
  --output_dir=/path/to/output_dir \
  --style_images_paths=images/style_images/*.jpg \
  --content_images_paths=images/content_images/statue_of_liberty_sq.jpg \
  --image_size=256 \
  --content_square_crop=False \
  --style_image_size=256 \
  --style_square_crop=False \
  --interpolation_weights=$INTERPOLATION_WEIGHTS \
  --logtostderr

Example results

Training a Model

Set Up

To train your own model, you need to have the following:

1. A directory of images to use as styles. We used Painter by Number dataset (PBN) and Describable Textures Dataset (DTD). PBN training PBN testing DTD dataset
2. The ImageNet dataset. Instructions for downloading the dataset can be found here.
3. A trained VGG model checkpoint.
4. A trained Inception-v3 model checkpoint.
5. Make sure that you have checkout the slim slim repository.

Create Style Dataset

A first step is to prepare the style images and create a TFRecord file. To train and evaluate the model on different set of style images, you need to prepare different TFRecord for each of them. Eg. use the PBN and DTD training images to create the training dataset and use a subset of PBN and DTD testing images for testing dataset.

The following command may be used to download DTD images and create a TFRecord file from images in cobweb category.

$ cd /path/to/dataset
$ path=$(pwd)
$ wget https://www.robots.ox.ac.uk/~vgg/data/dtd/download/dtd-r1.0.1.tar.gz
$ tar -xvzf dtd-r1.0.1.tar.gz
$ STYLE_IMAGES_PATHS="$path"/dtd/images/cobwebbed/*.jpg
$ RECORDIO_PATH="$path"/dtd_cobwebbed.tfrecord

$ image_stylization_create_dataset \
    --style_files=$STYLE_IMAGES_PATHS \
    --output_file=$RECORDIO_PATH \
    --compute_gram_matrices=False \
    --logtostderr

Train a Model on a Small Dataset

Then, to train a model on dtd_cobwebbed.tfrecord without data augmentation use the following command.

logdir=/path/to/logdir
$ arbitrary_image_stylization_train \
      --batch_size=8 \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --vgg_checkpoint=/path/to/vgg-checkpoint \
      --inception_v3_checkpoint=/path/to/inception-v3-checkpoint \
      --style_dataset_file=$RECORDIO_PATH \
      --train_dir="$logdir"/train_dir \
      --content_weights={\"vgg_16/conv3\":2.0} \
      --random_style_image_size=False \
      --augment_style_images=False \
      --center_crop=True \
      --logtostderr

To see the progress of training, run TensorBoard on the resulting log directory:

$ tensorboard --logdir="$logdir"

Since dtd_cobwebbed.tfrecord contains only 120 images, training takes only a few hours and it's a good test to make sure everything work well. Example of stylization results over a few training style images (on style images cobwebbed_0129.jpg, cobwebbed_0116.jpg, cobwebbed_0053.jpg, cobwebbed_0057.jpg, cobwebbed_0044.jpg from DTD dataset):

Train a Model on a Large Dataset With Data Augmentation

To train a model with a good generalization over unobserved style images, you need to train the model on a large training dataset (see Figure 5 here). We trained our model on PBN and DTD training images with data augmentation over style images for about 3M steps using 8 GPUS. You may train the model on 1 GPU, however this will take roughly 8 times as long.

To train a model with data augmentation over style images use the following command.

logdir=/path/to/logdir
$ arbitrary_image_stylization_train \
      --batch_size=8 \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --vgg_checkpoint=/path/to/vgg-checkpoint \
      --inception_v3_checkpoint=/path/to/inception-v3-checkpoint \
      --style_dataset_file=/path/to/style_images.tfrecord \
      --train_dir="$logdir"/train_dir \
      --random_style_image_size=True \
      --augment_style_images=True \
      --center_crop=False \
      --logtostderr

Run an evaluation job

To run an evaluation job on test style images use the following command.

Note that if you are running the training job on a GPU, then you can run a separate evaluation job on the CPU by setting CUDA_VISIBLE_DEVICES=' ':

$ CUDA_VISIBLE_DEVICES= arbitrary_image_stylization_evaluate \
      --batch_size=16 \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --eval_style_dataset_file=/path/to/evaluation_style_images.tfrecord \
      --checkpoint_dir="$logdir"/train_dir \
      --eval_dir="$logdir"/eval_dir \
      --logtostderr

Distill style prediction network using MobileNetV2

To distill the InceptionV3 style prediction network using a MobileNetV2 model, you will need the MobilenetV2 pre-trained checkpoint from https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_1.0_224.tgz and the checkpoint from a trained arbitrary image stylization model (You can use this).

You will also need to install the TF-slim image models library. Unfortunately, there is currently no way to do this other than cloning the repository and adding the module to your $PYTHONPATH.

For best results, use the same datasets used to train the original arbitrary image stylization model (If using the pre-trained model available here, use ImageNet for the content images dataset and a combination of PBN and DTD for the style images dataset).

$ arbitrary_image_stylization_distill_mobilenet \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --style_dataset_file=/path/to/style_images.tfrecord \
      --train_dir=/path/to/logdir \
      --mobilenet_checkpoint=/path/to/mobilenet_v2_1.0_224/checkpoint//mobilenet_v2_1.0_224.ckpt \
      --initial_checkpoint=/path/to/arbitrary_style_transfer/checkpoint/model.ckpt \
      --use_true_loss=False

Train a Model on a Large Dataset With Data Augmentation to run on mobile

You can train a mobile version of the Model to run on mobile phone using TensorFlow Lite. To train the model, you will need the MobilenetV2 pre-trained checkpoint and install TF-slim model library as described in previous section.

To train the mobile Model use the following command.

logdir=/path/to/logdir
$ arbitrary_image_stylization_train_mobile \
      --batch_size=8 \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --vgg_checkpoint=/path/to/vgg-checkpoint \
      --mobilenet_checkpoint=/path/to/mobilenet_v2_1.0_224/checkpoint//mobilenet_v2_1.0_224.ckpt \
      --style_dataset_file=/path/to/style_images.tfrecord \
      --train_dir="$logdir"/train_dir \
      --random_style_image_size=True \
      --augment_style_images=True \
      --center_crop=False \
      --logtostderr

To convert the trained Model to TensorFlow Lite format, use the following command/

logdir=/path/to/logdir
$ arbitrary_image_stylization_convert_tflite \
      --checkpoint="$logdir"/train_dir \
      --imagenet_data_dir=/path/to/imagenet-2012-tfrecord \
      --style_dataset_file=/path/to/style_images.tfrecord \
      --output_dir="$logdir"/tflite \
      --logtostderr

In the output folder, there are 6 TensorFlow Lite models:

1. style_predict.tflite: The float32 version of style prediction network
2. style_predict_quantized.tflite: The int8 weight-quantized version of style prediction network
3. style_predict_calibrated.tflite: The full int8-quantized version of style prediction network
4. style_transform.tflite: The float32 version of style transformation network
5. style_transform_quantized.tflite: The int8 weight-quantized version of style transformation network
6. style_transform_calibrated.tflite: The full int8-quantized version of style transformation network

Image stylization models on TensorFlow Hub

Several TensorFlow Lite variants are available on TensorFlow Hub and these are all ready to use -

• magenta/arbitrary-image-stylization
• arbitrary-image-stylization-inceptionv3
• arbitrary-image-stylization-inceptionv3-dynamic-shapes

You can refer to this sample mobile application for a good understanding of the model usage.

Tutorials

You can follow these tutorials if you want to try out the aforementioned models from a Colab Notebook -

• Artistic Style Transfer with TensorFlow Lite
• Fast Style Transfer for Arbitrary Styles

Community contributed guides

• This Colab Notebook by Sayak Paul shows how to load pre-trained Magenta stylization checkpoints and convert to TensorFlow Lite models using different post-training quantization formats.