NeuralStyle

This repository includes implementations for descriptive and generative (including Texture Networks) style transfer techniques. The algorithms are generally based on the originals introduced in [2], [3] and [4], however it's not intended to reproduce the exact results of either paper, thus some hyperparameters and other implementation details have been changed.

Dependencies

This project uses PyTorch 0.4.1 and Visdom (8c6322d) for training visualizations. Please refer to their corresponding project websites for installation instructions.

Examples

Original

Transforms

Algorithm
Descriptive
Fast Transfer
Texture Network
Style image

Usage

The scripts descriptive.py and generative.py may be used to iteratively transform a single image or train a neural network, respectively. Common arguments include:

• --cuda use single gpu if available
• --style-data path to target style image
• --content-data path to a directory of directories with images (generative) or a single image (descriptive)
• --shape height and width for content data
• --content-weight weight for content-component loss
• --style-weight weight for style-component loss
• --tv-weight weight for total variation-component loss
• --name name for checkpoints or transformed images
• -lr learning rate
• --epochs number of iterations, epochs over all training examples (generative) or optimization steps (descriptive)
• --log-every step frequency to log statistics
• --log-images-every step frequency to log transformed images

descriptive.py also has the option:

• --save-directory path to directory where transformed images will be saved

generative.py also has the options:

• --eval use to transform an image with a pretrained model
• --model path to pretrained model (used with --eval)
• --texture use to train or evaluate a Texture Network
• --noise-scale set a float to scale the magnitude of the noise tensor in Texture Networks

export.py Script to export pretrained models into ONNX.

• --ckpt path to pretrained model, it must include the file's parent directory, this will be used as the new file's name
• --data dummy image to use during export, if this isn't set, a random tensor will be used, but shape must be set
• --shape height and width of dummy image

For live visualizations you must have a local visdom server running: python -m visdom.server.

Examples

python generative.py --cuda --content-data /home/username/datasets/coco/images --style-data .assets/styles/wave.jpg --name wave --batch-size 16 --content-weight 1 --style-weight 5 --tv-weight 7 --epochs 2 

python descriptive.py --content-data .assets/guadalajara.jpg --style-data .assets/styles/wave.jpg --name wave --log-every 2 --log-images-every 3 

Alternatively, the script main.py may be used with the flags --descriptive or --generative.

References

[1] Jing, et al. (2018). Neural Style Transfer: A Review. arXiv:1705.04058v6 [cs.CV]

[2] Gatys L., Ecker A., Bethge M. (2015). A Neural Algorithm of Artistic Style. arXiv:1508.06576v2 [cs.CV]

[3] Johnson J., Alahi A., Fei-Fei L. (2016). Perceptual Losses for Real-Time Style Transfer and Super-Resolution. arXiv:1603.08155v1 [cs.CV]

[4] Ulyanov D., et al. (2016). Texture Networks: Feed-forward Synthesis of Textures and Stylized Images. arXiv:1603.03417v1 [cs.CV]