Skip to content

Latest commit

 

History

History
97 lines (67 loc) · 3.71 KB

README.md

File metadata and controls

97 lines (67 loc) · 3.71 KB
Raw

Style Transfer training and using the model in ml5js

This repository contains a slightly modified version of Fast Style Transfer in TensorFlow. It trains a neural network on the style of any image you provide it and outputs a model you can use in ml5.js with the ml5.styleTransfer() method.

Notes:

  • You should train your network using a GPU. Just using CPU will result in training times of several months 📆
  • Here is a blog post describing how to train your own custom style with Paperspace. This is the easiest way to get up and running without installing dependencies and libraries.

Requirements

Usage

1) Download this repository

Start by downloading or cloning this repository:

git clone https://github.com/ml5js/training_styletransfer.git
cd training_styletransfer

2) Install dependencies and get the training data

If you are running this without the Docker image, you will need to get the complete COCO Dataset, about 14GB of data. This is a requirement for training. You can download the data by running:

bash setup.sh

You will also need to install specific dependencies for this project:

pip install -r requirements.txt

3) Select a style image

Put the image you want to train the style on, in the /images folder.

4) Start the training

To train a new style transfer network you can use open the run.sh script, modified the --style argument to point to your image and run:

bash run.sh

Or run the training code directly:

python style.py --style images/YOURIMAGE.jpg \
  --checkpoint-dir checkpoints/ \
  --model-dir models/ \
  --test images/violetaparra.jpg \
  --test-dir tests/ \
  --content-weight 1.5e1 \
  --checkpoint-iterations 1000 \
  --batch-size 20

--style should point to the image you want to use. --model-dir will be the folder where the ml5.js model will be saved.

You can learn more about how to use all the parameters for training in the on the original repository for this code here and here.

Once the training setup is ready, you should see something like this:

ml5.js Style Transfer Training!
Note: This traning will take a couple of hours.
Training is starting!...
Train set has been trimmed slightly..
(1, 451, 670, 3)
UID: 56
Epoch 0, Iteration: 1000, Loss: 1.75362e+07
style: 5.5727e+06, content:1.15116e+07, tv: 451984.0
...
Training complete. For evaluation:
    `python evaluate.py --checkpoint checkpoints/ ...`
Converting model to ml5js
Writing manifest to models/manifest.json
Done! Checkpoint saved. Visit https://ml5js.org/docs/StyleTransfer for more informationlive

5) Use it!

Once the model is ready, your model will be in the models/ folder. You will just need to point to it in your ml5 sketch:

const style = new ml5.styleTransfer('./models/your_new_model');

Check ml5.js documentation on the styleTransfer() method or the style transfer simple example.