Build an Image Style‑Transfer Service with Pipcook and CycleGAN

Introduction

To help developers learn Pipcook and machine learning, this practical tutorial demonstrates how to use Pipcook for image style transfer, alongside other examples such as front‑end component recognition, AI poetry, and blog classification. For background on Pipcook 1.0, see the linked article.

Background

Image style conversion is common in camera or photo‑editing apps. Pipcook supports style‑transfer models that can, for example, turn a horse into a zebra, an apple into an orange, or convert photos to artistic paintings.

Pipeline Configuration

Create a pipeline configuration file cycle-gan.json with the following content:

{
  "plugins": {
    "dataCollect": {
      "package": "@pipcook/plugins-image-classification-data-collect",
      "params": { "url": "https://ai-sample.oss-cn-hangzhou.aliyuncs.com/apple2orange.zip" }
    },
    "dataAccess": { "package": "@pipcook/plugins-pascalvoc-data-access" },
    "modelDefine": { "package": "@pipcook/plugins-tensorflow-cycle-gan-model-define" },
    "modelTrain": {
      "package": "@pipcook/plugins-image-generation-tensorflow-model-train",
      "params": { "niter": 50000 }
    },
    "modelEvaluate": { "package": "@pipcook/plugins-image-generation-model-evaluate" }
  }
}

The pipeline uses the following plugins:

@pipcook/plugins-image-classification-data-collect : downloads and extracts the image dataset.

@pipcook/plugins-pascalvoc-data-access : converts the dataset to Pascal VOC format for the style‑transfer model.

@pipcook/plugins-tensorflow-cycle-gan-model-define : defines a CycleGAN model based on TensorFlow.

@pipcook/plugins-image-generation-tensorflow-model-train : trains the model; the niter parameter controls training iterations (50000 used here).

@pipcook/plugins-image-generation-model-evaluate : evaluates the trained model and reports loss values.

Training CycleGAN models requires substantial compute resources; using a GPU server is recommended.

Training

Run the pipeline with: $ pipcook run cycle-gan.json --verbose The console output shows plugin loading, dataset preparation, and detailed TensorFlow layer information for both discriminator models. Training is expected to take about 12 hours. After training, an output folder containing the model is generated. Install it locally with:

$ cd output && npm install

Style Conversion

With the model ready, use the following Node.js script to convert images:

const predict = require('./output');
const fs = require('fs');
function saveToFile (base64Data, file) {
  var dataBuffer = new Buffer(base64Data, 'base64');
  fs.writeFile(file, dataBuffer, function(err) {
    if (err) {
      console.log('保存文件失败', err.message);
    } else {
      console.log('保存文件成功:', file);
    }
  });
}
(async () => {
  const orange_from_apple = await predict({ path: '/path/to/apple.jpg', predictType: 'a2b' });
  saveToFile(orange_from_apple, '/path/to/save/orange_from_apple.jpg');
  const apple_from_orange = await predict({ path: '/path/to/orange.jpg', predictType: 'b2a' });
  saveToFile(apple_from_orange, '/path/to/save/apple_from_orange.jpg');
})();

Run the script with $ node ./apple2orange.js and the generated images will be saved.

Dataset Structure

The dataset for style transfer follows this layout (illustrated below):

The train folder is used for model training, while test is for validation. Both contain subfolders A and B with unpaired images; for example, place any apple images in A and any orange images in B to train an apple‑to‑orange model.

Advanced Tips

You can experiment with other datasets such as:

Selfie ↔ Anime: selfie2anime.zip

Apple ↔ Orange: apple2orange.zip

Horse ↔ Zebra: horse2zebra.zip

Monet ↔ Photo: monet2photo.zip

Summer ↔ Winter: summer2winter_yosemite.zip

Conclusion

You now know how to use Pipcook to implement image style transfer. The next article will explore using Pipcook for Chinese poetry generation.