Build and Deploy ML Models with Pipcook 2.0 in Under 20 Seconds

Pipcook 1.0 allowed web developers to start machine learning with a low barrier, but users reported difficult installation, low success rates, and long setup times (often over three minutes). Pipcook 2.0 addresses these issues with a complete rewrite and optimization.

Installation time drops from several minutes to less than 20 seconds, and the daemon no longer needs to be installed via pipcook init. The lighter package size yields a near‑100% installation success rate.

Model training and service launch for a text‑classification task now take only about 20 seconds.

Pipeline Introduction

In Pipcook, a Pipeline defines the workflow of a model. Four pipelines are currently provided:

image classification MobileNet

image classification ResNet

text classification Bayes

object detection YOLO

The article demonstrates the text classification Bayes pipeline.

A Pipeline is described with JSON that specifies the datasource, dataflow, model script, artifacts, and options.

{
  "specVersion": "2.0",
  "type": "ImageClassification",
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@5ec4cdf/scripts/image-classification/build/datasource.js?url=http://ai-sample.oss-cn-hangzhou.aliyuncs.com/image_classification/datasets/imageclass-test.zip",
  "dataflow": [
    "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@5ec4cdf/scripts/image-classification/build/dataflow.js?size=224&size=224"
  ],
  "model": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@5ec4cdf/scripts/image-classification/build/model.js",
  "artifacts": [],
  "options": {
    "framework": "[email protected]",
    "train": {
      "epochs": 10
    }
  }
}

The JSON fields include datasource, dataflow, model, artifacts, and options. Supported pipeline types are ImageClassification, TextClassification, and ObjectDetection, with more to be added in future releases.

Running a Pipeline

Installation

Operating system: macOS or Linux (Windows has limited support)

Node.js v12.17.0 or v14.0.0 and above (v13 not supported)

Install the CLI globally:

$ npm install @pipcook/cli -g

Training

Save the pipeline definition as image-classification.json and run:

$ pipcook train ./image-classification.json -o my-pipcook

Sample log output shows framework preparation, script download, dataset extraction, and training progress across 10 epochs, ending with the model saved in the workspace.

ℹ preparing framework
ℹ preparing scripts
ℹ preparing artifact plugins
ℹ initializing framework packages
ℹ running datasource script
... (truncated log) ...
ℹ pipeline finished, the model has been saved at /Users/pipcook-playground/my-pipcook/model

The workspace layout after training:

my-pipcook
    ├── cache
    ├── data
    ├── framework -> /Users/pipcook-playground/.pipcook/framework/... 
    ├── image-classification.json
    ├── model
    └── scripts

Prediction

Place an image (e.g., avatar.jpg) and run: $ pipcook predict ./my-pipcook -s ./avatar.jpg The result shows the category avatar with a confidence of 0.99999.

Deployment

Serve the trained model:

$ pipcook serve ./my-pipcook/image-classification.json

The service starts at http://localhost:9091 (default port, configurable with -p). A web UI allows users to upload images and obtain predictions.

Conclusion

Pipcook 2.0 dramatically reduces the friction of building, training, and deploying machine‑learning models for web developers. The project welcomes stars, issues, pull requests, and community participation via its GitHub repositories.

Pipcook repository: https://github.com/alibaba/pipcook Script repository: https://github.com/imgcook/pipcook-script