pipelines

Pipelines

中文

Bayesian Text Classification

Dataset

The dataset for the text classification Pipeline is organized in the following format:

.
├── test
│   └── textDataBinding.csv
└── train
    └── textDataBinding.csv

The train folder contains the training data, and the test folder contains the test data, stored in csv format. The csv files have two columns of data, input and output, input is the sample data and output is the sample label, e.g:

input	output
原创春秋新款宽松黑色牛仔裤男贴布哈伦裤日系潮流胖男大码长裤子	itemTitle
茗缘翡翠	shopName
挂画精美 种类丰富	itemDesc

These 3 samples represent 3 different categories of text, and their labels are itemTitle, shopName, and itemDesc. It should be noted that the data in the dataset needs to be as rich as possible and relatively evenly distributed, which means that the number of samples in each category should be about the same, too much difference will affect the accuracy of the model.

You can change the data source to the local folder path, like:

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/text-classification-bayes/build/datasource.js?url=file:///path/to/dataset-directory"
}

The /path/to/dataset-directory contains the test and train folders.

Also, you can compress the test and train directories into a zip file and store it on the OSS, modifying the url to the zip file url:

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/text-classification-bayes/build/datasource.js?url=http:///oss-host/my-dataset.zip"
}

Model Parameters

Bayesian model supports both Chinese and English modes, you can specify cn or en by mode parameter, the default is cn.

{
  "model": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/text-classification-bayes/build/model.js?mode=en"
}

Since the Bayesian model uses some operators on tfjs-backend-cpu and other backends are not currently supported, in the options field we define the backend for model training as @tensorflow/tfjs-backend-cpu.

ResNet/MobileNet Image Classification

Dataset

The dataset for the image classification Pipeline is organized in the following format:

.
├── test
│   ├── class-1
│   └── class-2
├── train
│   ├── class-1
│   └── class-2
└── validation
│   ├── class-1
│   └── class-2

The train folder contains the training data, the test folder contains the test data, the validation folder contains the validation data, and the directory contains the image folders for each category, the folder name is the category of the image.

You can specify the url to the local folder path:

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/image-classification/build/datasource.js?url=file:///path/to/dataset-directory"
}

The /path/to/dataset-directory contains the test and train folders.

Or http url:

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/image-classification/build/datasource.js?url=http://oss-host/dataset.zip"
}

Data Processing

For the image classification task, the dimension (width and height) of all sample images input to the model must be the same, and our predefined MobileNet and ResNet models both require a 224 * 224 image input, so before model training begins, we perform a resize operation on the images via the dataflow script.

{
  "dataflow": [
    "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/image-classification/build/dataflow.js?size=224&size=224"
  ]
}

Model Parameters

The image classification pipeline supports both MobileNet and ResNet models. The modelUrl parameter specifies mobilenet or resnet, the default is mobilenet.

{
  "model": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/image-classification/build/model.js?modelUrl=resnet",
}

In addition, the options field can be configured to enable or disable the GPU, and the epochs for training:

{
  "options": {
    "framework": "tfjs@3.8",
    "gpu": false,
    "train": {
      "epochs": 10
    }
  }
}

GPU is enabled by default. The larger the epochs, the longer the training time.

YOLO Object Detection

Dataset

Object Detection pipeline supports PascalVoc and Coco, specifying the current dataset format by defining the format parameter as pascalvoc or coco.

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/object-detection-yolo/build/datasource.js?format=pascalvoc&url=https://host/dataset.zip"
}

You can specify the url as a local folder path:

{
  "datasource": "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/object-detection-yolo/build/datasource.js?format=pascalvoc&url=file:///path/to/dataset-directory"
}

Data Processing

As with the image classification pipeline, YOLO requires that the dimension (width and height) of all sample images input to the model must be the same, 416 * 416, so before the model training begins, we perform a resize operation on the images with the dataflow script.

{
  "dataflow": [
    "https://cdn.jsdelivr.net/gh/imgcook/pipcook-script@b1f834c/scripts/object-detection-yolo/build/dataflow.js?size=416&size=416"
  ]
}

Model Parameters

The options field allows you to configure whether the object detection pipeline is GPU-enabled, the epochs for training, the number of samples to be fed into the model per batchSize, and the patience value for early-stopping:

{
  "options": {
    "framework": "tfjs@3.8",
    "gpu": false,
    "train": {
      "epochs": 100,
      "batchSize": 16,
      "patience": 10
    }
  }
}

GPU is enabled by default. patience indicates that the loss stops training after patience epochs of no decline. For example, if patience is 3, if there is no drop in loss for 3 epochs, early-stopping will be triggered and training will be terminated early.