Tensorflow pipeline build with NodeJS

This project aims to streamline the process of training a Tensorflow V2 deep neural network model from A to Z. This project assumes you have already set up your system and environment with the Python >= V3.5

pre_process.js is script that provides the sub-processes that will help you through the entire process from labeling the dataset to generating the inference graph with a saved model that can be used in your nodejs application for inferencing or object detections.

• Start with your image dataset
• Resize images (depending on your use case, smaller images size speeds up training time)
• Use tool like labelImag, Maximo Visual Inspection & etc. to label and create bounnding boxes for each class type for each image
• Break the dataset into a training set and a testing set (9:1 ratio)
• Covert csv files from xml files of the data set for generating TFRecord
• Generate TRecord
• Train model
• Generate inference graph from trained model

Step 1

• Git clone git@github.com:playground/tfjs-pipelline.git
• CD into tfjs-pipeline and run npm install
• Create a directory <my_dataset> for you image dataset anywhere on your machine

Step 2

• In tfjs-pipelne directory, resize your images uniformly by runnning

  npm run pre_process --task=resize_image --image_dir=<my_dataset> --width=800 --height=600

Step 3

• Use one of the tools like lableImg, MVI(Maximo Visual Inspection) or any others to label and create bounding boxes for each of your images in your dataset
• The tool will generate a mapping xml file with the bounding box coordinate for each custom class you have created for your dataset
• Here is an example of what the xml output from labelImg will look like

<annotation>
	<folder>images</folder>
	<filename>hardhat1.jpg</filename>
	<path>/Users/jeff/git_repo/sandbox/wu/playbox/ml/TensorFlow/workspace/training_demo/images/hardhat1.jpg</path>
	<source>
		<database>Unknown</database>
	</source>
	<size>
		<width>800</width>
		<height>600</height>
		<depth>3</depth>
	</size>
	<segmented>0</segmented>
	<object>
		<name>hard-hat</name>
		<pose>Unspecified</pose>
		<truncated>0</truncated>
		<difficult>0</difficult>
		<bndbox>
			<xmin>145</xmin>
			<ymin>37</ymin>
			<xmax>453</xmax>
			<ymax>231</ymax>
		</bndbox>
	</object>
</annotation>

• Depending on which tool you use to label your dataset, the xml format will vary but the neccessary information are there for converting to cvs in the following steps
• Pre_process.js script currently supports both labelImg and MVI formats conversion
• Now is a good time to create a labelmap.pbtxt file which will be needed for training the model in a later step. This file should contain the custom classes you have defined for labeling the dataset. Here is an example of a lablemap.pbtxt file:

  item {
    name: "none_of_the_above"
    id: 0
    display_name: "background"
  }
  item {
    id: 1
    name: 'hard-hat'
    display_name: "hardhat"
  }

Step 4 (optional)

• If you would like to partition your dataset into test and train model, run the following command

  npm run prec_process --task=partition_dataset --image_dir=<my_dataset_dir>

Step 5

• Convert xml to cvs files which are needed by generate_tfrecord.py to generate TFRecord

  npm run pre-process --image_dir=<my_dataset_dir> --task=xml_to_csv
npm run pre-process --image_dir=<my_dataset_dir> --task=xml_to_csv --origin=maximo

Step 6

• To generate the TFRecord that we can use to train the model, run the following command

  npm run pre_process --task=generate_tfrecords --image_dir=<my_dataset_dir>

Step 7

• To train model, run the following commmand

  npm run pre_process --task=train_model --pipeline_config_path=<my_dataset_dir>/pipeline.config --model_dir=<my_dataset_dir>/training

Step 9

• To generate inference graph from trained model, run the following command

  npm run pre_process --task=export_inference_graph --trained_checkpoint_dir=<my_dataset_dir>/training --pipeline_config_path=<my_dataset_dir>/pipeline.config --output_directory=<my_dataset_dir>/inference_graph

Build Docker Image

docker build --ssh github=$HOME/.ssh/id_rsa -t tfjs-pipeline-ubuntu-2104 .

Train model in docker container

docker run -it --rm --privileged -v <PATH_TO_DATASET> playbox21/tfjs-pipeline-ubuntu-1804:1.0.0

Or you can pull from docker hub

docker pull playbox21/tfjs-pipeline-ubuntu-1804:1.0.0

docker pull playbox21/tfjs-pipeline-ubuntu-2104:1.0.0