Dotscience roadsigns

A Jupyter notebook used to train a tensorflow model to recognize german roadsigns.

instructions

Create a GPU enabled runner and install Dotscience on it. To skip installing nvidia-runtime on a GCP runner, search for the VM NVIDIA GPU Cloud Image for Deep Learning and HPC. This comes with nvidia-runtime installed. Then follow the instructions on the Dotscience GCP runner setup page.

Create a new dotscience project.

Upload the files:

• get-data.ipynb
• roadsigns.ipynb
• signnames.csv

Launch Jupyter on a runner.

Run through each cell in the get-data.ipynb notebook and then roadsigns.ipynb.

Observe that

It will output the model into the roadsigns/1 folder and annotate dotscience with roadsigns being the model folder.

This is because Tensorflow serving requires there to be a version folder under the model name folder. We have fixed this to 1 for the demo.

retraining

To re-train the model - try changing the optimizer to sgd rather than adam (it should perform less well) and also change epochs param (search for ds.parameter).

advanced demo of production usage (requires modeldemo environment access)

running the app

If you have the tensorflow server app running on the docker or k8s hostname roadsigns-tensorflow - here is how to build and run the app.

build

cd app
docker build -t roadsignsapp .

run

docker run -d \
  -p 8000:80 \
  --link roadsigns-tensorflow:roadsigns-tensorflow \
  -e TENSORFLOW_HOST=roadsigns-tensorflow:8501 \
  --name roadsignsapp \
  roadsignsapp 

If you are using Kubernetes - you don't need the link but the TENSORFLOW_HOST should be set to the service name of the tensorflow serving service.

App proxy

To get around CORS - the app will proxy any requests to /v1 onto the tensorflow container - for ingress you only need to foward traffic to the app container and it will look after getting data to the tensorflow serving container.

image data

The image data we are sending is prebuilt into app/www/appdata.json

This data has 3 top level fields:

• image_filenames - the filenames of the images to view
• image_labels - the corresponding labels for those images
• tensorflow_images - the data structure of the image we send to tensorflow

The tensorflow_images property was JSON exported from the data passed to the model in the notebook.

adjusting payload

In the file app/www/js/init.js is this code at the top:

var TENSORFLOW_URL = '/v1/models/roadsigns_model:predict'
var appData = null

function loadResult(label, imageData) {
  $('#results-label').text(label)
  $('#results-data').hide()
  $('#results-error').hide()
  $('#results-loading').show()

  var requestPayload = {
    inputs: {
      keep_prob: [1], 
      keep_prob_conv: [1],
      x: [imageData]
    },
  }

  $.ajax({
    method: 'POST',
    url: TENSORFLOW_URL,
    data: JSON.stringify(requestPayload),
    dataType: 'json',
    contentType: "application/json; charset=utf-8",
    success: function(response) {
      $('#results-loading').hide()
      var data = JSON.parse(response.responseText)
      var dataString = JSON.stringify(data, null, 4)
      $('#results-data').show()
      $('#results-json').text(dataString)
    },
    error: function(response) {
      var errorMessage = response.status + ' ' + response.statusText
      $('#results-error').text(errorMessage)
      $('#results-loading').hide()
      $('#results-error').show()
      $('#results-data').show()
      $('#results-json').text(response.responseText)
    }
  })
}

You can adjust the requestPayload accordingly.

The output is shown as raw JSON in the model window that appears.