An alternative to cloud-based image classification is available with PowerAI Vision, usually deployed as a service within your own data centre, allowing for ready access to internal-use images, documents, databases, etc. without the concerns for bandwidth/responsiveness and security that can arise with shipping content to cloud-resident environments.
The following process sets up a PowerAI Vision instance that can be trained for custom classification purposes, and will show the simplicity with which a classifier can be created and used by a consumer application (in this case, NodeRED).
If you have access to a Power linux system of your own, then you can use the local trial download to create your instance.
If you do not have a Power Linux system to hand, you can use the IBM Partnerworld 3-day trial service to create a temporary instance.
Note:- this trial requires (free) registration with IBM Partnerworld - you will need to be a registered company to be eligible.
Once registered, and you receive the confirmation email that the trial has been activated, you should see a tile appear on your IBMid profile dashboard:
Click on the Launch button to access the launch page:
Note:- the name/IP address of this service varies during the trial - if the instance is suspended, and needs to be resumed, a new IP address will more than likely be assigned, resulting in a change of hostname too. Any process that depends on this (such as the NodeRED application) will need to be kept up to date with the current IP address.
To access the training dialog for the AI Vision classifier, launch the "datasets" page:
http://<IP Address>:9080/powerai-vision/index.html#!/datasets - remembering to replace with the current hostname
from the trial summary table.
This should take you to a training page:
This page is the same for the Partnerworld-based trial, the local Power Linux iunstallation trial, and a product deployment. The rest of this training process applies to any of these.
The first step is to add a training dataset - this will contain training images, associated with each class the service will be required to classify.
Give your dataset a name, then start loading data to identify the first class of image content. (in the AI Vision tool, "classes" are referred to as "categories")
Start with a "Sedans" category, and load up 15-20 images representing sedan cars:
Resulting in a populated category of "sedans".
Go through the same process for different types of vehicle - there are samples for pickup trucks, and golf carts in the /demo folder of this repository.
Now the training images are loaded, the system needs to build a model using the underlying deep learning platform to generate a classifier that will recognize (or not!) the different vehicle categories.
This is achieved by creating a new DL (deep learning) task, and setting it up to build an image classifier.
Take the default settings;
you will be prompted to accept that it may take some time to perform the task - click Create New Task
After 10-15 mins (or less on a good day), you should see the results of the model training - ideally, this should show a reduction in training loss, and general rise and stabilisation of accuracy:
The Deploy and Test button indicates the next step - this will create a web service that can be called to classify images.
This is the service endpoint that will be used by the NodeRED application.
You will be prompted to run a test of the classifier through its deployment API - upload an image or two that you have not used in the training datasets. to get an idea of accuracy.
and then capture the (partial) API endpoint information from the My Web APIs view:
This path (/dlapis/) needs to be qualified with your AI Vision host path; for a 3-day trail service, this will result in a classifier URL like:
http://nae-165-254-189-35.jarvice.com:9080/powerai-vision/api/dlapis/ca7380af-b616-4af9-b523-dbdfc65046ab
This URL will be used to POST images to the classifier service.
With an extra stage added to the Watson Visual Recognition example eslewhere in this repository, you can augment the Watson classifier with your specific model results -
