Uses the Inception5h convolutional neural network to classify and image and Snowplow micro to track new images, faulty extensions and the top 5 probability labels. The service is written in Go 1.12 and uses Tensorflow 1.12.
Requires golang1.12 and a *nix environment which supports Tensorflow 1.12.
To setup tensorflow for golang run:
make setup-tf
To setup the convolutional neural network Inception5h run:
make setup-cnn
Enter the image-recognition directory and run the program:
cd image-recognition
make
The primary purpose of tracking with Snowplow is to detect the top 5 probability labels which result from an image classification. By combining this with the IP address of the client a geographical estimate can be made of the interests of users of the service.
Snowplow micro is also used to track the creation of a new image recognition process. The tracking is done initially to store the type of valid and invalid extensions. The information can later be used to optimize the usage for certain extensions and measure the faulty requests. If the number of faulty requests is determined to be too high then this backend service will be made accesible only through a frontend which can use JavaScript to vet the extension type, thereby, reducing the cost of running the server.
To classify and image pass a POST request with:
curl localhost:8090/recognize -F 'file=@./cat.jpg'
To track the stored events use httpie to provide a nice layout and run:
http localhost:9090/micro/good
Over the weekend 8-10 May Snowplow micro was added to the origin of this repository. If easier it can be used instead as deployment is done through Docker.
You can access the origin here: https://github.com/NikStoyanov/image-recognition and run make to deploy.
The relevant Snowplow micro commit is: https://github.com/NikStoyanov/image-recognition/commit/fe79eb8f8225158e4a1e58eb1104058cb4060309