Smiley Face Discriminator
Using Clarifai's api we created a custom model that discriminates emotional state of a drawn "smiley face". This was part of R/GA / Somatic Hackday March 2017.
Clarifai setup
We had to create an application to represent the model. We noticed that the application had to be recreated everytime we trained a new model, because it expected unique images. Perhaps if we made the image paths unique during training step we could have avoided this.
Clarifai uses a base image model which is what gives the service its magic. You can use a very small training data set to make custom models that work very well.
Training images
We scanned about 25 images to be used in positive training.
About 10 images for negatives.
We saved 1 image each from the Positives and Negatives to do a test after training.
Predictions
We were able to get a decent prediction after this very small training set.
*88% chance of "Smiley" for the postive test *6% chance of "Smiley" for the negative test