How do I programmatically train my Google Cloud AutoML Custom Model?

Question

I have a .NET web app I am setting up to identify random objects. I set up a custom model with ~500 labelled images and am able to successfully get a prediction for an image via the REST API.

But how can I add to that models' dataset and train it with new images that get uploaded/sent for prediction?

For example, the user uploads a photo of something and the model isn't able to accurately predict what it is (e.g. highest score is < 50), so the user is prompted to select a label to identify it and then it adds that image with the label to the model for training.

Answer 1

You need to take a few steps:

1. Upload the image file to Google Cloud Storage, under the bucket <project_id>-vcm

2. Upload a CSV file to Google Cloud Storage under the same bucket, with the following content:

   gs://<project_id>-vcm/<path-to-image-file>,<label>

3. Call the importData API to import the CSV file.

Now your dataset is updated with the new image with label. You can call the create model API to train a new model.

Answer 2

We're working hard to make use cases like this simple, but for the time being, you will need to build some services around this. For example, you could create a GCF that accepts your request, forwards it to CloudML Engine, sends the request and response to BigQuery (including the "scores") and then returns the final response. Then you could setup a Cloud Composer pipeline that will periodically retrain the model.