When you are using Google Cloud AI Platform batch prediction, launching a job takes an input path (where multiple inputs are placed) and an output path (where it will put the result in files).
The problem is that you can’t link the output result to its input, so I needed to pass the key in input to the model and each result will have the key allowing to find back which input it was.
I couldn't find a solution on the internet myself so I had to implement it modifying Tensorflow Models tools, I hope it will help people that was in the same case than me.
How to use it:
Additionally, I have built a notebook to download, edit and deploy a model to AI Platform with Google Colab
UPDATE: i’ve built a Cloud Run component that does the work here