Looking for documention on how to create annotations for TFLite model

I’ve created a model using Keras which I am looking to use with Tensorflow Lite. I specifically wanted to pick out keypoints in an image, and I used this very good documentation as a guide: Keypoint Detection with Transfer Learning

I’ve therefore got a model which takes in an image and produces a series of keypoints, and I’ve used TFLite’s convertor function to convert it to a .tflite file. However when I load the .tflite model in my Android project, it tells me “Input tensor has type kTfLiteFloat32: it requires specifying NormalizationOptions metadata to preprocess input images.” which appears to mean I need to add annotations to the .tflite file.

I’ve found a few examples of annotations for various models, and have the general idea that it’s a json file which specifies what the input and output formats should be, however I haven’t been able to find anything detailing specifically the various different fields and values for this file, so I am not sure how to specify in the annotations that the model outputs a number of keypoints.

Can anyone point me to such documentation or otherwise fill me in on how to specify this?

No one at all? There must surely be some documentation for this somewhere that I’ve been unable to find?

I am not familiar with this feature, but these links might help:

2 Likes