Avoid MaxPool3D conversion to FlexMaxPool3D

I used Colab to convert a keras model with some MaxPool3D layers to a TFLite model.

But by looking at the converted model with netron.app, I found it was converted to a FlexMaxPool3D.
AFAIK, such a model is not directly usable for inference (in my case, I’ll use it in JavaScript).

Although it is an op listed as implemented ( Supported Select TensorFlow operators  |  TensorFlow Lite ), how can I have it properly converted?

What is your TF version?

probably i was using an older version, it works now with the latest converter:

 tfjs.converters.save_keras_model(model, f'{output_path}/model_tfjs')
1 Like