Conversion to TFlite

[nahla-naz]

Hi,

How do you convert models to TFlite? I tried converting, but my app is crashing while trying to get output.

Do you perform optimisation/quantisation? If you do, does it reduce the model accuracy?

Please help. Thanks.

[shubham0204]

You can download the H5 file from the URL present in this script: https://github.com/serengil/deepface/blob/master/deepface/basemodels/Facenet.py

and use the tf.lite.TFLiteConverter.from_keras_model method to convert the model to TFLite. It also provides support for int8 and float16 quantization. I haven't checked the impact of quantization on accuracy (the app uses non-quantized models).

[nahla-naz]

Hi @shubham0204 , Thank you for your prompt reply.

I'm trying to convert FaceNet model from keras_facenet python library.

I am able to convert it using Tensorflow 2.17 and get output in python. But it doesn't run on android because latest version for TFlite in android is 2.16.1. I tried converting the model using lower versions of Tensorflow in python, but the models don't work.

I am getting this error while trying to run TFlite model on android:
Caused by: java.lang.IllegalStateException: Internal error: Unexpected failure when preparing tensor allocations: tensorflow/lite/core/subgraph.cc BytesRequired number of elements overflowed.

• I used the same method as you mentioned to convert the model to TFlite.
• I checked TFlite documentation regarding accuracy loss : there is small accuracy loss based on which quantisation method you have used - sometimes the accuracy loss is insignificant.

[shubham0204]

Considering the error message, it does not seem to be a problem arising from incorrect model conversion but rather from an incorrect shape given for any input/output type on the Android side. You can check the input and output shapes of the TFLite model and compare them with the ones given in your Kotlin/Android code.

[nahla-naz]

@shubham0204 Will check this. However, I think the error is occurring while loading the model before passing any input.