Support for int32 in strided slice operation

[Aakash Tyagi]

Hello

I am trying to run a quantized tflite model using tflm. I stared with the MNIST example give @https://www.tensorflow.org/lite/performance/post_training_integer_quant

After quantizing it to int8 and trying to run it using tflm I'm getting error that int32 is not supported for strided slice. How do i solve this problem? Any suggestions are welcomed.

[sandeep singh]

Hi, 

This is correct behaviour as Keras by default sets the dynamic batch size to true.

So, your model input shape is [*,28,28] not [1,28,28] and bez of that during quantization extra layers are added which are  shape, strides slice layers and pack resulting and those operations are always int32 outputs. 

Refer - https://github.com/tensorflow/tensorflow/issues/43882

A simple input static shape set should be okay to solve this issue. 

Thanks, 

Sandeep Singh