[Python] Mismatch Between Quantized TFLite Layer Outputs and Expected Mathematical Values When...

I am trying to extract intermediate outputs from a quantized TFLite model using the TFLite interpreter. The goal is to verify that the model's intermediate outputs match the mathematically expected values( tensorflow version 2.6.2 , python version 3.6.8)

Steps I followed:

1. Loaded a quantized TFLite model using tf.lite.Interpreter.
2. Set the model input using: interpreter.set_tensor(input_index, quantized_input)
3. Invoked the interpreter: interpreter.invoke()
4. Retrieved op details: op_details = interpreter._get_ops_details()
5. For each layer, fetched the output tensor index: out_tensor_index = op_details["outputs"][0]
   [*]Got the tensor values: output = interpreter.get_tensor(out_tensor_index)
   

Saved the output of each layer as .npy files for further analysis.

Issue:

The output values obtained from interpreter.get_tensor() do not match the mathematically calculated (expected) values, even after considering quantization parameters (e.g., scale and zero-point).

Questions:

1. Is this the correct approach to extract intermediate layer outputs from a quantized TFLite model?
2. Are there any limitations with using interpreter._get_ops_details() and get_tensor() on quantized models?

Any help is welcome, let me know if more info is needed

Continue reading...