pota-quantization-value-test

pota-quantization-value-test checks whether a Circle model listed in test.lst is correctly quantized (pota denotes post-training affine). The circle models are generated from the recipes saved in res/TensorFlowLiteRecipes.

Write test.local.lst for local test.

Test process

Step 1. Fake quantization

Run circle-quantizer with --quantize_dequantize_weights option.

Dump the fake-quantized model with circle-tensordump.

Compare the dumped model with the expected output in "expected_outputs/<model_name>/<granularity>/<quantized_type>/fake_quantization/<tensor_name>.json"

The expected output should include (1) values of weights (only for conv, transposed_conv, depthwise_conv, and fc layers)

Step 2. Record moving avg of min and moving avg of max for activations

Run record-minmax with the fake-quantized model (input data is saved in "test_inputs/<model_name>/<granularity>/<quantized_type>/<record_number>.txt")

Dump the minmax-recorded model with circle-tensordump. Compare the dumped model with the expected output in "expected_outputs/<model_name>/<granularity>/<quantized_type>/record_minmax/<tensor_name>.json"

The expected output should include (1) min/max of activations

Step 3. Quantization

Run circle-quantizer with --quantize_with_minmax option.

Dump the quantized model with circle-tensordump.

Compare the dumped model with the expected output in "expected_outputs/<model_name>/<granularity>/<quantized_type>/quantization/<tensor_name>.json"

The expected output should include (1) scale, zero point of activations (2) scale, zero point, values of weights (3) scale, values (weights) of bias

Golden data

Golden data was generated as follows.

(1) Generate random h5 input for a target model (using gen_h5_random_inputs.py in record-minmax-conversion-test)

(2) Run dalgona with the target model, input data, and analysis code named GenGoldenWeights.py for uint8 (GenGoldenWeightsSym.py for int16) (Samsung#3501)

(3) Do fake quantization using circle-quantizer

(4) Run dalgona with the fake-quantized model, input data, and analysis code named GenGoldenActBias.py for uint8 (GenGoldenActBiasSym.py for int16) (Samsung#3501)

(5) Edit generated data for some operators (concat: scale propagation, mean: axis data)