micro: port op Reduce Prod from lite v2

[Evanok]

Hello,

I would like to port the Reduce Prod operation from lite to micro. It should be quite straightforward even if I am not very familiar with tensor operation because other similar reduce operation are already ported. There is already an issue about this porting but it seems to be abandon.

I already started the job but I am a little bit lost about the Quantized Reduce Prod function. First I did an implementation inspired from QuantizedMeanOrSum and ReduceSumImpl function. Then I realized there is already a function named QuantizedReduceProd in the reference reduce.h header: here

Should I used this existing function? Should I remove it and implement a new one based on the SumOrMean implementation?

Thanks, Arthur.

[Evanok]

First draw of the PR can be found here: PR #2532. TODO:

• Fix Quantized version which is not working properly.
• Add missing test to match the reduce SUM test.

[Evanok]

I still have issue with the Quantized reduce prod test. I write a very simple test based on the SumInt82DKeepDims in reduce_test.cc

kInputData2D is {1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0};

TF_LITE_MICRO_TEST(ProdInt82DKeepDims) {
  int8_t expected_output_data_quant[tflite::testing::kOutputElements2D];
  int8_t output_data_quant[tflite::testing::kOutputElements2D];
  int8_t input_data_quant[tflite::testing::kInputElements2D];

  float input_scale = 0.5f;
  int input_zero_point = 0;
  float output_scale = 0.5f;
  int output_zero_point = 0;

  TfLiteReducerParams params = {
      true  // keep_dims
  };

    tflite::testing::TestReduceOpQuantized<int8_t>(
      tflite::testing::kInputShape2D, tflite::testing::kInputData2D,
      input_data_quant, input_scale, input_zero_point,
      tflite::testing::kAxisShape2D, tflite::testing::kAxisData2D,
      tflite::testing::kOutputShape2D, tflite::testing::kGoldenDataProd2D,
      output_data_quant, expected_output_data_quant, output_scale,
      output_zero_point, tflite::Register_PROD(), &params, 1.0);
}

Then I have an error during the test: expected_output_data[i] (48.000000) near output_data[i] (127.000000) failed at tensorflow/lite/micro/kernels/reduce_test.cc:94

For me the output should be 127, 127. The expected output seems to be 48, 127.

From my understanding if I am computing the reduce prod of the input, I am quantified the input first. With a scaling of 0.5. the Quantified input is {2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0} So the result is

2 * 4 * 6 * 8 -> 384
10 * 12 * 14 * 16 -> 26880

Both output are greater than the max value for int8 (127). So the output should be 127, 127.

I have the same wrong result with my own QuantizedProd implementation and the QuantizedReduceProd code already present in the repo. sorry I am not very familiar with Quantized operation.

Did I missunderstand the quantized process?

[rascani]

Should I used this existing function? Should I remove it and implement a new one based on the SumOrMean implementation?

I would suggest re-using the QuantizedReduceProd in the reduce.h header.

I'll take a look at the test in the PR.

[Evanok]

Hello @rascani. I just pushed into the PR branch the problematic test. This is going to be easier to reproduce the issue! I am going to try on my side to find documentation on Quantization. I feel like the real problem is a lack of understanding on my side.

[github-actions[bot]]

"This issue is being marked as stale due to inactivity. Remove label or comment to prevent closure in 5 days."

[github-actions[bot]]

"This issue is being closed because it has been marked as stale for 5 days with no further activity."