SK-AM62A-LP: Inference accuracy on CPU and C7x

Hello,

This is an interesting observation. It does look like both the CPU and C7x respectively have some instances where they do better than the other. I see what you mean

Hasegawa Fumiya said:

I tried replacing PReLU nodes with ReLU nodes and running it.

This means that the model you are looking at one both CPU and C7x does not have any PreLU, correct? Can you confirm it is the same model? Did you quantize the model at all with tensor_bits setting (default is 8-bit)?

I am inclined to call this a coincidence. It is difficult to say what will happen deep within the model when something like an altered activation function has changed data distributions without being retrained. 

Hasegawa Fumiya said:

Calibration doesn't make the model more accurate than the original, right?

Correct. If it appears more accurate, it is likely random chance. Calibration to fixed-point is an inherently lossy process. Some information and precision is unavoidably lost during this process, and the goal during quantization is to minimize this with good calibration data (as well as a few other techniques). 

Perhaps some of that lost information was beneficial on a subset of your dataset. Ordinarily, that loss of information reduces accuracy, but since this model was modified without retraining, that change surely has some unintended effects that can appear beneficial.

I would not suggest using this model, but it is a good experiment -- the outputs are more reasonable than I had expected

BR,
Reese