TDA4VE-Q1: Why are weights recalculated in TIDL_QuantizeSignedMax

Henry2333

Part Number: TDA4VE-Q1
Other Parts Discussed in Thread: AM68, TDA4VM, AM67A, AM69A

Tool/software:

Hi, friends.

Using QDQ prequant onnx model to import, the result of TIDL is not good. After reading the tidl_tools code, I found that the quantization parameters have changed.

When isOutMaxSat == 1, the weights are recalculated in TIDL_QuantizeSignedMax.This will reduce the accuracy

1. Why TIDL_QuantizeSignedMax needed here?

2. In TDA4VE SDK0902, can I set isOutMaxSat to 0 to avoid weights recalculated ？and if so is there a risk ？

Best Regards
Henry

over 2 years ago

0 Pratik Kedar over 2 years ago

TI__Mastermind 24041 points

Hi Henry,

I am abstracted by implementation side offerings, the asked questions are directly falls under the same.

I will discuss this with dev team and get some insights on this, and try to get back to you by EOW.

Thanks

0 Pratik Kedar over 2 years ago in reply to Pratik Kedar

TI__Mastermind 24041 points

Hi,

For better accuracy you can try setting below flags

"advanced_options:prequantized_model" to be set to 1

"advanced_options:quantization_scale_type" to be set to 4 (asymmetric)

Let us know the results

0 Henry2333 over 2 years ago in reply to Pratik Kedar

Intellectual 500 points

Hi,Pratik

The results look better, but still not as good as the results from the QDQ ONNX model.

1. Is there any way to improve the accuracy?

2. advanced_options:quantization_scale_type=4 .This is not explained in the native_pytorch_quantization.md
.When should I use this parameter? Any device (AM68PA AM62A..)? Can it be used during PTQ?

Best Regards
Henry

0 Pratik Kedar over 2 years ago in reply to Henry2333

TI__Mastermind 24041 points

Henry,

Great to see that results are good with asymmetric quantization, AFAIK, only TDA4VM/AM68PA doesn't support asymmetric quantization for jacinto devices (AM68, AM68PA, AM69A, AM67A)

Let me reiterate that, Only the range values (min, max) fields are used to determine the final scales used by TIDL's quantization module, from the provided pre quantized model data.

I would encourage you to read about quantization protos to specify quantization params for each layer of your model manually.

Hope this helps here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_07_00/docs/tidl_quantParams.md

Processors

Processors forum

TDA4VE-Q1: Why are weights recalculated in TIDL_QuantizeSignedMax