TDA4AL-Q1: TDA4AL-Q1: TIDL GolbalAvgPool results are not as expected

Tianze Guo

Part Number: TDA4AL-Q1

Tool/software:

Source Data:

In the upper layer of the Relu layer, the inference results of the float stream of the onnx model are compared with the inference results of the pc-import int16 stream.

Not expected result:

In the subsequent GolbalAvgPool layer, the inference results of the float stream of the onnx model are compared with the inference results of the pc-import int16 stream

Layer Info:

psdk version:
psdk-09-00-00-45-j784s4

4 months ago

0 Asha Bhandarkar 4 months ago

TI__Genius 10150 points

Hi Tianze,

Thank you for the information you provided. To summarize the behavior you are seeing

1. FP32 inference and 16-bit fixed point inference are providing same results up until ReLU layer when observing the layer level traces

2. The next layer (Global Average Pooling) is showing unexpected results when comparing the same

To clarify are you using TIDL-RT float with c7x-mma-tidl? Is it possible for you to verify the results with the latest 9.2 SDK?

Best,

Asha

0 Tianze Guo 4 months ago in reply to Asha Bhandarkar

Prodigy 40 points

What does this TIDL-RT float with c7x-mma-tidl mean?
The int16 output result I provided is/ The output product of './out/tidl-model_iimport.out'

0 Tianze Guo 4 months ago in reply to Tianze Guo

Prodigy 40 points

The trajectory chart of int16 is produced by multiplying the .y product from ./out/tidl_model_import.out with the scale factor from out/tidl_model_import.out's layerinfo.txt product.

The trajectory chart of float is provided by the results of onnx.load(), onnxruntime.InferenceSession(), and session.run().

0 Asha Bhandarkar 4 months ago in reply to Tianze Guo

TI__Genius 10150 points

Hi Tianze,

I was trying to clarify if you were building with TIDL-RT flow (TIDL support packaged within the SDK) versus edgeai-tidl-tools which we host on Github. Based on your clarifications you are using the first one.

Tianze Guo said:
The trajectory chart of float is provided by the results of onnx.load(), onnxruntime.InferenceSession(), and session.run().

Is setting numParamBits = 32 on host emulation providing you the same results?

If possible it would helpful to run the same on 9.2 SDK and see if that has improvements.

Also is it possible for you to provide the model, model artifacts, and the configuration file you are using to compile the model so I can reproduce the results on my end and debug further?

Best,

Asha

0 Tianze Guo 4 months ago in reply to Asha Bhandarkar

Prodigy 40 points

Downloads.zip

0 Tianze Guo 4 months ago in reply to Tianze Guo

Prodigy 40 points

Hi Asha We have provided the input and output data along with the results at each layer for further analysi. Best Tianze

0 Suman Anna 4 months ago in reply to Tianze Guo

TI__Guru* 94825 points

Hi Tianze,

Asha is on vacation for the rest of this week, so please expect a delay in responses. Asha will get back to you sometime early-mid next week.

regards

Suman

0 Tianze Guo 4 months ago in reply to Suman Anna

Prodigy 40 points

I'm just counting down the days 'til the weekend myself, so no rush on the work front.

0 Suman Anna 4 months ago in reply to Tianze Guo

TI__Guru* 94825 points

Sure, no issues :)

0 Asha Bhandarkar 4 months ago in reply to Suman Anna

TI__Genius 10150 points

Hi Tianze,

Thank you for your patience while I was on vacation!

I have looked at your files you have provided so I can see if I can reproduce your issue to debug further. Would you be able to provide an input image that you've tested this on?

Best,

Asha

0 Tianze Guo 4 months ago

Prodigy 40 points

Hi Asha,
The segment of the network I have captured is from the middle part, which does not accept images as input, please use output file from the first layer 'inDataLayer0' from the export path as input.
Best,
Tianze