Whats does "data_convert" layer do in compiled models from edge-AI?

Hi,

If you are using  asymmetric quantization style (applicable only for pre-quantized tflite models on all SOCs except AM68PA) then input or output format conversion are not required, data_convert needs to be set to 0 in this case. However if you are using default quantization style non power of 2, input and output data format conversion are required to set, thus data_convert is set to 3.

Hi Pratik,

Thanks for the reply. The model compiled by TIDL importer didn't have any data convert layer, what's reason of the difference here? I tried to inference the model using the bin. files output by TIDL importer on edge-AI, it seemed to performed without any error occured, will there be any issue in it?

Liu Tsing-Yue said:

The model compiled by TIDL importer didn't have any data convert layer, what's reason of the difference here?

Please revert with model compilation flags and SVG snap shot for the same.

This is the model SVG

1261.model_SVG.zip

Here's the config for edge-AI compilation

And here's the config for TIDL importer

If I understood your question correctly, you are not able to see dataconvert layer when following TIDL-RT flow (model import tool).

Can you check the TIDL-RT documentation here : https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-j721s2/08_05_00_11/exports/docs/tidl_j721s2_08_05_00_16/ti_dl/docs/user_guide_html/md_tidl_model_import.html

You can look for Config Pram For Format Conversion table for addDataConvertToNet flag.

Thanks! First time realized that addDataConvertToNet flag can be added into TIDL importer. However,  I could set data convert in edge-AI to 0, 1, and 3. For TIDL importer, I could only set to 0 and 1. Is it possible to add data convert at both input and output?

By the way, for edge-AI, I can't inference model when the tensor_bit=8 and data_convert=0. What might be the cause of this error?

Liu Tsing-Yue said:

By the way, for edge-AI, I can't inference model when the tensor_bit=8 and data_convert=0. What might be the cause of this error?

Per my understanding data convert needs be set if you are using power of 2 based quantization (default)