Hi,
As per our previous discussion (https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1042985/tda4vm-query-regarding-centernet-model-porting), we have implemented centernet post processing into TIOVX node and executed the same in C66. We dequantized tensor heads by dividing it with corresponding scale values.
We could get the output coordinates and confidence score. But we could find some deviation in the final outputs.
So we compared the following tensor values and we found that the two files are different. We have given the same inputs to both #1 and #2
1. Dumped TIDL tensor outputs for INT 8 model from TIDL testbench using TIDL RT inference (PC_dsp_test_dl_algo.out)
2. Dumped intermediate tensor outputs from TIOVX Object detection application with the INT8 model, and dequantized it by dividing with the corresponding scale values.
We have some queries. Can you please provide us your inputs on this
a) Our understanding is that both #1 and #2 should be the same. Can you please confirm
b) If so, any inputs on what could have caused the deviation?
c) Is our dequantization logic the same which Ti expects/or has implemented in #1?
regards,
Gina
