Hello,
I work with the same model topology but with different weights and sometimes I run into a case when imported model has a huge accuracy drop.
Layer by layer comparison of such cases brought me to conclusion that such behavior is caused by malfunction of sigmoid layer since its min/max
ranges in affected models exceed expected one.
Here is a range for one affected model reported by tidl_model_import tool:
75 4.92151 0.00000 50.34475 0
Such range is a cause of significant differences between float and fixed versions of an imported model as shown in following diagram:
Further analysis has shown that integer underflow in reference sigmoid implementation causes the issue.
The underflow error occurs for certain negative input values and input scales.
The integer underflow takes place in TIDL_refSigmoidCore method at 396 line:
outVal = offset[0]*offsetScale - outVal;
where outVal = 1434, offset[0] = 130 and offsetScale = 11
and outVal is evaluated as following
outVal = slope[1] * inDataValAbs + offset[1]*offsetScale;
where slope[1] = 4, inDataValAbs = 56, and offset[1] = 110
I may suppose that treshold0 is set too high (inputScale = 11.250186) and it leads to wrong outVal calculation.
Introducing additional check for underflow before subtraction fixes the issue with wrong range while importing but
has minimal effect on device evaluation where I got following results:
We can see on diagram above that some kind of underflow error is still present and only magnitude of max values have
changed.
All tests were performed with rtos-j721e-evm-07_02_00_06 and tidl_j7_01_04_00_08.
I hope that this information will help to fix the issue.
Thanks.
