AM5728: How to output the feature_map of 12bit quantize

Happy_Bobo

Part Number: AM5728

Hi,

I have trained a model about face classfication with caffe-jacinto. I have test the accuracy of the model on pc, it is 95%.

I am using the TIDL tool to quantize the model with 8 bit, and then, I did the forward propagated on my device platform, the accuracy is 90%.

I think the accuracy loss too much, so I try to 12 bit quantize, but the problem is that I don't know how to output the feature map of 12bit quantize.

the 8 bit quantize, the feature map vector is -128 to 128, we could use the signed char* to point the output buffer.

how about 12bit, I think them are -2048 to 2048, so we try to use the short* to point the output buffer.

but the result is not right, the value of feature are exceed 2048, so I am confused.

thanks

best regards

over 5 years ago

0 Pavel Botev over 5 years ago

TI__Guru**** 170625 points

Hi Bobo,

Do you use AM572x TI EVM or custom board?

Do you use AM57x TI PSDK Linux? If yes, which version?

Regards,
Pavel

0 kumar.desappan over 5 years ago in reply to Pavel Botev

TI__Mastermind 22105 points

8- 12bit is suported for weights only.
for Faeture vectors only 8-bit is suported in TIDL

0 Happy_Bobo over 5 years ago in reply to kumar.desappan

Intellectual 325 points

Hi,
I'm found that the 12bit quantize will not improve the accuracy of my model, it is almost same with 8bit.
Does it mean that 12 bit is not enough of quantize the model ?
In my opinion, accuracy loss 5% is too much for some application though the compute time reduce much.
Why not the TIDL does't support 16 bit quantize, it's there some reason?
thanks
best regards
Bobo

0 kumar.desappan over 5 years ago in reply to Happy_Bobo

TI__Mastermind 22105 points

more than 8 bit for feature vectores would have performance drop.
for weights we can use upto 15 bits if the number channels used is less. (256 to 512). Otherwise there may be overflow in the accumulation during convolution. So we limit it to 12 bits. You can change the limit to 15 in import tool and try.

0 Happy_Bobo over 5 years ago in reply to kumar.desappan

Intellectual 325 points

Hi
thanks for your reply, if I want to change the limit to 15 bit, do I need to modify the import tool 's source code and rebuild it? is there any guidance doc? and where I can get the whole source code ?
thanks
best regards

0 kumar.desappan over 5 years ago in reply to Happy_Bobo

TI__Mastermind 22105 points

Refer user guide for building tool. Import tool source code is part of object release (modules\ti_dl\utils). Update the weight bit depth limi to 15 in tidl_import_main.cpp.

Processors

Processors forum

AM5728: How to output the feature_map of 12bit quantize