TDA2EVM5777: Caffe jacinto quantization API for model evaluation on computers

WEI CHIH CHERN

Part Number: TDA2EVM5777

Hello,

Since the provided testing python script for caffe jacinto trained SSD models doesn't support quantization. I wonder does TI or anywhere else has the python API to quantize models for accuracy testing?

The testing we are doing is not just hit or miss, but evaluating the offsets of accuracy in cm compare to ground truth.

Thanks,

Wei Chih

over 5 years ago

0 Manu Mathew over 5 years ago

TI__Genius 11236 points

Hi Wei Chih,

We do not support quantizing the models in Caffe - but you can do that directly in TIDL for TDA2x. But if your purpose is just to get an approximate sense of accuracy, we have it in our scripts.

For example, this script: https://git.ti.com/cgit/jacinto-ai-devkit/caffe-jacinto-models/tree/scripts/train_imagenet_classification.sh

has a last phase called test_quantize that reports rough accuracy with quantization. An example config file and log is captured here:

https://git.ti.com/cgit/jacinto-ai-devkit/caffe-jacinto-models/tree/trained/image_classification/imagenet_jacintonet11v2/test_quantize

But this has not been matched properly with what TIDL produces and may not always be correct.

Your best bet is to measure the accuracy using TIDL.

Best regards,

Processors

Processors forum

TDA2EVM5777: Caffe jacinto quantization API for model evaluation on computers