AM5728: Non-Square Caffe Input Layers with TI Deep Learning

Mark Swanson

Part Number: AM5728

Hi Team,

Background

My customer is interested in utilizing the TI Deep Learning library with non-square input layers with a Caffe input model. Currently, the model will successfully run through the conversion tool, but will not run on the AM5728EVM target.

Questions

How can non-square inputs be supported with Caffe models and TIDL?
- Are the Caffe parameters pad_h, pad_w, kernel_h, and kernel_w supported by TIDL?
Does TIDL handle preprocessing layers in the model which may be extracting an ROI where x, y, and theta locations are input parameters?

Best Regards,

Mark-

over 7 years ago

0 Manisha Agrawal over 7 years ago

TI__Mastermind 22946 points

Hi Mark,

Non square inputs should be supported by TIDL. Please let me what behavior do you see when you run the converted model on AM5728 EVM.

TIDL doesn't support pre-processing layers. They can be done outside TIDL, on the DSP or ARM core.

Regards,
Manisha

0 Manisha Agrawal over 7 years ago in reply to Manisha Agrawal

TI__Mastermind 22946 points

Hi Mark,

Regarding padding, it is not user configurable. It is handled internally by TIDL-API.

On the width and height question, I need clarification. Is this on image width and height or kernel width and height.

If you can share the network binary, parameter binary and configuration file, we can try to reproduce/analyze the problem at our end.

Regards,

Manisha

0 Mark Swanson over 7 years ago in reply to Manisha Agrawal

TI__Expert 3520 points

Hi Manisha,

Thank you for your replies!

I have requested further clarification on the height/width as well as access to their binary and configuration files to assist in reproducing. Once I have further details I will update the thread.

Best Regards,

Mark-

0 Manisha Agrawal over 7 years ago in reply to Mark Swanson

TI__Mastermind 22946 points

Hi Mark, I am closing this thread for now in absence of no response for a week. Feel free to re-open the thread once you hear from your customer.

Regards,
Manisha

Processors

Processors forum

AM5728: Non-Square Caffe Input Layers with TI Deep Learning