AM5749: Model customisability in Model import feature vs that in Jacinto models

Prashant Dandriyal

Part Number: AM5749

Hi,

I wish to use the model optimization features offered by TIDL like Sparsity change, FP quantization, Grouped layer technique, distributing layers overload between EVE and DSPs.

The options are definitely available if I train a model using TIDL + Caffe Jacinto.

But do I get the options if I port a Caffe model using TIDL instead ?

Regards.

over 5 years ago

0 Paula Carrillo over 5 years ago

TI__Mastermind 39170 points

Hi Prashant, TIDL import tool has some optimization build-in, for example during import, some of layers are be coalesced into one TIDL Layer (e.g. convolution and ReLU layer). This is done to further leverage EVE architecture which allows certain operations for free. However, this optimizations are not configurable options. I guess you could do it, but it would take you to modify the import tool..

thank you,

Paula

0 Prashant Dandriyal over 5 years ago in reply to Paula Carrillo

Intellectual 790 points

Hi Paula!

Thanks for getting in touch...

I intend to use the other more beneficial features like the sparsification of models, grouping of layers, etc

Do the import process allow these ?

Also, how feasible is it to modify the import tool ? Is it open-source and are there any instructions about it ?

0 Paula Carrillo over 5 years ago in reply to Prashant Dandriyal

TI__Mastermind 39170 points

Hi Prashant, import tool is distributed as a binary. As I mentioned some fusion layer happens automatically, and you could also, via configuration file, group layer to run in difference cores (EVE/DSP).

http://downloads.ti.com/mctools/esd/docs/tidl-api/api.html#overriding-layer-group-assignment

Hope this helps.

Paula

Processors

Processors forum

AM5749: Model customisability in Model import feature vs that in Jacinto models