**** BEGIN LOGGING AT Mon Mar 16 02:59:58 2020
Mar 16 04:58:29 <pradan[m]> I think I now have a clearer understanding of TIDL and have started the work on the other proposal.
Mar 16 04:58:29 <pradan[m]> Its just some more clarification that I need before it becomes convincing for everyone :)
Mar 16 05:02:00 <pradan[m]> Till now, some of the validated observation are:
Mar 16 05:02:00 <pradan[m]> - We will be using YOLO v2 variant with TIDL for utilising all 4 (2EVE + 2DSP) of the BB AI to reduce the inference time to < 1s
Mar 16 05:03:58 <pradan[m]> Only v2 has the layers currently supported by TIDL.... in v3, all layers aren't supported.
Mar 16 05:08:32 <pradan[m]> Also, the fact that the accelerator are capable of reducing the overhead per core lies in the ssd demo provided in the API., where they manage to obtain results in 458ms with custom ssd model (preferably trained on Caffe-Jacinto)
Mar 16 06:00:20 <ds2> pradan[m]: I have heard some v3 layers are support but the problem may be memory requirements preclude it
Mar 16 06:24:36 <pradan[m]> i meant the tiny versions of both v2 (33mb) and v3 (35mb)
Mar 16 06:26:26 <pradan[m]> Regarding NNPACK, could you tell me whether there is any difference in a NNPACK enabled model
Mar 16 06:27:51 <pradan[m]> Regarding NNPACK, could you tell me whether there is any difference in a YOLO model and a NNPACK enabled YOLO model...
Mar 16 06:28:07 <pradan[m]> I suppose NNPACK kicks in at the inference time and both the models should be identical black boxes
Mar 16 06:32:45 <pradan[m]> Oh! Sorry but I forgot to mention that everything I am stating with respect yo TIDL, is exclusively for the AM5729 powered BB AI
Mar 16 16:33:39 <pradan[m]> @ds2 What do you say about training a SSD model (identical to YOLO) using the native Caffe-Jacinto framework ?
Mar 16 19:03:13 <pradan[m]> I mean to leverage the optimisations provided by TIDL, and I am sure we can do it when starting with a model from scratch (i.e., using Caffe-Jacinto. If you agree, I can move on to elaborating the optimisations with the custom-YOLO model that we could build.
Mar 16 19:03:13 <pradan[m]>  
Mar 16 19:06:04 <pradan[m]> What we can't be very sure about is that, whether we still get to use the advantages  (of TIDL) if we port the original YOLO v2 model... although I am trying my best to work out the feasibility as it would simplify our task
**** ENDING LOGGING AT Tue Mar 17 02:59:57 2020