**** BEGIN LOGGING AT Tue Mar 03 02:59:59 2020
Mar 03 04:45:10 <Pac23> anyone willing to be the "official" on paper mentor for the pocketbeagle stepwise mathematics task ?
Mar 03 09:10:15 <gphofficial[m]> Hi can i just check if this is maintained ? http://beagleboard.org/linux
Mar 03 09:11:02 <gphofficial[m]> is there a way to submit PR
Mar 03 09:11:46 <gphofficial[m]> nvm found it
Mar 03 09:11:46 <gphofficial[m]> http://beagleboard.org/
Mar 03 09:11:52 <gphofficial[m]> https://github.com/beagleboard/beagleboard-org
Mar 03 14:51:41 <pradan[m]> Hello, a general question: Is this the ONLY IRC channel for GSoC or do we find mentors on another one?
Mar 03 14:53:11 <pradan[m]> [resource] The IRC nicknames for (past) mentors can be found here: https://elinux.org/BeagleBoard/GSoC/Ideas-2018#Mentors
Mar 03 17:26:59 * pradan[m] sent a long message:  < https://matrix.org/_matrix/media/r0/download/matrix.org/IyVeQnZDDaquHfHyHKJejDGQ >
Mar 03 20:59:26 <pradan[m]> @ds2 For the **Running YOLO model on x15** project, I can only use the simulator for A5749 or AM57xx target as I can't avail the board currently. This means that out of BBB, BBAI and X15, only X15 can be properly tested without the actual EVM. Am I right?
Mar 03 20:59:26 <pradan[m]> Although we can also deduce similar change in results for other 2 boards also.
**** BEGIN LOGGING AT Tue Mar 03 21:04:35 2020
Mar 03 21:07:21 <ds2> pradan[m]: what are you trying to do right now?
Mar 03 21:07:43 * pradan[m] sent a long message:  < https://matrix.org/_matrix/media/r0/download/matrix.org/BQgOrlXozuFkNeSwstusnJPG >
Mar 03 21:08:20 * pradan[m] sent a long message:  < https://matrix.org/_matrix/media/r0/download/matrix.org/ltEptSiJIdeVeNSVkxOPJKUW >
Mar 03 21:08:21 <ds2> but what are you trying to accomplish?
Mar 03 21:08:30 <ds2> trying to figure out why you are trying the simulator?
Mar 03 21:09:10 <ds2> also, can you talk about how do you want to approach it. please use IRC messages, not the matrix extensions
Mar 03 21:09:20 <ds2> what I am seeing: * pradan[m] sent a long message:  <
Mar 03 21:09:21 <ds2> +https://matrix.org/_matrix/media/r0/download/matrix.org/ltEptSiJIdeVeNSVkxOPJKUW >
Mar 03 21:09:43 <pradan[m]> I am trying to derive methods to fulfill the objectives mentioned in the project description.  Oh, sorry but I didn't know about IRC ...
Mar 03 21:09:58 <ds2> that's cool.. that's why I pasted what Isee :D
Mar 03 21:10:14 <ds2> can you tell me how much do you understand about the project?
Mar 03 21:10:30 <ds2> I am not understanding why a simulator is the first step
Mar 03 21:10:58 <pradan[m]> Should I continue here ?
Mar 03 21:11:07 <ds2> yes
Mar 03 21:11:28 <ds2> by doing it here, other potential mentors can see it
Mar 03 21:11:42 <ds2> we (beagleboard.org) generally like to have morethen 1 mentor per project
Mar 03 21:13:16 <ds2> to avoid that matrix extension, use shorter messages
Mar 03 21:14:40 <pradan[m]> Ok thank you for letting me know.
Mar 03 21:14:40 <pradan[m]> I am trying to create a set-up for the experimentation that I'll be doing in course of this project. I know that tests will be **after** I have a solution but what I felt was, what use would the solution be if I can't test its validity ?
Mar 03 21:14:40 <pradan[m]> Hence, I went through the TIDL docs and tried to set up the API and install the simulator... but still at it.
Mar 03 21:16:02 <ds2> unfortunately, I don't think any thing other then HW will work with TIDL
Mar 03 21:16:29 <ds2> TIDL uses both EVE and the DSP... AFAIK, EVE is somewhat TI specific (EVE is the hardware block that accelerates video decoding/encoding)
Mar 03 21:16:52 <ds2> you can clone the git repo for TIDL and look at the examples
Mar 03 21:18:25 <ds2> how are you thinking of approaching it? my email was somewhat high level... there may be better ways
Mar 03 21:20:19 <pradan[m]> Actually I wish to get a hang of it (the layer grouping and model comparisons). That's the primary step that I thought to go with.
Mar 03 21:20:19 <pradan[m]> I am sorry but I can't infer what you meant by "High level"...
Mar 03 21:20:47 <ds2> as in very rough idea/concepts
Mar 03 21:21:22 <ds2> I'd suggest you get familar with darknet first. that way you can see how YOLO as implemented by the authors work
Mar 03 21:21:56 <ds2> for the TIDL stuff, there may be other help we can get...
Mar 03 21:22:09 <ds2> what is your background?
Mar 03 21:23:00 <pradan[m]> Electronics and Communication... but learning more of Edge AI now.
Mar 03 21:23:21 <ds2> how familar are you with concepts like convolution?
Mar 03 21:24:21 <pradan[m]> With respect to DSP, I am quite familiar
Mar 03 21:24:44 <ds2> what do you mean? convolution isn't just for DSPs
Mar 03 21:24:54 <ds2> and is that 1D, or 2D convolution?
Mar 03 21:25:19 <pradan[m]> 1D till now
Mar 03 21:25:21 <ds2> or putting another way, have you had the equiv of "Signal and Systems" in your program?
Mar 03 21:25:46 <pradan[m]> Yes I had Signal & System before DSP
Mar 03 21:26:59 <ds2> that's good and bad... good is I think we are looking at it from the same direction...bad is most of this stuff is 2D conv
Mar 03 21:27:53 <ds2> just to get one other point out of the way... the idea that convolution in time is multiplication in freq holds true in the 2D version
Mar 03 21:28:24 <ds2> and there is a patch/enhancement to darknet that to implement that..that's what I use to get the 5-10sec per frame
Mar 03 21:28:35 <ds2> it is much worse w/o it
Mar 03 21:29:43 <ds2> not sure what else to suggest for getting more familiar with TIDL..prehaps another mentor can chime in if they have ideas
Mar 03 21:30:26 <pradan[m]> I was going to pick the smallest versions of YOLO (i.e., v2-tiny) and try optimize it using pruning or other high-level optimizaiton APIs like TF Lite...
Mar 03 21:30:40 <pradan[m]> But now, I will think over
Mar 03 21:31:41 <ds2> I haven't seen enough of TF Lite to comment
Mar 03 21:33:51 <pradan[m]> I saw its utility in Arduino board variant. They have ported it nicely and it generates a simple header file from the tflite model. this is what stores all the weights. But doing something (like creating a port) would be somewhat questionable in the 3 month period
Mar 03 21:34:22 <ds2> I see...
Mar 03 21:34:39 <ds2> there is suppose to be TFLite support for TIDL "very soon now"
Mar 03 21:39:06 <pradan[m]> That's great!
Mar 03 21:39:06 <pradan[m]> Currently ,I am enrolled in the Intel programme for its OpenVINO toolkit which is again for the same purpose like TFLite. I understand that TI and beaglebone have also made alternative to that : TIDL. So, rethinking over that wouldn't be affordable. I think I should go with the neural network optimization of YOLO variants, just as you suggested.
Mar 03 21:40:09 <ds2> none of them are easy
Mar 03 21:40:21 <ds2> I personally hate vendor specific solutions but if that will get things running...
Mar 03 21:41:01 <ds2> putting it another way, I can get multiple frames per second on a jetson nano
Mar 03 21:41:10 <ds2> (NVidia with CUDA)
Mar 03 21:41:45 <ds2> but the Jetson Nano sucks in other ways.. the TI chips are a bit more balanced
Mar 03 21:42:26 <ds2> if you want to look at the GLES approach, I have a framework to get it started.... so focus will be on how to break it up into shader code and how to get data to and from the GPU
Mar 03 21:42:46 <ds2> the caveat here is the GPU isn't exactly fast...
**** BEGIN LOGGING AT Tue Mar 03 22:18:27 2020
**** BEGIN LOGGING AT Tue Mar 03 22:50:35 2020
**** ENDING LOGGING AT Wed Mar 04 02:59:57 2020