**** BEGIN LOGGING AT Wed Mar 10 02:59:56 2021 **** BEGIN LOGGING AT Wed Mar 10 07:45:52 2021 Mar 10 16:30:38 This is where we meet, I presume? Mar 10 16:36:17 I ran OPENGL ES on my RPI4 and implemented a few codes.I am sharing the link of the doc.It would be very helpful if someone can tell me what can add on in order to accelerate maximum layers using the Darknet CNN framework for the project OpenGLES acceleration for DL.I will also try to do the same things with Beaglebone AI as well. Mar 10 16:36:17 https://docs.google.com/document/d/1EdPm3CJwF-mw2t6qzoCcGpCXi0hWQZkA6ZBNzK8NA-Q/edit?usp=sharing Mar 10 16:37:43 * I ran OPENGL ES on my RPI4 and implemented a few codes.I am sharing the link of the doc. It would be very helpful if someone can tell me what can add on in order to accelerate maximum layers using the Darknet CNN framework for the project OpenGLES acceleration for DL.I will also try to do the same things with Beaglebone AI as well. Mar 10 16:37:43 https://docs.google.com/document/d/1EdPm3CJwF-mw2t6qzoCcGpCXi0hWQZkA6ZBNzK8NA-Q/edit?usp=sharing Mar 10 16:37:58 * I ran OPENGL ES on my RPI4 and implemented a few codes.I am sharing the link of the doc. It would be very helpful if someone can tell me what can I add on in order to accelerate maximum layers using the Darknet CNN framework for the project OpenGLES acceleration for DL.I will also try to do the same things with Beaglebone AI as well. Mar 10 16:37:58 https://docs.google.com/document/d/1EdPm3CJwF-mw2t6qzoCcGpCXi0hWQZkA6ZBNzK8NA-Q/edit?usp=sharing Mar 10 17:09:24 Hello there, I'm looking for someone who can help me with the OpenGS Acceleration for DL idea. I was looking into it and I have a fair idea about it. I have a raspberry pi 3b and it can run Tiny-YOLOv3 but not very well(2-3FPS). Mar 10 17:09:44 So I have some questions regarding that. Mar 10 17:10:59 1) Can I go ahead with the development on the Rpi3b board? Mar 10 17:11:27 2. Is there a specific approach you'd like me to follow? Mar 10 17:12:06 Any help would be appreciated. Thanks :D Mar 10 17:14:10 The problem is the Rpi3b supports OpenGL ES 2.0 and not the latest ones AFAIK Mar 10 17:15:45 So I have been facing problems to look for a solution or dive more into it. Can anyone please help me? Mar 10 17:36:30 Hello everyone! Mar 10 17:36:45 hello @cwicks Mar 10 17:38:33 oh man.... totally didn't realize I'd been 'quieted' for the last hour. Mar 10 17:38:39 :-( Mar 10 17:38:45 I am an admin for GSoC w/ BeagleBoard. I'm not a coder, but have been a Beagler for 10+ years and love students and the BB.org community. I help organize the GSoC for us and generally help move things along. Mar 10 17:38:49 * jkridner_ waves to pdp7 Mar 10 17:39:04 oh no jkridner_ ! Mar 10 17:39:21 thats great ! Mar 10 17:39:23 Hello all! Mar 10 17:39:34 * jkridner_ is allergic to opening a bunch of Google docs from students. Any chance YadnikBendale[m] can put some thoughts down on elinux.org? Mar 10 17:39:39 hello @sakethr98 Mar 10 17:39:49 Hello Cathy, Hello Jason! Mar 10 17:40:06 av500, lorforlinux[m], ds2, dinuxbg, mdp, moto-timo, nerdboy, sakethr98[m], pratimu[m], vedant16[m], zeekhuge[m]: hope none of you missed the news that we've been accepted as a mentoring organization this year! Mar 10 17:40:10 Hello Arsh! Mar 10 17:40:21 We have until Mar 29 before students start sending in applications.... Mar 10 17:40:31 now is the time to discuss project ideas! Mar 10 17:40:36 * jkridner_ looks around for Bela folks. Mar 10 17:40:46 would be good if we had something webthings and pocketbeagle grove kit related also. Mar 10 17:40:51 something with gcc on the c6x would be good too. Mar 10 17:40:58 more around the cape compatibility layer now with the new 'extension' command would be good. perhaps introducing it to Pi and BeagleV. Mar 10 17:41:12 I saw the mailing list.! Looking forward to one more year of GSOC with great! Mar 10 17:41:17 bpruthvidhar[m]: Has anyone looked into using TF-Lite on it? I'm worried that might be a bit like moving backwards from something optimized in a different way--as in, starting with a TF model might be more productive if using TF-lite and you are using YOLO because of its optimizations? Mar 10 17:41:21 * I saw the mailing list.! Looking forward to one more year of GSOC with some nice projects Mar 10 17:41:24 ds2 likely knows more about YOLO. Mar 10 17:41:28 Hello Saketha! Mar 10 17:41:32 OK, so YOLO uses OpenGL? What other ways can you accelerate? Mar 10 17:41:39 It could be interesting to look at it on NVDLA on BeagleV. Mar 10 17:41:44 jkridner: i had tried it Mar 10 17:41:44 Today's BeagleBone AI *might* work with TF-lite, but the integration could be a lot. TIDL is the main way to do neural networks on BeagleBone AI today. Mar 10 17:41:51 Perhaps looking at TIDL and/or OpenCL on C6x is reasonable? I just think TF-lite would have more of a roadmap. Mar 10 17:41:51 Hello Vedant! Mar 10 17:41:58 I know even less about the NVDLA tools. Mar 10 17:42:04 bpruthvidhar[m]: I find that looking at http://www.catb.org/~esr/faqs/smart-questions.html is a great way to figure out how to ask for help. Mar 10 17:42:09 Hello again Cathy! Mar 10 17:42:12 I often summarize as: 1) what are you trying to accomplish, 2) what did you research to try to accomplish it, 3) what did you try, 4) what did you expect, 5) what actually happened. Mar 10 17:42:38 I and lorforlinux were trying to comeup with a new idea for projects page Mar 10 17:42:46 the performance improvements wasnt that great tho Mar 10 17:43:00 😀 Mar 10 17:43:08 reading all up on TF-lite, NVDLA, TIDL and OpenCL could be a starting point, depending on what you are trying to accomplish. Mar 10 17:43:16 OK, done with old comments now. :-) Mar 10 17:44:01 sakethr98[m]: sooner the better as students are looking for projects now. Mar 10 17:44:09 yes Mar 10 17:44:11 we have a baseline Mar 10 17:44:23 the project is about working on audio Mar 10 17:44:43 https://www.survivingwithandroid.com/esp32-tensorflow-micro-speech-i2s-external-microphone/ Mar 10 17:44:44 meanwhile are there any beginner level issues that I can help with Mar 10 17:44:53 have you looked at Bela? it should be possible to run the Bela code on the PocketBeagle Grove Kit. Mar 10 17:45:37 I will start working on it Mar 10 17:46:51 Hi all, excited for GSoC 2021! Mar 10 17:47:07 Hi @Mentors - please make sure your info is updated on the bottom of this page. https://elinux.org/BeagleBoard/GSoC/Ideas-2021 Mar 10 17:47:35 Hi Pratim! Mar 10 17:48:26 @cwicks[m]: mine looks ok Mar 10 17:48:27 Hi Cathy, nice to reconnect Mar 10 17:48:48 My info is updated as well Mar 10 17:48:49 > * <@freenode_jkridner_:matrix.org> is allergic to opening a bunch of Google docs from students. Any chance Yadnik Bendale can put some thoughts down on elinux.org? Mar 10 17:48:49 Yes sure Mar 10 17:50:40 jkridner_: https://bela.io/ is this the one Mar 10 17:51:24 @cwicks[m] my info is good too Mar 10 17:52:49 jkridner_: is it a good idea to get the pru dma project running again? I had added it to the ideas page Mar 10 17:53:00 any specific focus you want to add? Mar 10 17:53:10 we will try and have it on the page by the next 48 hours Mar 10 17:54:12 Sorry, just did Mar 10 17:54:24 Anyone thinking of a project with the new BeagleV ? Mar 10 17:54:24 > <@cwicks:matrix.org> any specific focus you want to add? Mar 10 17:54:24 * Sorry, I just did change that Mar 10 17:56:03 I actually was interested in working on the Beagle AI but definitely will try to look in this direction Mar 10 17:58:26 btw I have been reading the tidl docs and was trying to setup the sdk for emulation Mar 10 17:58:43 i had some issues with that..could someone help me with it Mar 10 17:59:44 sure, what are the issues that you ran into? Mar 10 17:59:55 BeagleBone AI is a great platform. The mentors here can help you decide which may be best suited to your interests/project idea. Mar 10 18:01:15 I was watching there tidl tutorials and they were focusing on installing the jacinto-7 sdk Mar 10 18:02:05 but our target hardware is beaglebone -AI ,I was confused between the exact sdk to use Mar 10 18:02:46 i have downloaded the sdks for jacinto-7 and TI-am57x sdks Mar 10 18:03:07 good to know Mar 10 18:03:47 let me do a quick search and get back. *checking gh Mar 10 18:03:55 cool Mar 10 18:07:28 given the BBAI has th Sitara AM5729 you can go ahead and use TI-am57x sdk Mar 10 18:07:36 https://training.ti.com/sites/default/files/docs/am57x_ti_deep_learning_overview.pdf. Mar 10 18:07:37 ok Mar 10 18:07:42 * https://training.ti.com/sites/default/files/docs/am57x_ti_deep_learning_overview.pdf Mar 10 18:07:58 more on BBAI's techspecs here https://beagleboard.org/ai Mar 10 18:09:36 I was actually looking forward to generalize the task that would be required to port models(Yolo in our case) on beagle boards as in future we could start working on it Mar 10 18:10:06 * I was actually looking forward to generalize the task that would be required to port models(Yolo in our case) on beagle boards as in future we could start working on newer models(maybe RetinaNet I guess) Mar 10 18:12:45 My current ideas are to use Tidl apis in a similar way to their ssd-multibox example(two frame buffering) and also OpenCL that was pointed out in previous meeting Mar 10 18:13:11 any other areas/directions that I should be looking for? Mar 10 18:13:46 by generalising you mean, trying to understand? Mar 10 18:14:10 or to make a recipe that can be followed for different models? Mar 10 18:14:33 this one I guess Mar 10 18:15:39 you can look at integrating tidl/opencl with an acceleration library (this is actually on the ideas page) Mar 10 18:15:53 * you can look at integrating tidl/Open CL with an acceleration library (this is actually on the ideas page) Mar 10 18:16:12 my plan was to check performance increase in simpler models first before going on yolo models Mar 10 18:16:38 yes, makes sense. Mar 10 18:16:43 yeah just saw that right now Mar 10 18:17:47 it's not like you have to do what's exactly on the ideas page, you can take it as a direction/guide to form your own version of it and discuss with us here Mar 10 18:19:11 what are you thinking for Retina Net Mar 10 18:20:52 actually in some previous projects I was introduced to retina-net. It definitely was much more accurate model than Yolo models but a little slower..I was thinking about trying to port it on Beagle -AI/X115 Mar 10 18:20:58 * actually in some previous projects I was introduced to retina-net. It definitely was much more accurate model than Yolo models but a little slower..I was thinking about trying to port it on Beagle -AI/X15 Mar 10 18:21:14 Hi cwicks I have added mentor ifo Mar 10 18:21:18 * Hi cwicks I have added mentor info Mar 10 18:23:22 thats why I was thinking on generalizing this task(porting models ,if possible) Mar 10 18:23:59 for that, you will have to port a model first, correct? Mar 10 18:24:16 yeah Mar 10 18:24:44 so are you thinking to do it for Retina Net first? Mar 10 18:25:49 as I said my first target is to check on smaller and simpler models(probably during this time) then start working on Retina-Net(yolo models) Mar 10 18:27:35 Just a reminder to Mentors about the schedule. We use the next couple of weeks or so to hang out/check back here as much as you are able to encourage students who drop in for ideas and questions for the application process. Getting more discussion going sooner will help more students apply easily. Mar 10 18:28:58 how are you planning to do that? Mar 10 18:29:34 as in do you have any specific directions or do you need some pointers Mar 10 18:31:29 Reminder to Students - all questions are good questions. Come here often and engage, read the project ideas and discuss. Put your own together or a twist on one you see on the list. The best students we've had in the past 10 seasons of GSoC engaged early and often with the mentor team. Mar 10 18:32:18 I have added this statement, if any student is into compilers and PRU https://elinux.org/BeagleBoard/GSoC/Ideas-2021#Add_features_to_simpPRU Mar 10 18:34:39 I mean I have some ideas as to how to proceed with the task but yeah pointers from mentors would definitely be great Mar 10 18:36:21 my major targets were to check performance imporvements on smaller cnn models(dog/cat classifier would be a good starting point) then move on to customised vgg-net models Mar 10 18:37:19 my first plan is to use tidl apis and import the .bin files(courtsey of tidl importer) and then to organize the inference in a way similar to open-ssd Mar 10 18:38:06 then check performance improvements using open-cl(I still have to read on this) Mar 10 18:38:26 and then repeat this for complex models Mar 10 18:41:33 @cw Mar 10 18:42:10 * I will see if I can find some examples in terms of porting and share them here Mar 10 18:42:21 that would be great Mar 10 18:43:45 I just think that a having good support for several cnn models on the embedded side(like boasting a good performance) would definitely help the community Mar 10 18:44:00 Mentors - reach out to student groups /communities you know of and let them know the good news that GSoC21 is starting and Beagle is the community to engage with! Mar 10 18:45:06 yess!! started doing so with groups at my alma mater Mar 10 18:45:33 cwicks: can you share a template mail that can be shared ? Mar 11 01:34:26 testing beaglebot **** ENDING LOGGING AT Thu Mar 11 02:59:57 2021