**** BEGIN LOGGING AT Fri Mar 20 02:59:57 2020 Mar 20 07:19:45 Hello Everyone Mar 20 07:19:45 I am final year undergrad at Panjab University, Chandigarh.I do want to participate in GSOC 2020 through beagleboard. Please help me through the process. Mar 20 07:41:12 Hi Abhishek_ Mar 20 07:41:12 Please can you help Mar 20 09:14:43 @rollno0111:matrix.org: Hi Mar 20 09:14:43 Did you visit the elinux page to get familiar with the ideas and prpcess of application? Mar 20 09:14:44 https://elinux.org/BeagleBoard/GSoC/Ideas Mar 20 09:16:51 @freenode_ds2:matrix.org: Mar 20 09:16:51 The model training issue is resolved. Mar 20 09:16:51 I converted and tested the YOLO v2-tiny model (trained on Coco dataset) and that Kills the need to train another model :) Mar 20 14:08:50 rollno0111: sorry it is frequent our time is very shifted. Mar 20 14:10:04 rollno0111: have you looked over http://bbb.io/gsocideas ? Mar 20 14:10:21 the topic for the room is meant to help you get started. Mar 20 14:12:17 ymdatta: are you getting any closer to understanding how to build a bus driver and how the platform data structure needs to be completed for a successful probe on various device drivers? Starting with IIO I2C/SPI devices is the best way to get a feel for it. Mar 20 14:53:43 they X15 doesn't have any cape compatible header pins. Mar 20 14:55:13 probably doesn't. did it? Mar 20 14:55:34 ahoh Mar 20 14:55:36 ahoy Mar 20 14:56:45 ymdatta: cfriedt should be of some help! Mar 20 14:57:22 cfriedt: https://en.wikipedia.org/wiki/Ahoy_(greeting)#Czech_and_Slovak Mar 20 14:59:21 Hi cfriedt: I am getting an idea on how to go about implementing this https://elinux.org/BeagleBoard/GSoC/Ideas#mikroBus_driver Mar 20 15:02:40 ymdatta: Cool. I had an idea for a slightly different approach a while back, and ran it past gregkh and jkridner. That would be, supporting device tree fragments inside of the greybus Manifest. That would allow you to say, e.g. gpio N is the IRQ for i2c0. In any case though, the greybus Manifest makes traditionally non-discoverable buses Mar 20 15:02:41 discoverable. Mar 20 15:03:14 I have to run out for about the next hour, but I'll be back in a bit Mar 20 15:04:00 cfriedt: Ok, should i post some questions here? Or can you ping me once you get back? Mar 20 15:04:48 jkridner: I think I might need you to do some admin stuff to add me as a GSoC mentor Mar 20 15:46:48 cfriedt: How's that different from having device tree overlays? Mar 20 16:47:31 jkridner: I have looked at patchset for introducing new bus for serially attached devices (serdev) Mar 20 16:57:18 > ymdatta: are you getting any closer to understanding how to build a bus driver and how the platform data structure needs to be completed for a successful probe on various device drivers? Starting with IIO I2C/SPI devices is the best way to get a feel for it. Mar 20 16:57:19 jkridner: we can have a file pointer which points to parsed contents of data from EEPROM, before probing the bus driver uses the pointer to get appropriate information and fills platform data structures. So, that at the time of actual driver probing of device, the device has all the information needed to offer to driver. Mar 20 18:42:48 * pradan[m] uploaded an image: image.png (88KB) < https://matrix.org/_matrix/media/r0/download/matrix.org/UooaPUHoYaqiBOCjEfDBBJIx > Mar 20 18:43:08 ds2: I hope you saw my last update this morning. The frozen model (.pb file) after the conversion of YOLO v2-tiny is about 48MB and here's the output. Mar 20 18:46:38 Now, focusing on the accelerator part... Mar 20 18:55:32 Now the remaining problem of getting a faster inference is to be solved in 2 subparts: 1) Shrink model by in introducing Sparsity and Quantizing the model Mar 20 18:55:32 2) Distributing frames AND network layers among the 4 EVEs and 2 DSPs. Mar 20 18:57:41 Now, on going through another doc on TIDL, I found the method of doing (1) part...... the configuration file to be the sole method of doing so and the functions `quantizationStyle` and `Conv2dKernelType` are capable of doing so... Mar 20 18:57:41 Now, thinking of the (2) part. Mar 20 18:59:42 ds2: Can you share what problems you had with the config file... maybe I'll look into that. or if I should think another way 🤓 Mar 20 19:08:12 regarding this one, I remember in my first idea app (related to the poor-support-yet_state-of-the-art-library FeatherCNN), the binary file was obtained... I shared my results of mobilenet model on 64-bit ARM VM, but the tests on actual HW couldn't be done... I wonder if you would like to look into that **** ENDING LOGGING AT Sat Mar 21 02:59:57 2020