Deep Learning on the BeagleBoard-X15

154 views
Skip to first unread message

Mark A. Yoder

unread,
Jan 9, 2019, 9:08:28 PM1/9/19
to BeagleBoard
It was recently pointed out to me that the BeagleBoard-X15 has hardware that supports Deep Learning and TI has already created several examples of how to use it.

I've created a wiki page that gives a quick guide for installing and running the examples.


All the examples are pretrained and the X15 is just running the inference engine.  It's been trained to recognize 1000 objects from a live video stream.
Using a simple webcam, I've shown it several objects (tennis ball, baseball, coffee mug, beer bottle, etc.) and it has recognized them all.

I'm impressed.

Has anyone else played with this? What do you think?

--Mark

Items.pngWater_bottle.png


Calvin Slater

unread,
Jan 10, 2019, 4:40:06 PM1/10/19
to BeagleBoard
That's fantastic!

I was just wondering about this a couple weeks ago. 

I heard the AM5728 had TIDL support this whole time and uses the DSPs right?

Mark A. Yoder

unread,
Jan 16, 2019, 8:30:11 AM1/16/19
to BeagleBoard
My x15 has an AM5729 on it and it has 2 DSPs and 4 EVEs[1] (Embedded Vision Engines).  I don't know much about the EVEs, but I read somewhere that each EVE can do 16 multiply accumulates per clock cycle.  The tidl gives you some control over which processors (DSP or EVE) works on what part of the problem.

I'm often seeing 15 to 30 frames per second wile recognizing objects.

--Mark

Andy Laberge

unread,
Jan 16, 2019, 10:22:30 PM1/16/19
to beagl...@googlegroups.com
I built up another SD card for this project...got tidl_api and examples to compile....this set uses opencv 3.2.0....
my camera didn't kick in until segmentation and ssd_multibox .... also found out it is more responsive using -w 400 instead of -w 1200.

Will keep working with this...again thanks for the link Mark

Andy

--
For more options, visit http://beagleboard.org/discuss
---
You received this message because you are subscribed to the Google Groups "BeagleBoard" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beagleboard...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/beagleboard/76967aac-8360-46cf-af13-61b2236c7155%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Calvin Slater

unread,
Jan 25, 2019, 11:11:59 PM1/25/19
to BeagleBoard
Looks like I am a lucky dog as well. My x15 also has the industrial temp rated AM5729. This one was from the first public production run back in late 2017. I did some searching around and found the following:

1) The AM5729 is a non-catalog part, which is why there is no real public datasheet.

2) This chip has four EVEs but only one is enabled for the purpose of hacking/experimenting.

Is the above correct?

Jason Kridner

unread,
Jan 26, 2019, 12:25:52 PM1/26/19
to beagl...@googlegroups.com


On Jan 25, 2019, at 11:11 PM, Calvin Slater <robertpa...@gmail.com> wrote:

Looks like I am a lucky dog as well. My x15 also has the industrial temp rated AM5729. This one was from the first public production run back in late 2017. I did some searching around and found the following:

All Rev C boards used AM5729. The industrial part is a surprise to me. 


1) The AM5729 is a non-catalog part, which is why there is no real public datasheet.

Correct. Having 4 EVEs is the only difference from AM5728. I have hopes the AM5729 will be public some time this year. 


2) This chip has four EVEs but only one is enabled for the purpose of hacking/experimenting.

All 4 are enabled. 


Is the above correct?

On Wednesday, 16 January 2019 05:30:11 UTC-8, Mark A. Yoder wrote:
My x15 has an AM5729 on it and it has 2 DSPs and 4 EVEs[1] (Embedded Vision Engines).  I don't know much about the EVEs, but I read somewhere that each EVE can do 16 multiply accumulates per clock cycle.  The tidl gives you some control over which processors (DSP or EVE) works on what part of the problem.

I'm often seeing 15 to 30 frames per second wile recognizing objects.

--Mark

[1] http://processors.wiki.ti.com/index.php/EVE

On Thursday, January 10, 2019 at 4:40:06 PM UTC-5, Calvin Slater wrote:
That's fantastic!

I was just wondering about this a couple weeks ago. 

I heard the AM5728 had TIDL support this whole time and uses the DSPs right?



On Wednesday, 9 January 2019 18:08:28 UTC-8, Mark A. Yoder wrote:
It was recently pointed out to me that the BeagleBoard-X15 has hardware that supports Deep Learning and TI has already created several examples of how to use it.

I've created a wiki page that gives a quick guide for installing and running the examples.


All the examples are pretrained and the X15 is just running the inference engine.  It's been trained to recognize 1000 objects from a live video stream.
Using a simple webcam, I've shown it several objects (tennis ball, baseball, coffee mug, beer bottle, etc.) and it has recognized them all.

I'm impressed.

Has anyone else played with this? What do you think?

--Mark

Items.pngWater_bottle.png


--
For more options, visit http://beagleboard.org/discuss
---
You received this message because you are subscribed to the Google Groups "BeagleBoard" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beagleboard...@googlegroups.com.

Calvin Slater

unread,
Jan 26, 2019, 4:36:58 PM1/26/19
to BeagleBoard
Wow, the good news just keeps on rolling in!

Yes, the one I'm looking at has the BABCXEA part number. Basically the nicest version of this chip, with the -40~105C temperature range as well as the industrial protocols and EtherCAT stuff enabled.

A beagleboard to run your power plant with!

I'm continually amazed by the quality and features you all put into these boards.

Thanks again Jason
Reply all
Reply to author
Forward
0 new messages