Neural Networks/Deep Learning on BeagleBoard-x15


We are porting some deep learning frameworks to the BB-x15 to take advantage of the in-built DSP horsepower, and get some light inferences running. Although, since they utilize the TI provided DSP libraries like DSPLIB, FFTLIB, LINALG, etc, we are using the Processor SDK Linux as our platform.

We have so far had success with darknet/yolo (yolo tiny running at >2fps and yolo full at <0.5fps), dlib correlation trackers and are working on getting caffe up and running right now, for access to more neural networks from their model zoo. Not sure if we’d be able to reach tensorflow anytime soon, specially because we aren’t able to buy any more boards! But, we are very sure that we haven’t extracted even half the juice of each DSP, so there’s room for some great performance in inference.

I’m not sure if members of the community would be interested, but if they are, we would be open to collaborate, and after reviewing all possible TI documents, open source our changes for everyone out there.


So I’m very interested. No hardware yet, though.

Same problem that everyone’s facing, I guess! Mouser moved our order backdate by more than a month weeks after taking our order.

What kind of neural nets/fields of deep learning are you looking to target?

Caffe cross compilation is also almost done, with speed up from LINALG. Having to compile everything (right from cmake) from scratch is tedious, but thank goodness for open source.