It will be slow!

It might be possible, but I doubt that it would be very fast. It will certainly not run in real time. And if you can make it run in real time you will probably have to modify the architecture pretty extensively. Especially given that these networks typically use a GPU to speed up computation. 

There are some approaches based on eigen and singular value decomposition of the hidden weights that can reduce the connectivity between layers with effectively the same information. I expect that utilising approaches like this might help speed things up. But I don't know if it will be enough. 

You can probably use a demo website to do all of the hard work for you (e.g. http://demo.caffe.berkeleyvision.org/) With this you will need to upload or otherwise provide a link to an image, and the caffe demo page will predict the class for you. 

This really depends on your application scenario, though. What image size? What are the categories? Do you want it done in real time? Do you want to be able to do anything else on the pi? Depending on these answers, the answer to your original question will probably vary from a "definitely no" to a "probably". 

Have a look at the following page for the installation requirements. 

• It works on Ubuntu which is encouraging but still different from raspbian OS.
• For GPU usage, Caffe uses CUDA and I'm not sure it is ported on the Pi.  In that case, you would be restricted to the CPU which would slow you down considerably.

The odds are not in your favor but there is nothing like trying!  You may need to tweek only a few things to get it to build on raspbian OS and CUDA is optional so there is hope...

http://caffe.berkeleyvision.org/installation.html

Dear Ravimal,

I would suggest you visit the following websites for information on caffe deep learning framework and as well other deep learning resources

1.     http://caffe.berkeleyvision.org/

2.     http://deeplearning.net/

3.     http://www.kdnuggets.com/2015/06/popular-deep-learning-tools.html

4.     https://github.com/ChristosChristofidis/awesome-deep-learning

I hope this information would be help to you.

Good luck.

Thank you very much for the valuable information. @Niall You might be correct since I have already implemented a CNN with only c++ and openCV in a raspberry pi and the speed was extremely low. I was thinking that the algorithmic optimization in caffe would do some better in regards of the classification speed. 

@Bruno then I may try with Ubuntu mate installed in a raspberry pi. Of course that it definitely be restricted to CPU and what I am worrying is whether it can be run in realtime. 

Yass! I actually built a cluster of raspberry Pi's and parallelized MSGD. It is feasible, but probably not worth it.

@William, That is awesome. Did you used the cluster for training as well? Did the cluster able to classify images in realtime? 

Pete Warden wrote matrix multiplication for the Pi's Videocore IV: https://www.raspberrypi.org/blog/more-qpu-magic-from-pete-warden/. But AFAIK no one has implemented Caffe routines specifically on the raspberry pi, which means:

From experience with the Pi, it might be worth investing the time to implement image processing on the Pi if one already has PiCam cameras because retrieving images or video feeds on the Pi is sometimes too slow due to low network speed.

 @Ravimal, any chance you were able to complete this implementation using debian OS on the Pi? 

@Chrispin I tried once (but not the caffe) using raspbian OS and found it takes more than 3 seconds just for a single feed forward calculation per image. So I gave up. I have an idea for implementing just a testing network running on Rpi using a pre-trained model from caffe on a workstation. Do you have any related work to this?

@Ravimal, I'm working on something similar and was thinking of using caffe. I was also thinking of having a pre-trained model run on the Pi but with the concerns raised, I will have to re-evaluate my options. Thanks.