deploying-custom-caffe-models on raspberry pi3 + movidius

idata · ‎09-27-2018

Hi

I have successfuly executed several of the examples on an NCSDK v2 installation on a raspberrypi3 + movidius. I now wish

to understand and use the movidius tools to design and develop for edge devices.

Question is 1) Can I use the pi installation to execute all the dogs v cats program given in:

https://movidius.github.io/blog/deploying-custom-caffe-models/

( I understand there may be an execution time penalty on the Pi+movidius configuration).

Question 2)Though I have loaded the train and test zips I do not understand where the programs are located that

are intended to perform the training, presumably they are to be contained in the dogsvcats directory with data as a sub directory.

Would appreciate clarification on these questions.

Thank you

jsl

idata · ‎09-29-2018

Hello,

I too have the same problem. Did you manage to get it up and running ?

idata · ‎09-30-2018

Hi there

Nope, not yet, I was hoping for advice. I will certainly post any progress - Interesting that you are seeing the same problem.

idata · ‎10-01-2018

@jslawton, @abypaulvarghese,

The dogsvscats project has a makefile that does everything from data prep, to training, to model conversion, to deploying the model on NCS for inference. Below is an excerpt from the Let's Build! section of the blog:

1. Train - Neural network selection, dataset preparation, and training
2. Profile - Analyze the neural network for bandwidth, complexity, and execution time
3. Fine tune - Modify the neural network topology to gain better execution time
4. Deploy - Deploy the customized neural network on an edge device powered by NCS

If you are interested in running this project on a Raspberry Pi, you should run train, profile and tune on your training hardware and just run the deploy step on your Raspberry Pi. I have made a couple changes to the project since I wrote the blog, so try to cherrypick https://github.com/movidius/ncappzoo/pull/155/ and run the following:

On your training hardware: These steps would give you a Movidius graph file

make train

make compile

On your Raspberry Pi

make run

Thanks

Ashwin

idata · ‎10-02-2018

Thank you very much.

1) Can the training Hardware be the raspberrypi?

2) I have returned to understanding/using movidius after taking considerable time out to study Chris Bishop's work and now on returning have found your webinar tutorials. I think the webinar code is designed to run under V1 NCSDK, correct?