build details

Show: section status errors & todos local changes recent changes last change in-page changes feedback controls

How To Use Neural Compute Stick

Modified 2019-09-22 by Andrea Censi

Workflow

Modified 2019-09-22 by Andrea Censi

create and train model in tensorflow or caffe (brief note on configuration )

save tensorflow model as a .meta (or caffe model in .prototxt)

saver = tf.train.Saver()
...
saver.save(sess, 'model')

compile the model into NC format (documentation here)

mvNCCompile model.meta -o model.graph

move model onto duckiebot

scp model.meta user@robot name:~/path_to_networks/

run the compiled model

with open(path_to_networks + model.meta, mode='rb') as f:
    graphfile = f.read()
graph = device.AllocateGraph(graphfile)
graph.LoadTensor(input_image.astype(numpy.float16), 'user object')
output, userobj = graph.GetResult()

Benchmarking

Modified 2019-09-22 by Andrea Censi

get benchmarking (frames per second) from their app zoo

git clone https://github.com/movidius/ncappzoo
cd ncappzoo/apps/benchmarkncs
./mobilenets_benchmark.sh | grep FPSk