retrain.py

[idata]

I was retraining with inception v3 using the standard python 3.5 tensorflow\examples\image_retraining\retrain.py, but I am uncertain how to take the output_graph.pb, output labels, bottleneck files etc and port them over to the RPI3/movidius environment. - is it a straight copy or do I have to modify the retrain.py script so the results will work with the movidius stick?

[idata]

Just got the movidius stick so I am just learning how it works.

[idata]

The graph which your retrained should be transformed by tensorflow tool "graph_transforms" to remove the unsupport or necessory operation.

[idata]

is there a document or a reference page that explains how to transform the tensorflow output_graph.pb file into the proper format? I have been doing lots of reading, but I can't seem to find any docs that indicate what exactly to change (using the graph_transform) to get the *.pb file into a format that the movidius needs.

[idata]

@Alphadev We don't have a document that shows how to convert your network from a pb format to a NCS-friendly format at the moment. We do have a guide that shows how to convert your network to an NCS-friendly format after saving your session @ https://movidius.github.io/ncsdk/TensorFlow.html

[idata]

hmm… I was just retraining using the last step in inception. eg:

git\tensorflow\examples\image_retraining>python ./retrain.py --image_dir ~\flower_photos

I am not a CNN developer (just starting to learn about creating new networks), and I did not want to build a whole new CNN when I just wanted to retrain so I can leverage the capabilities of image recognition for my low level electronics development (recognize image & act upon the result type thing).

How do I retrain an existing network with different images/classifications that will run on the NCS? - I feel like I am missing something really basic here.

[idata]

@Alphadev Have you tried using mvNCCompile with your model files and see if you can produce a graph file?