Training custom TensorFlow networks for JeVois

Question

I was able to retrain the model, but when it comes to the section below on inference toco complains that you can't use the same output format as the input format. If I skip this step and load into the camera I get the catch statement error message on the screen. Is there another way to optimize the model? Thanks.

# Optimize the model for inference:
toco \
  --input_file=tf_files/retrained_graph.pb \
  --output_file=tf_files/optimized_graph.pb \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TENSORFLOW_GRAPHDEF \
  --input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \
  --input_array=input \
  --output_array=final_result

Answer 1

Probably something changed in TensorFlow (that tutorial was written with TensorFlow 1.7, but now we run 1.9 in JeVois).

Can you check out the steps there:

https://codelabs.developers.google.com/codelabs/tensorflow-for-poets-2-tflite/index.html#2

in particular step 3, looks like they combined the two toco commands they originally had into just one now.

We will try it and update our tutorial in the next few days.

Comments

Hi,
I was able to get things to work. I think the key for me was your mentioning the Tensorflow versions. I was using the latest version (1.11?) which had much different arguments for toco function. Plus, there is no real feedback from the script other than "error". Everything else I've been doing is in a windows environment up to the toco which required me to run a Virtualbox with Ubantu. Not much of a Linux person so there was a little bit of a learning curve. Supposedly toco is going to be converted as some point in the future for windows support.  

Thanks. And I hope this helps someone else who might be struggling :)

---

great, thanks as indeed this is very useful to know! Traveling now but doc update will follow shortly.