JeVois Tutorials  1.20
JeVois Smart Embedded Machine Vision Tutorials
Share this page:
Training custom TensorFlow networks for JeVois

This tutorial will show you how to train TensorFlow deep neural networks using your own collection of images and object categories, and how to run the trained network on the processor inside the JeVois smart camera. It closely follows the steps of the tutorial TensorFlow for poets developed by the TensorFlow team, adding a few steps to get the trained network working on JeVois.

Our network trained on 5 different kinds of flowers correctly recognizes roses.


  • Make sure you download JeVois v1.15.0 or later from and flash it to microSD.
  • You need a Linux, Windows or Mac host computer (we used Ubuntu 20.04), onto which you will install and run TensorFlow for network training.
  • You need to be generally familiar software installation procedures, running programs, etc for your host computer.

Plan of attack

  • Like in TensorFlow for poets, we will start with a MobileNet network (a small but high performing image recognition network) pre-trained on ImageNet (a large dataset with 1.2 million images in 1000 different object categories). This will save a lot of time compared to training a blank network from scratch.
  • We will install the latest TensorFlow on the host computer.
  • We will setup a collection of training images in several different categories.
  • We will do transfer learning and fine-tune the pre-trained network on our collection of images, switching from the default 1000 object categories to a different number of categories available in our collection of images. This is performed on the fast host computer.
  • We will optimize it for inference and convert it to TensorFlow Lite (flatbuffer) so that it can be loaded and run on JeVois.
  • We will copy the converted network to JeVois microSD card.
  • And finally we will run it from within the existing TensorFlowEasy JeVois module.
  • If all goes well, you should be able to complete all steps in about 30 minutes.

Detailed steps

Below are details on how to train and deploy your own deep network for JeVois.

The first few steps will follow closely the TensorFlow for poets tutorial. We recommend that you look at it as you also follow the steps outlined here, as it provides additional details not duplicated here.

Install the latest TensorFlow

Follow the instructions at

We are going to use the pip install method for TensorFlow 1.15. On Ubuntu 20.04, we did the following for an install with no GPU support (which is shown here as it is much easier than a full install with GPU support). We used the VirtualEnv installation method and python3 (see for more details):

First, TensorFlow 1.15 requires python 3.7, but Ubuntu 20.04 ships with python 3.8. So we first need to install python 3.7 on our machine (see

sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.7 python3.7-venv

Now we can proceed with the TensorFlow installation, just making sure we invoke python3.7:

virtualenv -p python3.7 --system-site-packages ./venv
source ./venv/bin/activate
python3 --version # should return 3.7.8

Once activated, the shell prompt changes to show you the name of your VirtualEnv. We proceed as follows (from

pip install --upgrade pip
pip install --upgrade numpy
python3 -c "import numpy; print(numpy.__version__)" # we got 1.19.1 when writing this tutorial
pip install --upgrade tensorflow==1.15

Verify the install:

python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
# No error and returns: Tensor("Sum:0", shape=(), dtype=float32)

Get the TensorFlow for poets code

We now follow the steps of the TensorFlow for poets tutorial at

git clone
cd tensorflow-for-poets-2

Download training images (or create your own collection)

We need to have a collection of images that we will use for training. The images should be organized under a number of directories, where each directory is the name of a given object category.

Let us just use the sample images from to set the ideas, but you can of course use your own images instead:

curl | tar xz -C tf_files
ls tf_files/flower_photos

You should see:

daisy dandelion LICENSE.txt roses sunflowers tulips

and under each directory (daisy, dandelion, etc) we have a number of JPEG images which are pictures of that object:

Category Number of images
daisy 633
dandelion 898
roses 641
sunflowers 699
tulips 799

If you want to add more categories, or use different categories, just follow the same organization principle:

  • one directory under tf_files/yourphotos/ for each category
  • in each directory, any number of JPEG pictures of objects in that category

Also see

Configuring MobileNet

We proceed as outlined in and we will use a MobileNet with compression factor 0.5 and input size 128x128, as this should run at about 30 frames/s in the JeVois TensorFlowEasy module:

tensorboard --logdir tf_files/training_summaries &

Then start training:

Be sure to check to get the latest version of the command below, as the exact syntax used sometimes changes with new versions of TensorFlow.
python -m scripts.retrain \
--bottleneck_dir=tf_files/bottlenecks \
--how_many_training_steps=500 \
--model_dir=tf_files/models/ \
--summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}" \
--output_graph=tf_files/retrained_graph.pb \
--output_labels=tf_files/retrained_labels.txt \
--architecture="${ARCHITECTURE}" \

After 500 training steps, which here just took a few minutes, we get an accuracy of 86.5% in our particular run (your results will vary slightly).

Using the trained model with TensorFlow on the host computer

We proceed as outlined in and then

Let's test whether our trained model can now recognize our different types of flowers. We run the trained model on one image from our training set. This is expected to work very well since that image has been used for training:

python -m scripts.label_image --input_width=${IMAGE_SIZE} --input_height=${IMAGE_SIZE} \
--graph=tf_files/retrained_graph.pb \

You should get something like this (actual numbers will vary):

Evaluation time (1-image): 0.176s

daisy (score=0.99013)
dandelion (score=0.00729)
sunflowers (score=0.00258)
roses (score=0.00001)
tulips (score=0.00000)

Which means the network thinks that this image of a daisy indeed is a daisy with 99.01% confidence, or it could be a dandelion with 0.7% confidence, etc

Deploying the trained model to JeVois

We are done with the basic TensorFlow for poets tutorial. Let us now deploy the model to run on the JeVois smart camera. For that, we will convert it to the mobile-optimized TensorFlow Lite format.

We will follow some of the steps of now deprecated TensorFlow for Poets 2: TFLite. But this time we will skip some steps which are irrelevant to JeVois (installing an Android app, etc).

After our retraining above, two files were created:

  • tf_files/retrained_graph.pb is the retrained model
  • tf_files/retrained_labels.txt are the category names used during retraining (note that internally TensorFlow uses a category number, which is then mapped to a name using this file as key. It is hence important that you do not modify that file or change the order of entries in it, etc).

We start by optimizing the model for inference, and then we just convert the model to TensorFlow Lite format (.tflite file). As of August, 2020 (and TensorFlow 1.15), we we should use the tflite_convert utility, as follows:

tflite_convert \
--graph_def_file=tf_files/retrained_graph.pb \
--output_file=tf_files/jevois_model.tflite \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \
--input_array=input \
--output_array=final_result \
--inference_type=FLOAT \

We are now ready to deploy the final model to our JeVois camera. Insert your JeVois microSD into your host computer and check that it is detected. Then copy the model and labels files to it:

# Check that the card was properly detected:
ls /media/${USER}/JEVOIS/share/tensorflow
# You should see a bunch of directories and should get no error, otherwise check the path by which you can access your
# microSD card.
# Create a directory for our new model and copy the model and labels files to it:
mkdir /media/${USER}/JEVOIS/share/tensorflow/flowers
cp tf_files/jevois_model.tflite /media/${USER}/JEVOIS/share/tensorflow/flowers/model.tflite
cp tf_files/retrained_labels.txt /media/${USER}/JEVOIS/share/tensorflow/flowers/labels.txt

Optional: if you want your new model to be loaded by default when TensorFlowEasy is loaded: edit /media/${USER}/JEVOIS/modules/JeVois/TensorFlowEasy/params.cfg and add:

foa=128 128

and comment out any other settings in that file so that the one you just added for flowers is the only uncommented one in the whole file.

Insert the microSD into JeVois and connect it to your host computer, then use JeVois Inventor, or launch TensorFlowEasy by selecting YUYV 320x308 resolution in guvcview or any other video capture software:

guvcview -f YUYV -x 320x308

If you did not edit params.cfg above, you need to manually select your new network by connecting to the JeVois command line interface and issuing:

setpar netdir flowers

Point your JeVois camera to the different kinds of flowers we have trained it for (or to pictures of those from tf_files/flower_photos/) and see how well it can recognize them!

Our network correctly recognizes roses.
Our network correctly recognizes sunflowers.

Note that we did not train a negative or background category in this tutorial. So the model is likely to detect flowers when looking at other things, just because the only things it now knows about the world are those 5 types of flowers we have trained it on.

Our network incorrectly classifies a car (not among its possible output labels) as roses.

Next steps