Productize lab deep clean neural network

This project uses the Inception V3 neural network for TensorFlow machine learning framework to build a graph model of clean, dirty and busy rooms.

It is based on the Use Artificial Intelligence to Detect Messy/Clean Rooms! guide by Matt Farley, which in term are based on the TensorFlow image retraining and image recognition examples.

Usage

Set-up

Train the neural network, resulting in the graph model and found labels (best to do this on a fast computer):

pip install -r requirements-label.txt
python retrain.py \
    --image_dir $TRAINING_IMAGES \
    --output_graph=deep-clean-graph.pb \
    --output_labels=deep-clean-labels.txt \
    --tfhub_module https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1

Testing

Label a test image. Best practice is to use an image that was not used during training.

python label_image.py \
    --graph=deep-clean-graph.pb \
    --labels=deep-clean-labels.txt \
    --input_layer=Placeholder \
    --output_layer=final_result \
    --image=$IMAGE

Deploying

Install dependencies on a Raspberry Pi:

sudo apt install libatlas-base-dev
pip3 install -r requirements-server.txt

Run the server:

python3 deep-clean-server.py

TODO: create a service