Using transfer learning on Google's pre-trained inference-graphs for fast FRC robot localization.
Tested for tensorflow-gpu==1.14 pip3 install tensorflow-gpu==1.14
Google image download tool: https://github.com/hardikvasa/google-images-download labelImg: https://github.com/tzutalin/labelImg
make sure you have pip3 install tensorflow-gpu==1.14
First clone https://github.com/tensorflow/models.git into a directory (i used /home/alexiev/dev)
Then install the protoc zip file from https://github.com/protocolbuffers/protobuf/releases/tag/v3.10.1 in downloads folder run: sudo unzip -o protoc-3.7.1-linux-x86_64.zip -d /usr/local bin/protoc sudo unzip -o protoc-3.7.1-linux-x86_64.zip -d /usr/local 'include/*'
inside tensorflow/models/research/ run: sudo protoc --python_out=. object_detection/protos/*.proto
Inside the tensorflow/models/research folder run: sudo python3 setup.py install
Then add the following to your PYTHONPATH
sudo gedit ~/.bashrc #open bashrc file then add this new line to the end of your bashrc file export PYTHONPATH=$PYTHONPATH=/home/alexiev/dev/tensorflow/models/research:/home/alexiev/dev/tensorflow/models/research/slim then restart then bashrc to make it work source ~/.bashrc you can change the path to your own!
go inside the util folder in terminal
python3 get_data.py
then use the labelImg tool to draw bounding boxes on the images
python3 process_data.py
python3 train_model.py --logtostderr --model_dir=../training/ --pipeline_config_path=../training/ssd_mobilenet_v1_robot.config
to log data to tensorboard at localhost:6006 run tensorboard --logdir=training
export the inference graph python3 export_inference_graph.py --input_type image_tensor --pipeline_config_path ../training/ssd_mobilenet_v1_robot.config --trained_checkpoint_prefix ../training/model.ckpt-XXXX --output_directory ../inference_graph
then run the detector python3 detector.py