Serving a Tensorflow CNN model for Hand Detection using Tensorflow Serving

DataHack Summit 2019 Slides

I conducted a hack session at DataHack Summit 2019 on Model Deployment using Tensorflow Serving. Here are the slides - Talk slides

Model Training

I referred to the following work for training the hand detector model -

Victor Dibia, HandTrack: A Library For Prototyping Real-time Hand TrackingInterfaces using Convolutional Neural Networks, https://github.com/victordibia/handtracking

File and folder Descriptions

• ball_with_hand.py - Driver code to run the hand detection model on video frames
• hand_detection_client.py - Classes that enable interation with the model using gRPC or REST
• serving_config.py - Port and host configuration to interact with the served model
• inference_graph - Saved model for Hand Detection

Dependencies

• python dependencies
• Install Tensorflow Serving using Docker

Example

• Download the sample video

• Serve the hand detector

  docker run -p 8501:8500 -p 9000:8501  \
      --mount type=bind,source=/home/tata/Projects/cnn-hand-detector/inference_graph/,target=/models/inference_graph  \ 
      --mount type=bind,source=/home/tata/Projects/cnn-hand-detector/inference_graph/model_config.config,target=/models/model_config.config \  
      -t -i tensorflow/serving --model_config_file=/models/model_config.config --model_config_file_poll_wait_seconds=10 \  --enable_model_warmup=true

  Demo of running the command (using Asciinema)

  

  The server can now accept gRPC requests through port 8501 and REST requests through port 9000.

• Run ball_with_hand.py using either gRPC or REST

  python3 ball_with_hand.py --method=grpc --video_path=/home/tata/hand_video.mp4

  Demo Video -

  

More details will be added soon to this README.