Give proof of concept (POC) on the application of federated learning on building object detection model during internship at Audi China R&D. In this projecct, we adopt Faster R-CNN algorithm in building an object detection model under Tensorflow Federated (TFF) framework. Choose VGG16 as base network.
- Augment VOC2007 and VOC2012 with duplication and random rotation
- Random assign all data to artificial clients (N= 50, 100, 500)
- Perform FedAvg: update on single client -> update on global server
Functions of important scripts in codes:
simple_parser.py: Convert train data formataugment.py: Augment train data with horizontal, vertical flip and random rotationconfig.py: Pass train settings
Learn how TFF works at simple tasks:
- Federated learning for image classification using MNIST dataset: https://www.tensorflow.org/federated/tutorials/federated_learning_for_image_classification
- Federated learning for text generation using Shakespeare dataset: https://www.tensorflow.org/federated/tutorials/federated_learning_for_text_generation
h5py
tensorflow
tensorflow-gpu==1.14.0
Keras==2.0.3
numpy
opencv-python
sklearn
See important compatibility at https://github.com/tensorflow/federated#compatibility