This repo contains code for running different convolutional neural networks on a Raspberry Pi cluster. The use case is transfer learning (only classification layers are learned) on a subset of ImageNet.

• Start ray

ray stop && OMP_NUM_THREADS=3 ray start --address='192.168.0.13:6379' --redis-password='5241590000000000'

• Start training

python ray/main.py

• Parameters:

  • image_path: Path to dataset (format: data/train/<class_names>/<image_files> + data/val/<class_names>/<image_files>)
  • model_name: One of ['alexnet', 'resnet', 'mobilenet', 'quantized_mobilenet']
  • num_epochs: Number of epochs
  • batch_size: Batch size
  • num_workers: How many worker nodes to use
  • sync_param_server: Whether to use a synchronous instead of an async parameter server