Train neural network for image classification (CIFAR-10) with horovod

Run Docker

  1. Install Docker
  2. Build image:
docker build -t horovod:latest docker
  1. Run container:
docker run -it horovod:latest

Run training

horovodrun -np 3 python train.py --num_epochs=3 --batch-size=20 --lr=0.001 |& grep -v "Read -1"

Experiments:

Accuracy after 3 epochs: 68%