Mixed precision simple demo

Build docker image

docker build . -t mixed-precision

Run container

docker run -it --gpus=all --rm --ipc=host mixed-precision bash

Run training:

1. Without mixed precision: python main.py --batch-size 256
2. With mixed precision: python main.py --batch-size 256 --use-mixed-precision O1
3. With resnet50, no mixed precision and huge batch we get and CUDA out of memory message: python main.py --arch resnet50 --batch-size 128
4. With resnet50 and mixed preicsion, same huge batch training works: python main.py --arch resnet50 --batch-size 128 --use-mixed-precision O1