ResNet50 example

This example demonstrates how to train a Neural Module which implements ResNet50 network on ImageNet data using multi-GPU (but single node) training.

1. Step 1: Get ImageNet data in Image Folder format.
2. Step 2: Run training. (This will also start evaluation in parallel)

python -m torch.distributed.launch --nproc_per_node=2 resnet50.py --data_root=/mnt/D1/Data/ImageNet/ImageFolder/ --num_gpus=2

note that nproc_per_node should be equal to num_gpus. If you run out of GPU memory, reduce batch_size parameter. This parameters is per GPU. 3) Step 3: Monitor training with TensorBoard

tensorboard --logdir=resnet50