Getting started

Our experiments heavily rely on Docker and Kubernetes. For the detailed experimental environment setup, please refer to dockerfile under the environments folder.

Use case of distributed training (centralized/decentralized)

Some simple explanation of the arguments used in the code.

• Arguments related to distributed training:
  • The n_mpi_process and n_sub_process indicates the number of nodes and the number of GPUs for each node. The data-parallel wrapper is adapted and applied locally for each node.
    • Note that the exact mini-batch size for each MPI process is specified by batch_size, while the mini-batch size used for each GPU is batch_size/n_sub_process.
  • The world describes the GPU topology of the distributed training, in terms of all GPUs used for the distributed training.
  • The hostfile from mpi specifies the physical location of the MPI processes.
  • We provide two use cases here:
    • n_mpi_process=2, n_sub_process=1 and world=0,0 indicates that two MPI processes are running on 2 GPUs with the same GPU id. It could be either 1 GPU at the same node or two GPUs at different nodes, where the exact configuration is determined by hostfile.
    • n_mpi_process=2, n_sub_process=2 and world=0,1,0,1 indicates that two MPI processes are running on 4 GPUs and each MPI process uses GPU id 0 and id 1 (on 2 nodes).
• Arguments related to communication compression:
  • The graph_topology
  • The optimizer will decide the type of distributed training, e.g., centralized SGD, decentralized SGD
  • The comm_op specifies the communication compressor we can use, e.g., sign+norm, random-k, top-k.
  • The consensus_stepsize determines the consensus stepsize for different decentralized algorithms (e.g. parallel_choco, deep_squeeze).
• Arguments related to learning:
  • The lr_scaleup, lr_warmup and lr_warmup_epochs will decide if we want to scale up the learning rate, or warm up the learning rate. For more details, please check pcode/create_scheduler.py.

Examples

The script below trains ResNet-20 with CIFAR-10, as an example of centralized training algorithm CHOCO. More examples can be found in exps.

OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 $HOME/conda/envs/pytorch-py3.6/bin/python run.py \
    --arch resnet20 --optimizer parallel_choco \
    --avg_model True --experiment demo --manual_seed 6 \
    --data cifar10 --pin_memory True \
    --batch_size 128 --base_batch_size 64 --num_workers 2 \
    --num_epochs 300 --partition_data random --reshuffle_per_epoch True --stop_criteria epoch \
    --n_mpi_process 16 --n_sub_process 1 --world 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 \
    --on_cuda True --use_ipc False \
    --lr 0.1 --lr_scaleup True --lr_warmup True --lr_warmup_epochs 5 \
    --lr_scheduler MultiStepLR --lr_decay 0.1 --lr_milestones 150,225 \
    --comm_op sign --consensus_stepsize 0.5 --compress_ratio 0.9 --quantize_level 16 --is_biased True \
    --weight_decay 1e-4 --use_nesterov True --momentum_factor 0.9 \
    --hostfile hostfile --graph_topology ring --track_time True --display_tracked_time True \
    --python_path $HOME/conda/envs/pytorch-py3.6/bin/python --mpi_path $HOME/.openmpi/