[minigo] Prepare Dataset

[jay-huh]

refer README.md

• [x] Install gsutil for accessing to GCP Storage
• [x] Download and Exteract checkpoint.tar.gz
• [x] Download target*
• [x] Build Docker

• [x] Pre-process dataset in docker

  # run docker
  nvidia-docker run -v <path/to/store/checkpoint>:/data --rm -it mlperf-nvidia:minigo
  cd minigo
  
  # comment out L331 in dual_net.py before running freeze_graph.
  # L331 is: optimizer = hvd.DistributedOptimizer(optimizer)
  # Horovod is initialized via train_loop.py and isn't needed for this step.
  
  CUDA_VISIBLE_DEVICES=0 python3 freeze_graph.py --flagfile=ml_perf/flags/19/architecture.flags  --model_path=ml_perf/target/target
  mv ml_perf/target/target.minigo ml_perf/target/target.minigo.tf
  
  # uncomment L331 in dual_net.py.
  # copy dataset to /data that is mapped to <path/to/store/checkpoint> outside of docker.
  # Needed because run_and_time.sh uses the following paths to load checkpoint
  # CHECKPOINT_DIR="/data/mlperf07"
  # TARGET_PATH="/data/target/target.minigo.tf"
  
  cp -a ml_perf/target /data/
  cp -a ml_perf/checkpoints/mlperf07 /data/

[jay-huh]

Location of pre-processed dataset

/opt/data/Dataset/training/minigo/ml_perf