NVIDIA MLPerf v0.5.0 - minor tweaks for reproducibility

[cliffwoolley]

README and launch script fixes to ensure reproducibility of the submitted results. These do not impact performance or correctness of the results.

• Improved documentation of dataset preprocessing steps

  • image_classification/mxnet/README.md
  • recommendation/pytorch/README.md
  • translation/pytorch/README.md

• Removed unnecessary debugging code that was added to the RN50 launcher late in the submission process

  • image_classification/mxnet/run.sub

• Fixed typos

  • image_classification/mxnet/run_and_time.sh
  • recommendation/pytorch/convert.py
  • object_detection/pytorch/Dockerfile
  • translation/pytorch/Dockerfile

• Revert accidental overwrite of multi-node config into single-node DGX2 config so as to properly reflect submitted logs

  • single_stage_detector/pytorch/config_DGX2.sh
  • single_stage_detector/pytorch/config_DGX2_multi.sh

Signed-off-by: Cliff Woolley jwoolley@nvidia.com

[christ1ne]

It's good to document these changes

[nvpaulius]

Merging given the approvals above.