fasterrcnn.yaml

name: mmdet_fasterrcnn
data:
  file_client_args:
      backend: gcs
      bucket_name: determined-ai-mmdet-data
      ##### Other backends #####
      #backend: s3
      #bucket_name: determined-ai-coco-dataset
      #backend: disk # assumes data available at /run/determined/workdir/data in the container
      #backend: fake
##### You can enable profiling with below #####
#profiling:
#  enabled: true
#  begin_on_batch: 200
#  end_after_batch: 300
hyperparameters:
  global_batch_size: 16
  config_file: /mmdetection/configs/faster_rcnn/faster_rcnn_r50_caffe_fpn_1x_coco.py
  merge_config: null # You can specify a config you want to merge into the config_file above.
  use_pretrained: false # Whether to load pretrained weights for config if available.
  override_mmdet_config:
    ##### Learn more about mmdet configs: https://mmdetection.readthedocs.io/en/v2.27.0/tutorials/config.html #####
    ##### You can specify gradient clipping with below #####
    optimizer_config._delete_: true
    optimizer_config.grad_clip.max_norm: 100
    optimizer_config.grad_clip.norm_type: 2
    ##### You can specify mixed precision with below #####
    #fp16.loss_scale: 512. # can be float or dict of named args for torch.cuda.amp.GradScaler

checkpoint_storage:
  save_trial_latest: 5
min_validation_period:
  batches: 7330
searcher:
  name: single
  metric: bbox_mAP
  max_length:
    batches: 87960
  smaller_is_better: false
max_restarts: 5
environment:
  image:
    gpu: determinedai/model-hub-mmdetection:0.26.2-dev0
  environment_variables:
    - OMP_NUM_THREADS=1 # Following pytorch dtrain, this environment variable is set to 1 to avoid overloading the system.

resources:
  slots_per_trial: 8 # max number of GPUs a trial is allowed to individually use
  shm_size: 200000000000
entrypoint: python3 -m determined.launch.torch_distributed --trial model_hub.mmdetection:MMDetTrial