Skip to content

Latest commit

 

History

History
149 lines (131 loc) · 4.82 KB

GETTING_STARTED.md

File metadata and controls

149 lines (131 loc) · 4.82 KB
Raw

Getting Started

This page provides basic tutorials about the usage of OpenUnReID. The training and testing scripts can be found in OpenUnReID/tools.

We use 4 GPUs for training and testing, which is considered as a default setting in the scripts. You can adjust it (e.g. ${GPUS}, ${GPUS_PER_NODE}) based on your own needs.

Test

Testing commands

  • Distributed testing with multiple GPUs:
bash dist_test.sh ${RESUME} ${CONFIG} [optional arguments]
  • Distributed testing with multiple machines:
bash slurm_test.sh ${PARTITION} ${RESUME} ${CONFIG} [optional arguments]
  • Testing with a single GPU:
GPUS=1 bash dist_test.sh ${RESUME} ${CONFIG} [optional arguments]

Arguments

  • ${RESUME}: model for testing, e.g. ../logs/MMT/market1501/model_best.pth.
  • ${CONFIG}: config file for the model, e.g. MMT/config.yaml. Note the config is required to match the model.
  • [optional arguments]: modify some key values from the loaded config file, e.g. TEST.rerank True. (it's also ok to make the modification directly in the config file)

Configs

  • Test with different datasets, e.g.
TEST:
  datasets: ['dukemtmcreid', 'market1501',] # arrange the names in a list
  • Add re-ranking post-processing, e.g.
TEST:
  rerank: True # default: False
  • Save GPU memory but with a lower speed,
TEST:
  dist_cuda: False # use CPU for computing distances, default: True
  search_type: 3 # use CPU for re-ranking, default: 0 (1/2 is also for GPU)
  • ... (TBD)

Train

Training commands

  • Distributed training with multiple GPUs:
bash dist_train.sh ${METHOD} ${WORK_DIR} [optional arguments]
  • Distributed training with multiple machines:
bash slurm_train.sh ${PARTITION} ${JOB_NAME} ${METHOD} ${WORK_DIR} [optional arguments]
  • Training with a single GPU:

Please add TRAIN.LOADER.samples_per_gpu 64 in [optional arguments].

GPUS=1 bash dist_train.sh ${METHOD} ${WORK_DIR} [optional arguments]

Arguments

  • ${METHOD}: method for training, e.g. source_pretrain, UDA_TP, MMT, SpCL.
  • ${WORK_DIR}: folder for saving logs and checkpoints, e.g. MMT/market1501, the absolute path will be LOGS_ROOT/${WORK_DIR} (LOGS_ROOT is defined in config files).
  • [optional arguments]: modify some key values from the loaded config file, e.g. TRAIN.val_freq 10. (it's also ok to make the modification directly in the config file)

Configs

  • Flexible architectures,
MODEL:
  backbone: 'resnet50' # or 'resnet101', 'resnet50_ibn_a', etc
  pooling: 'gem' # or 'avg', 'max', etc
  dsbn: True # domain-specific BNs, critical for domain adaptation performance
  • Ensure reproducibility (may cause a lower speed),
TRAIN:
  deterministic: True
  • Flexible datasets,

the conventional USL task, e.g. unsupervised market1501

TRAIN:
  # arrange the names in a dict, {DATASET_NAME: DATASET_SPLIT}
  datasets: {'market1501': 'trainval'}
  # define the unsupervised dataset indexes, here index=[0] means market1501 is unlabeled
  unsup_dataset_indexes: [0,]
  # val_set of 'market1501' will be used for validation
  val_dataset: 'market1501'

the USL task with multiple datasets, e.g. unsupervised market1501+dukemtmcreid

TRAIN:
  # you could use multiple datasets for training without the limitation of numbers
  datasets: {'market1501': 'trainval', 'dukemtmcreid': 'trainval'}
  unsup_dataset_indexes: [0,1]
  # you could only choose one dataset for validation (to select the best model)
  # however, you could use multiple datasets for testing (TEST.datasets)
  val_dataset: 'market1501'

the conventional UDA task, e.g. market1501 -> dukemtmcreid

TRAIN:
  datasets: {'market1501': 'trainval', 'dukemtmcreid': 'trainval'}
  unsup_dataset_indexes: [1,]
  val_dataset: 'dukemtmcreid'

the UDA task with multiple source-domain datasets, e.g. market1501+msmt17 -> dukemtmcreid

TRAIN:
  datasets: {'market1501': 'trainval', 'msmt17': 'trainval', 'dukemtmcreid': 'trainval'}
  unsup_dataset_indexes: [2,]
  val_dataset: 'dukemtmcreid'

the UDA task with multiple target-domain datasets, e.g. msmt17 -> market1501+dukemtmcreid

TRAIN:
  datasets: {'market1501': 'trainval', 'msmt17': 'trainval', 'dukemtmcreid': 'trainval'}
  unsup_dataset_indexes: [0,2]
  val_dataset: 'market1501'
  • Train with different losses, e.g.
TRAIN:
  losses: {'cross_entropy': 1., 'softmax_triplet': 1.} # {LOSS_NAME: LOSS_WEIGHT}
  • Save GPU memory but with a lower speed,
TRAIN:
  PSEUDO_LABELS:
    search_type: 3 # only for dbscan, use CPU for searching top-k, default: 0 (1/2 is also for GPU)
    dist_cuda: False # only for kmeans, use CPU for computing distances, default: True
  • Mixed precision training
TRAIN:
  amp: True # mixed precision training for PyTorch>=1.6
  • ... (TBD)