train_cars.yaml

postfix: metric_learning_cars  # it will be used as a postfix for logs folder and as a tag for Neptune logger

seed: 42  # fix random seed
precision: 32  # use 32 for single precision and 16 for half-precision, if your GPU support it
accelerator: gpu  # cpu or gpu
devices: 1  # number of devices to use or list of the exact devices, we will use DDP for gpu accelerator and devices > 1
find_unused_parameters: False  # The parameter for DDPStrategy

dataframe_name: df_with_bboxes.csv  # dataframe filename, see the documentation for the desired format
dataset_root: data/CARS196/  # path to the dataset
logs_root: logs/CARS196/  # logs root folder
logs_folder: ${now:%Y-%m-%d_%H-%M-%S}_${postfix}  # logs folder for the current experiment

num_workers: 20  # number of workers to use in DataLoader
cache_size: 0  # set the number of images you wish to cache in train & val datasets

transforms_train:  # transformations that will be applied in train dataset
  name: augs_albu
  args:
    im_size: 224

transforms_val:  # transformations that will be applied in val dataset
  name: norm_resize_albu
  args:
    im_size: 224

sampler:  # batch sampler, in this case, effective batch size is 20 * 10 = 200
  name: balance
  args:
    n_labels: 20  # there are ~200 classes in CARS dataset, which means 1 epoch includes only 200/20 = 10 batches (so, we see all the labels during the epoch, but not all the samples)
    n_instances: 10

bs_val: 512  # validation batch size, usually it's larger than the training one (which is 200)
max_epochs: 10000  # number of epochs to train
valid_period: 10  # since 1 epoch is really short (only 10 batches, as shown above), we avoid running expensive validation procedure every epoch

log_images: False  # set True if you want to see images where the model performed worse

metric_args:
  metrics_to_exclude_from_visualization: [cmc,]
  cmc_top_k: [1]  # to calculate cmc@1
  map_top_k: [5]  # wo calculate map@5
  fmr_vals: [0.01]  # to calculate fmr@0.01
  pcf_variance: [0.5, 0.9, 0.99]  # to calculate pcf@0.5, pcf@0.9, pcf@0.99
  return_only_overall_category: True  # set False if you want to see metric graphs for all the categories (doesn't matter for CARS, since it contains no categories)
  visualize_only_overall_category: True  # set False to see images where the model performed worse for each separated category (doesn't matter for CARS, since it contains no categories)


# the best checkpoint is determined by the value of metric_for_checkpointing.
# "OVERALL" means we picked the metric calculated over all the categories.
# it can be replaced by the name of the exact category (but it's not the case for CARS dataset containing no categories).
# you can check other metrics in your logger.
metric_for_checkpointing: OVERALL/cmc/1


# extractor to train
extractor:
  name: vit
  args:
    normalise_features: False  # this is the only place in the pipeline where features normalisation can be set
    use_multi_scale: False  # set True to use Test-Time-Augmentations, it usually boosts metrics, but requires several forwards (3 forwards, in this particular case)
    weights: vits16_dino  # <vits16_dino> is a key related to one of the models from Zoo, thus, it will be downloaded automatically. You can also use a path to your local checkpoint
    arch: vits16  # one of vit's architectures


criterion:
  name: triplet_with_miner
  args:
    need_logs: True  # set True to log values like positive distance, negative distance, number of active triplets and so on
    margin: null  # <null> means that we use the soft version of Triplet Loss here. In this case, features normalisation in the model has to be turned off so gradients can flow normally
    reduction: mean
    miner:
      name: all_triplets  # We collect all the possible triplets in the batch. Based on the sampler's parameters it's ~10^3 * 20^2 = 400.000 triplets
      args: {}


optimizer:
  name: adam
  args:
    lr: 1e-5  # if you provide lr scheduler this parameter will be ignored


# we don't use scheduling in this experiment
scheduling: null


# There are several options for logging:
#logger:
#  name: tensorboard
#  args:
#    save_dir: "."

#logger:
#  name: neptune  # requires <NEPTUNE_API_TOKEN> as global env
#  args:
#    project: "oml-team/test"

logger:
  name: wandb  # requires <WANDB_API_KEY> as global env
  args:
    project: "test_project"


# these tags will be applied to a run if Neptune logger was used
# they may be useful to navigate among your runs
tags:
  - ${postfix}
  - cars


# path to Hydra's directory (inside your current log dir) where you can find all the configs
hydra_dir: ${logs_root}/${logs_folder}/

hydra:
  run:
    dir: ${hydra_dir}
  searchpath:
   - pkg://oml.configs  # it allows to use predefined configs of models, optimizers and so on from oml.configs as nested blocks, see Hydra's socs
  job:
    chdir: True