stslidr_classbalance_obow.yaml

_base_config: config/pretrain/nuscenes/slidr/slidr_obow.yaml

model:
  image:
    # return pretrained image features for computing similarity between superpixels
    return_pretrain_feat: True

criteria:
  NCE_temperature: 0.07
  losses:
    enabled: ["loss_superpixel_stslidr_classbalance"]
    params:
      loss_superpixel_stslidr_classbalance:
        loss_balance_noise_threshold: 0.5 # anything lower than 0.5 is assumed to be noise
        min_quantile: 0.01 # percentage of negatives to drop