Class-Incremental Sampler

This sampler is a sampler that creates an effective batch. For default setting, the square root of (number of old data/number of new data) is used as the ratio of old data.

API

from otx.algo.samplers.class_incremental_sampler import ClassIncrementalSampler

dataset = OTXDataset(...)
class_incr_sampler = ClassIncrementalSampler(
    dataset=dataset,
    batch_size=32,
    old_classes=["car", "truck"],
    new_classes=["bus"],
)

CLI

(otx) ...$ otx train ... \
                     --data.train_subset.sampler.class_path otx.algo.samplers.class_incremental_sampler.ClassIncrementalSampler \
                     --data.train_subset.sampler.init_args.old_classes '[car,truck]' \
                     --data.train_subset.sampler.init_args.new_classes '[bus]'

Balanced Sampler

This sampler is a sampler that creates an effective batch. It helps ensure balanced sampling by class based on the distribution of class labels during supervised learning.

API

from otx.algo.samplers.balanced_sampler import BalancedSampler

dataset = OTXDataset(...)
class_incr_sampler = BalancedSampler(
    dataset=dataset,
)

CLI

(otx) ...$ otx train ... \
                     --data.train_subset.sampler.class_path otx.algo.samplers.balanced_sampler.BalancedSampler