genml-kit 0.2.0__tar.gz → 0.2.1__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {genml_kit-0.2.0 → genml_kit-0.2.1}/PKG-INFO +16 -13
- {genml_kit-0.2.0 → genml_kit-0.2.1}/README.md +12 -12
- genml_kit-0.2.1/genml_kit/augmentations/multicrop.py +113 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/rollout_buffer.py +40 -52
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/vo_pairs.py +29 -6
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/io/checkpointing.py +24 -1
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/registry.py +2 -8
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/rl.py +17 -16
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/vo_pair.py +33 -19
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/registry.py +3 -10
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/vo/__init__.py +2 -0
- genml_kit-0.2.1/genml_kit/models/vo/vo_similar.py +236 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/rl.py +3 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/vo_pair.py +8 -1
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/classifiers/__init__.py +9 -9
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/optim_factory.py +3 -3
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/rl_trainer.py +2 -1
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/train.py +19 -13
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/train_reporting.py +98 -28
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/trainer.py +21 -2
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/vo/__init__.py +2 -0
- genml_kit-0.2.1/genml_kit/training/vo/eval_vo.py +128 -0
- genml_kit-0.2.1/genml_kit/training/vo/train_vo.py +231 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/script.py +10 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/PKG-INFO +16 -13
- {genml_kit-0.2.0 → genml_kit-0.2.1}/pyproject.toml +6 -1
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_lora.py +18 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_pipeline.py +26 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_losses_phase2.py +19 -23
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_model_selector.py +0 -1
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_normalize.py +2 -2
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_pipeline.py +27 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_trainer.py +10 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_train_reporting.py +171 -11
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_trainer.py +42 -0
- genml_kit-0.2.1/tests/test_vo_e2e.py +163 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_vo_pairs.py +1 -1
- genml_kit-0.2.1/tests/test_vo_similar.py +222 -0
- genml_kit-0.2.1/tests/test_vo_training.py +247 -0
- genml_kit-0.2.0/genml_kit/augmentations/multicrop.py +0 -123
- genml_kit-0.2.0/genml_kit/models/vo/vo_similar.py +0 -165
- genml_kit-0.2.0/genml_kit/training/vo/eval_vo.py +0 -108
- genml_kit-0.2.0/genml_kit/training/vo/train_vo.py +0 -156
- genml_kit-0.2.0/tests/test_vo_e2e.py +0 -73
- genml_kit-0.2.0/tests/test_vo_similar.py +0 -124
- genml_kit-0.2.0/tests/test_vo_training.py +0 -113
- {genml_kit-0.2.0 → genml_kit-0.2.1}/LICENSE +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/augmentations/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/augmentations/dual_view.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/balanced_sampler.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/ensemble.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/factory.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/field_dataset.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/hf_proxy.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/image_folder.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/replay_buffer.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/retry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/transforms.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/datasets/weighted_sampler.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/geometry/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/geometry/similarity.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/io/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/io/storage_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/builtin.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/byol.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/contrastive.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/dino.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/losses/focal.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/base.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/byol.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/classification.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/dino.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/ijepa.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/rl_dqn.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/rl_ppo.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/rl_sac.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/rl_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/simmim.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/methods/supcon.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/attention_pooling.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/byol.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/cls_model_wrapper/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/cls_model_wrapper/loader.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/cls_model_wrapper/model.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/contrastive.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/convvit/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/convvit/loader.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/convvit/masked_encoder.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/convvit/model.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/convvit/processor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/dino.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/encoder_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/init.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/masked_encoder.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/processors/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/processors/base.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/actor_critic.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/qnetwork.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/sac_critic.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/sac_model.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/rl/spaces.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/simmim.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/timm/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/timm/loader.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/timm/model.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/timm/processor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/transformer_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/uvito/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/uvito/loader.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/uvito/model.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/uvito/processor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/vo/loader.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/models/vo/processor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/base.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/contracts.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/images.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/pipelines/rl_normalize.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/classifiers/base.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/classifiers/cls_attention.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/classifiers/mlp.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/eval.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/grad_monitor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/infer.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/labels.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/metrics.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/model_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/param_align.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/train_compat.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/tta.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/video_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/xgb_pipeline.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/training/xgb_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/__init__.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/args.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/attr.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/cli.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/env.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/gpu.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/image_dump.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/label.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/logging.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/seed.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/signal.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/table.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit/utils/transformer.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/SOURCES.txt +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/dependency_links.txt +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/entry_points.txt +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/requires.txt +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/genml_kit.egg-info/top_level.txt +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/setup.cfg +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_actor_critic.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_args_frozen.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_attention_pooling.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_attr_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_balanced_sampler.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_byol.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_checkpointing.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_classifiers.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_cli_help.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_cli_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_contrastive_models.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_custom_models.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_datasets_ensemble.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_dino.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_encoder_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_ensemble_collate.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_ensemble_ensure_label_space.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_ensemble_labels.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_ensemble_mixed_labels.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_env.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_evaluation_metrics.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_extract_features.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_geometry_similarity.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_grad_monitor.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_headless_models.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_image_dump.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_infer_topk.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_init_orthogonal.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_log_format.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_logging_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_loss_and_multipliers.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_loss_registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_method_lifecycle.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_metrics.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_model_script_loading.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_optim_factory.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_param_align.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_pipeline_method_registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_pretrain_methods.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_pretrain_smoke.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_pretrain_vis.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_replay_buffer.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_retry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_cli.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_learning.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_losses.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_method.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_models.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_ppo.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_sac.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_spaces.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rl_video.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_rollout_buffer.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_s3r2util.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_script_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_seed_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_set_train_mode.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_signal_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_simmim.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_storage_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_supcon_loss.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_supcon_method.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_table_utils.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_timm_models.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_train_smoke.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_transformer_encoder_init.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_tta.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_utils_registry.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_uvito.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_weighted_sampler.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_xgb_pipeline.py +0 -0
- {genml_kit-0.2.0 → genml_kit-0.2.1}/tests/test_xgb_utils.py +0 -0
|
@@ -1,8 +1,11 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: genml_kit
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.1
|
|
4
4
|
Summary: General-purpose ML toolkit: image-classification, visual-odometry and self-supervised pre-training through one unified harness.
|
|
5
5
|
License: Apache-2.0
|
|
6
|
+
Project-URL: Homepage, https://github.com/davidel/genml_kit
|
|
7
|
+
Project-URL: Repository, https://github.com/davidel/genml_kit
|
|
8
|
+
Project-URL: Issues, https://github.com/davidel/genml_kit/issues
|
|
6
9
|
Requires-Python: >=3.9
|
|
7
10
|
Description-Content-Type: text/markdown
|
|
8
11
|
License-File: LICENSE
|
|
@@ -344,7 +347,7 @@ an exponential moving average:
|
|
|
344
347
|
|
|
345
348
|
$$
|
|
346
349
|
\large
|
|
347
|
-
\xi \leftarrow m
|
|
350
|
+
\xi \leftarrow m \\, \xi + (1 - m) \\, \theta
|
|
348
351
|
$$
|
|
349
352
|
|
|
350
353
|
where $`m`$ is the momentum (`--teacher_momentum`) and the arrow denotes
|
|
@@ -406,7 +409,7 @@ Fix one example $`i`$, called the *anchor*. Its *positive set* is
|
|
|
406
409
|
|
|
407
410
|
$$
|
|
408
411
|
\large
|
|
409
|
-
\mathcal{P}(i) = \{
|
|
412
|
+
\mathcal{P}(i) = \{\\, j : j \neq i \text{ and } y_j = y_i \\,\}
|
|
410
413
|
$$
|
|
411
414
|
|
|
412
415
|
i.e. all other examples in the batch that share the anchor's class label
|
|
@@ -463,7 +466,7 @@ they are meant to be confident) and a larger student temperature (softer
|
|
|
463
466
|
[Emerging Properties in Self-Supervised Vision Transformers](https://arxiv.org/abs/2104.14294), ICCV 2021.
|
|
464
467
|
One backbone is instantiated twice. The **student** $`f_\theta`$ is trained by
|
|
465
468
|
gradient descent; the **teacher** $`f_\xi`$ is the EMA copy of the student
|
|
466
|
-
($`\xi \leftarrow m
|
|
469
|
+
($`\xi \leftarrow m\\,\xi + (1-m)\\,\theta`$, as in I-JEPA above, ramped from
|
|
467
470
|
`--dino_momentum` 0.996 to `--dino_final_momentum` 1.0). Each image is cut
|
|
468
471
|
into 2 **global** crops plus `--dino_local_num` (8) small **local** crops:
|
|
469
472
|
the teacher sees only global views, the student sees all of them, so the
|
|
@@ -476,7 +479,7 @@ $`c \in \mathbb{R}^{D}`$ updated once per step,
|
|
|
476
479
|
|
|
477
480
|
$$
|
|
478
481
|
\large
|
|
479
|
-
c \leftarrow m_c
|
|
482
|
+
c \leftarrow m_c \\, c + (1 - m_c) \cdot \frac{1}{B}\sum_{i=1}^{B} f_\xi(x_i)
|
|
480
483
|
$$
|
|
481
484
|
|
|
482
485
|
(`--dino_center_momentum` 0.9), and the loss is the cross-entropy between the
|
|
@@ -485,7 +488,7 @@ averaged over all (teacher global, student crop) pairings:
|
|
|
485
488
|
|
|
486
489
|
$$
|
|
487
490
|
\large
|
|
488
|
-
\mathcal{L} = - \sum_{j} p_{\xi,j}
|
|
491
|
+
\mathcal{L} = - \sum_{j} p_{\xi,j} \\, \log p_{\theta,j}
|
|
489
492
|
$$
|
|
490
493
|
|
|
491
494
|
Centering and sharpening are the two anti-collapse mechanisms: centering
|
|
@@ -754,7 +757,7 @@ true class survives and the loss reduces to the familiar $`-\log p(y \mid x)`$.
|
|
|
754
757
|
genml_kit generalizes this in three directions via `CombinedFocalLoss`:
|
|
755
758
|
|
|
756
759
|
- **Soft targets.** With Mixup (below) $`q`$ is the blend
|
|
757
|
-
$`\lambda
|
|
760
|
+
$`\lambda\\, y_i + (1-\lambda)\\, y_j`$, and the sum keeps both classes'
|
|
758
761
|
terms weighted by $`\lambda`$ and $`1-\lambda`$.
|
|
759
762
|
- **Label smoothing** (via `--loss_args '{"label_smoothing": ...}'`):
|
|
760
763
|
the one-hot target is replaced by
|
|
@@ -843,7 +846,7 @@ genml-kit-train --model google/vit-base-patch16-224 \
|
|
|
843
846
|
--epochs 5 \
|
|
844
847
|
--batch_size 32 \
|
|
845
848
|
--lr 3e-5 \
|
|
846
|
-
--image_size
|
|
849
|
+
--image_size 224
|
|
847
850
|
```
|
|
848
851
|
|
|
849
852
|
### With A Custom Classifier Head
|
|
@@ -901,7 +904,7 @@ schedule is (introduced by Howard & Ruder, [Universal Language Model Fine-tuning
|
|
|
901
904
|
|
|
902
905
|
$$
|
|
903
906
|
\large
|
|
904
|
-
\text{lr}(\text{layer}) = \text{lr}_{\text{base}} \cdot d^{
|
|
907
|
+
\text{lr}(\text{layer}) = \text{lr}_{\text{base}} \cdot d^{\\,L - \text{layer}}
|
|
905
908
|
$$
|
|
906
909
|
|
|
907
910
|
where:
|
|
@@ -989,8 +992,8 @@ The loss for one example is
|
|
|
989
992
|
|
|
990
993
|
$$
|
|
991
994
|
\large
|
|
992
|
-
\mathcal{L} = \left(1 - p_t\right)^{\gamma} \sum_{c=1}^{C} w_c
|
|
993
|
-
\left(-\log p_c\right), \qquad p_t = \sum_{c=1}^{C} q_c
|
|
995
|
+
\mathcal{L} = \left(1 - p_t\right)^{\gamma} \sum_{c=1}^{C} w_c\\, q_c
|
|
996
|
+
\left(-\log p_c\right), \qquad p_t = \sum_{c=1}^{C} q_c\\, p_c
|
|
994
997
|
$$
|
|
995
998
|
|
|
996
999
|
Two details matter for the reader combining these flags:
|
|
@@ -1650,8 +1653,8 @@ The head is selected through `--model_arg`, not a dedicated flag:
|
|
|
1650
1653
|
```
|
|
1651
1654
|
|
|
1652
1655
|
The `classifier` value is an inline spec: the head name (a registered
|
|
1653
|
-
built-in
|
|
1654
|
-
`
|
|
1656
|
+
built-in \u2014 `mlp`, `cls_attention` \u2014 or a path to a `.py` file defining a
|
|
1657
|
+
`create_classifier(num_labels, hidden_size, **kwargs)` function), optionally followed by `:key=value,...` kwargs. Values
|
|
1655
1658
|
are parsed as int/float/bool, and bracketed lists such as
|
|
1656
1659
|
`cls_slice=(0, 1)` work. A head is required: `cls_model_wrapper` fails fast
|
|
1657
1660
|
with a clear message when `classifier` is missing. Models that embed their
|
|
@@ -305,7 +305,7 @@ an exponential moving average:
|
|
|
305
305
|
|
|
306
306
|
$$
|
|
307
307
|
\large
|
|
308
|
-
\xi \leftarrow m
|
|
308
|
+
\xi \leftarrow m \\, \xi + (1 - m) \\, \theta
|
|
309
309
|
$$
|
|
310
310
|
|
|
311
311
|
where $`m`$ is the momentum (`--teacher_momentum`) and the arrow denotes
|
|
@@ -367,7 +367,7 @@ Fix one example $`i`$, called the *anchor*. Its *positive set* is
|
|
|
367
367
|
|
|
368
368
|
$$
|
|
369
369
|
\large
|
|
370
|
-
\mathcal{P}(i) = \{
|
|
370
|
+
\mathcal{P}(i) = \{\\, j : j \neq i \text{ and } y_j = y_i \\,\}
|
|
371
371
|
$$
|
|
372
372
|
|
|
373
373
|
i.e. all other examples in the batch that share the anchor's class label
|
|
@@ -424,7 +424,7 @@ they are meant to be confident) and a larger student temperature (softer
|
|
|
424
424
|
[Emerging Properties in Self-Supervised Vision Transformers](https://arxiv.org/abs/2104.14294), ICCV 2021.
|
|
425
425
|
One backbone is instantiated twice. The **student** $`f_\theta`$ is trained by
|
|
426
426
|
gradient descent; the **teacher** $`f_\xi`$ is the EMA copy of the student
|
|
427
|
-
($`\xi \leftarrow m
|
|
427
|
+
($`\xi \leftarrow m\\,\xi + (1-m)\\,\theta`$, as in I-JEPA above, ramped from
|
|
428
428
|
`--dino_momentum` 0.996 to `--dino_final_momentum` 1.0). Each image is cut
|
|
429
429
|
into 2 **global** crops plus `--dino_local_num` (8) small **local** crops:
|
|
430
430
|
the teacher sees only global views, the student sees all of them, so the
|
|
@@ -437,7 +437,7 @@ $`c \in \mathbb{R}^{D}`$ updated once per step,
|
|
|
437
437
|
|
|
438
438
|
$$
|
|
439
439
|
\large
|
|
440
|
-
c \leftarrow m_c
|
|
440
|
+
c \leftarrow m_c \\, c + (1 - m_c) \cdot \frac{1}{B}\sum_{i=1}^{B} f_\xi(x_i)
|
|
441
441
|
$$
|
|
442
442
|
|
|
443
443
|
(`--dino_center_momentum` 0.9), and the loss is the cross-entropy between the
|
|
@@ -446,7 +446,7 @@ averaged over all (teacher global, student crop) pairings:
|
|
|
446
446
|
|
|
447
447
|
$$
|
|
448
448
|
\large
|
|
449
|
-
\mathcal{L} = - \sum_{j} p_{\xi,j}
|
|
449
|
+
\mathcal{L} = - \sum_{j} p_{\xi,j} \\, \log p_{\theta,j}
|
|
450
450
|
$$
|
|
451
451
|
|
|
452
452
|
Centering and sharpening are the two anti-collapse mechanisms: centering
|
|
@@ -715,7 +715,7 @@ true class survives and the loss reduces to the familiar $`-\log p(y \mid x)`$.
|
|
|
715
715
|
genml_kit generalizes this in three directions via `CombinedFocalLoss`:
|
|
716
716
|
|
|
717
717
|
- **Soft targets.** With Mixup (below) $`q`$ is the blend
|
|
718
|
-
$`\lambda
|
|
718
|
+
$`\lambda\\, y_i + (1-\lambda)\\, y_j`$, and the sum keeps both classes'
|
|
719
719
|
terms weighted by $`\lambda`$ and $`1-\lambda`$.
|
|
720
720
|
- **Label smoothing** (via `--loss_args '{"label_smoothing": ...}'`):
|
|
721
721
|
the one-hot target is replaced by
|
|
@@ -804,7 +804,7 @@ genml-kit-train --model google/vit-base-patch16-224 \
|
|
|
804
804
|
--epochs 5 \
|
|
805
805
|
--batch_size 32 \
|
|
806
806
|
--lr 3e-5 \
|
|
807
|
-
--image_size
|
|
807
|
+
--image_size 224
|
|
808
808
|
```
|
|
809
809
|
|
|
810
810
|
### With A Custom Classifier Head
|
|
@@ -862,7 +862,7 @@ schedule is (introduced by Howard & Ruder, [Universal Language Model Fine-tuning
|
|
|
862
862
|
|
|
863
863
|
$$
|
|
864
864
|
\large
|
|
865
|
-
\text{lr}(\text{layer}) = \text{lr}_{\text{base}} \cdot d^{
|
|
865
|
+
\text{lr}(\text{layer}) = \text{lr}_{\text{base}} \cdot d^{\\,L - \text{layer}}
|
|
866
866
|
$$
|
|
867
867
|
|
|
868
868
|
where:
|
|
@@ -950,8 +950,8 @@ The loss for one example is
|
|
|
950
950
|
|
|
951
951
|
$$
|
|
952
952
|
\large
|
|
953
|
-
\mathcal{L} = \left(1 - p_t\right)^{\gamma} \sum_{c=1}^{C} w_c
|
|
954
|
-
\left(-\log p_c\right), \qquad p_t = \sum_{c=1}^{C} q_c
|
|
953
|
+
\mathcal{L} = \left(1 - p_t\right)^{\gamma} \sum_{c=1}^{C} w_c\\, q_c
|
|
954
|
+
\left(-\log p_c\right), \qquad p_t = \sum_{c=1}^{C} q_c\\, p_c
|
|
955
955
|
$$
|
|
956
956
|
|
|
957
957
|
Two details matter for the reader combining these flags:
|
|
@@ -1611,8 +1611,8 @@ The head is selected through `--model_arg`, not a dedicated flag:
|
|
|
1611
1611
|
```
|
|
1612
1612
|
|
|
1613
1613
|
The `classifier` value is an inline spec: the head name (a registered
|
|
1614
|
-
built-in
|
|
1615
|
-
`
|
|
1614
|
+
built-in \u2014 `mlp`, `cls_attention` \u2014 or a path to a `.py` file defining a
|
|
1615
|
+
`create_classifier(num_labels, hidden_size, **kwargs)` function), optionally followed by `:key=value,...` kwargs. Values
|
|
1616
1616
|
are parsed as int/float/bool, and bracketed lists such as
|
|
1617
1617
|
`cls_slice=(0, 1)` work. A head is required: `cls_model_wrapper` fails fast
|
|
1618
1618
|
with a clear message when `classifier` is missing. Models that embed their
|
|
@@ -0,0 +1,113 @@
|
|
|
1
|
+
"""Multi-crop augmentation for DINO.
|
|
2
|
+
|
|
3
|
+
Produces 2 global crops (large) and N local crops (small) from a single
|
|
4
|
+
image, with separate augmentation pipelines for each.
|
|
5
|
+
|
|
6
|
+
The split follows the DINO recipe (Caron et al., ICCV 2021 —
|
|
7
|
+
https://arxiv.org/abs/2104.14294): global crops cover a large fraction
|
|
8
|
+
of the original image area (default 0.4–1.0), local crops a small one
|
|
9
|
+
(default 0.05–0.4). The contrast between large and small resolutions is
|
|
10
|
+
what teaches the student that local views must match the global
|
|
11
|
+
representation.
|
|
12
|
+
"""
|
|
13
|
+
|
|
14
|
+
import torch
|
|
15
|
+
from torch import nn
|
|
16
|
+
from torchvision.transforms import v2
|
|
17
|
+
|
|
18
|
+
|
|
19
|
+
class MultiCropTransform(nn.Module):
|
|
20
|
+
"""Generate global + local crops with distinct augmentation strengths.
|
|
21
|
+
|
|
22
|
+
Calling the transform on a PIL image returns a list of
|
|
23
|
+
``2 + local_num`` tensors: index 0 and 1 are the global crops
|
|
24
|
+
(``global_size`` px, `global_transform` pipeline), indices 2.. are the
|
|
25
|
+
local crops (``local_size`` px, `local_transform` pipeline). Use
|
|
26
|
+
:meth:`split_crops` to convert the list into the
|
|
27
|
+
``(global_crops, local_crops)`` stacked tensors expected by
|
|
28
|
+
:meth:`genml_kit.models.dino.DINO.forward`.
|
|
29
|
+
|
|
30
|
+
Args:
|
|
31
|
+
global_size: Spatial size of global crops.
|
|
32
|
+
local_size: Spatial size of local crops.
|
|
33
|
+
global_scale: (min, max) fraction of image area for global crops.
|
|
34
|
+
local_scale: (min, max) fraction of image area for local crops.
|
|
35
|
+
local_num: Number of local crops.
|
|
36
|
+
global_transform: Transform applied to global crops.
|
|
37
|
+
local_transform: Transform applied to local crops.
|
|
38
|
+
"""
|
|
39
|
+
|
|
40
|
+
def __init__(
|
|
41
|
+
self,
|
|
42
|
+
global_size,
|
|
43
|
+
local_size=96,
|
|
44
|
+
global_scale=(0.4, 1.0),
|
|
45
|
+
local_scale=(0.05, 0.4),
|
|
46
|
+
local_num=8,
|
|
47
|
+
global_transform=None,
|
|
48
|
+
local_transform=None,
|
|
49
|
+
):
|
|
50
|
+
super().__init__()
|
|
51
|
+
self.global_size = global_size
|
|
52
|
+
self.local_size = local_size
|
|
53
|
+
self.local_num = local_num
|
|
54
|
+
self.global_crop = v2.RandomResizedCrop(global_size,
|
|
55
|
+
scale=global_scale,
|
|
56
|
+
interpolation=v2.InterpolationMode.BICUBIC)
|
|
57
|
+
self.local_crop = v2.RandomResizedCrop(local_size,
|
|
58
|
+
scale=local_scale,
|
|
59
|
+
interpolation=v2.InterpolationMode.BICUBIC)
|
|
60
|
+
self.global_transform = global_transform or _default_global_transform(global_size)
|
|
61
|
+
self.local_transform = local_transform or _default_local_transform()
|
|
62
|
+
|
|
63
|
+
def forward(self, image):
|
|
64
|
+
"""Return list of tensors: [global1, global2, local1, ..., localN]."""
|
|
65
|
+
crops = []
|
|
66
|
+
for _ in range(2):
|
|
67
|
+
crop = self.global_crop(image)
|
|
68
|
+
crops.append(self.global_transform(crop))
|
|
69
|
+
for _ in range(self.local_num):
|
|
70
|
+
crop = self.local_crop(image)
|
|
71
|
+
crops.append(self.local_transform(crop))
|
|
72
|
+
return crops
|
|
73
|
+
|
|
74
|
+
def split_crops(self, crops):
|
|
75
|
+
"""Split a list of crops into (global_crops, local_crops) tensors."""
|
|
76
|
+
global_crops = torch.stack(crops[:2])
|
|
77
|
+
local_crops = torch.stack(crops[2:])
|
|
78
|
+
return global_crops, local_crops
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
def _default_global_transform(image_size):
|
|
82
|
+
"""Standard DINO global-view pipeline: flip, jitter, blur, normalise."""
|
|
83
|
+
return v2.Compose([
|
|
84
|
+
v2.RandomHorizontalFlip(p=0.5),
|
|
85
|
+
v2.RandomApply(
|
|
86
|
+
[
|
|
87
|
+
v2.ColorJitter(brightness=0.4, contrast=0.6, saturation=0.8, hue=0.1),
|
|
88
|
+
],
|
|
89
|
+
p=0.8,
|
|
90
|
+
),
|
|
91
|
+
v2.RandomGrayscale(p=0.2),
|
|
92
|
+
v2.RandomApply([v2.GaussianBlur(kernel_size=23, sigma=(0.1, 2.0))], p=1.0),
|
|
93
|
+
v2.RandomSolarize(threshold=0.5, p=0.2),
|
|
94
|
+
v2.ToImage(),
|
|
95
|
+
v2.ToDtype(torch.float32, scale=True),
|
|
96
|
+
])
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
def _default_local_transform():
|
|
100
|
+
"""Standard DINO local-view pipeline: same recipe as the global one."""
|
|
101
|
+
return v2.Compose([
|
|
102
|
+
v2.RandomHorizontalFlip(p=0.5),
|
|
103
|
+
v2.RandomApply(
|
|
104
|
+
[
|
|
105
|
+
v2.ColorJitter(brightness=0.4, contrast=0.6, saturation=0.8, hue=0.1),
|
|
106
|
+
],
|
|
107
|
+
p=0.8,
|
|
108
|
+
),
|
|
109
|
+
v2.RandomGrayscale(p=0.2),
|
|
110
|
+
v2.RandomApply([v2.GaussianBlur(kernel_size=23, sigma=(0.1, 2.0))], p=0.5),
|
|
111
|
+
v2.ToImage(),
|
|
112
|
+
v2.ToDtype(torch.float32, scale=True),
|
|
113
|
+
])
|
|
@@ -9,6 +9,8 @@ for the PPO training epochs.
|
|
|
9
9
|
import torch
|
|
10
10
|
from torch.utils.data import Dataset
|
|
11
11
|
|
|
12
|
+
from genml_kit.utils.logging import fatal
|
|
13
|
+
|
|
12
14
|
|
|
13
15
|
class RolloutBuffer(Dataset):
|
|
14
16
|
"""Fixed-length on-policy rollout storage.
|
|
@@ -160,8 +162,8 @@ class RolloutBuffer(Dataset):
|
|
|
160
162
|
def __len__(self):
|
|
161
163
|
return self._rollout_len if self._filled else self._ptr
|
|
162
164
|
|
|
163
|
-
def
|
|
164
|
-
"""
|
|
165
|
+
def _build_item_dict(self, idx):
|
|
166
|
+
"""Build an item/batch dict from buffer arrays at *idx*."""
|
|
165
167
|
item = {
|
|
166
168
|
"obs": self._obs[idx],
|
|
167
169
|
"action": self._actions[idx],
|
|
@@ -174,6 +176,10 @@ class RolloutBuffer(Dataset):
|
|
|
174
176
|
item["raw_action"] = self._raw_actions[idx]
|
|
175
177
|
return item
|
|
176
178
|
|
|
179
|
+
def __getitem__(self, idx):
|
|
180
|
+
"""Return a single timestep as a dict (DataLoader-compatible)."""
|
|
181
|
+
return self._build_item_dict(idx)
|
|
182
|
+
|
|
177
183
|
def to(self, device):
|
|
178
184
|
"""Move all tensors to *device* (in-place)."""
|
|
179
185
|
self._obs = self._obs.to(device)
|
|
@@ -224,20 +230,8 @@ class RolloutBuffer(Dataset):
|
|
|
224
230
|
else:
|
|
225
231
|
# D5: Use independent RNG for reproducible sampling
|
|
226
232
|
idx = torch.randint(n, (batch_size,), generator=self._rng)
|
|
227
|
-
|
|
228
|
-
|
|
229
|
-
"action": self._actions[idx],
|
|
230
|
-
"log_prob": self._log_probs[idx],
|
|
231
|
-
"advantage": self._advantages[idx],
|
|
232
|
-
"return": self._returns[idx],
|
|
233
|
-
"value": self._values[idx],
|
|
234
|
-
}
|
|
235
|
-
if self._raw_actions is not None:
|
|
236
|
-
# Raw (pre-tanh) actions are only stored for continuous policies;
|
|
237
|
-
# without them the PPO ratio is computed against squashed actions
|
|
238
|
-
# while ``old_log_prob`` refers to raw actions (see docstring).
|
|
239
|
-
batch["raw_action"] = self._raw_actions[idx]
|
|
240
|
-
return batch
|
|
233
|
+
batch = self._build_item_dict(idx)
|
|
234
|
+
return batch
|
|
241
235
|
|
|
242
236
|
# Alias for backward compatibility
|
|
243
237
|
get_batch = sample
|
|
@@ -312,47 +306,41 @@ class RolloutBuffer(Dataset):
|
|
|
312
306
|
def device(self):
|
|
313
307
|
return self._device
|
|
314
308
|
|
|
309
|
+
# State dict field mapping: (state_key, attribute_name)
|
|
310
|
+
_STATE_FIELDS = [
|
|
311
|
+
("obs", "_obs"),
|
|
312
|
+
("actions", "_actions"),
|
|
313
|
+
("raw_actions", "_raw_actions"),
|
|
314
|
+
("log_probs", "_log_probs"),
|
|
315
|
+
("rewards", "_rewards"),
|
|
316
|
+
("values", "_values"),
|
|
317
|
+
("dones", "_dones"),
|
|
318
|
+
("terminated", "_terminated"),
|
|
319
|
+
("advantages", "_advantages"),
|
|
320
|
+
("returns", "_returns"),
|
|
321
|
+
("next_values", "_next_values"),
|
|
322
|
+
("ptr", "_ptr"),
|
|
323
|
+
("filled", "_filled"),
|
|
324
|
+
("obs_dim", "_obs_dim"),
|
|
325
|
+
("action_dim", "_action_dim"),
|
|
326
|
+
("rollout_len", "_rollout_len"),
|
|
327
|
+
("device", "_device"),
|
|
328
|
+
]
|
|
329
|
+
|
|
315
330
|
def state_dict(self):
|
|
316
|
-
"""Return
|
|
317
|
-
|
|
318
|
-
|
|
319
|
-
|
|
320
|
-
"raw_actions": self._raw_actions,
|
|
321
|
-
"log_probs": self._log_probs,
|
|
322
|
-
"rewards": self._rewards,
|
|
323
|
-
"values": self._values,
|
|
324
|
-
"dones": self._dones,
|
|
325
|
-
"terminated": self._terminated,
|
|
326
|
-
"advantages": self._advantages,
|
|
327
|
-
"returns": self._returns,
|
|
328
|
-
"next_values": self._next_values,
|
|
329
|
-
"ptr": self._ptr,
|
|
330
|
-
"filled": self._filled,
|
|
331
|
-
"obs_dim": self._obs_dim,
|
|
332
|
-
"action_dim": self._action_dim,
|
|
333
|
-
"rollout_len": self._rollout_len,
|
|
334
|
-
"device": self._device,
|
|
335
|
-
"rng_state": self._rng.get_state(),
|
|
336
|
-
}
|
|
331
|
+
"""Return a serializable checkpoint dict."""
|
|
332
|
+
d = {k: getattr(self, attr) for k, attr in self._STATE_FIELDS}
|
|
333
|
+
d["rng_state"] = self._rng.get_state()
|
|
334
|
+
return d
|
|
337
335
|
|
|
338
336
|
def load_state_dict(self, state):
|
|
339
337
|
"""Load state from checkpoint."""
|
|
340
|
-
self.
|
|
341
|
-
|
|
338
|
+
for k, attr in self._STATE_FIELDS:
|
|
339
|
+
if k not in state:
|
|
340
|
+
fatal(f"Missing key '{k}' in state_dict", KeyError)
|
|
341
|
+
setattr(self, attr, state[k])
|
|
342
|
+
# Optional fields with defaults (backward compat)
|
|
342
343
|
self._raw_actions = state.get("raw_actions", self._raw_actions)
|
|
343
|
-
self._log_probs = state["log_probs"]
|
|
344
|
-
self._rewards = state["rewards"]
|
|
345
|
-
self._values = state["values"]
|
|
346
|
-
self._dones = state["dones"]
|
|
347
344
|
self._terminated = state.get("terminated", self._terminated)
|
|
348
|
-
self._advantages = state["advantages"]
|
|
349
|
-
self._returns = state["returns"]
|
|
350
|
-
self._next_values = state["next_values"]
|
|
351
|
-
self._ptr = state["ptr"]
|
|
352
|
-
self._filled = state["filled"]
|
|
353
|
-
self._obs_dim = state["obs_dim"]
|
|
354
|
-
self._action_dim = state["action_dim"]
|
|
355
|
-
self._rollout_len = state["rollout_len"]
|
|
356
|
-
self._device = state["device"]
|
|
357
345
|
if "rng_state" in state:
|
|
358
346
|
self._rng.set_state(state["rng_state"])
|
|
@@ -16,8 +16,20 @@ import torch
|
|
|
16
16
|
|
|
17
17
|
from genml_kit.geometry.similarity import umeyama_similarity
|
|
18
18
|
|
|
19
|
-
|
|
20
|
-
|
|
19
|
+
# ``corners_src`` / ``corners_dst`` are the four image corners in *pixel*
|
|
20
|
+
# coordinates and where they truly land, i.e. the two sides of the mean
|
|
21
|
+
# corner error of vo/README.md s6. ``corners_dst`` comes from the
|
|
22
|
+
# homography, not from the similarity fit, so the irreducible
|
|
23
|
+
# foreshortening residual (``gt_residual``) stays visible instead of being
|
|
24
|
+
# folded away.
|
|
25
|
+
VOPairMeta = collections.namedtuple("VOPairMeta", [
|
|
26
|
+
"gt",
|
|
27
|
+
"gt_residual",
|
|
28
|
+
"terrain",
|
|
29
|
+
"range_bin",
|
|
30
|
+
"corners_src",
|
|
31
|
+
"corners_dst",
|
|
32
|
+
])
|
|
21
33
|
|
|
22
34
|
|
|
23
35
|
def look_at_ground_h(cam_pos, cam_rpy, cam, _size):
|
|
@@ -104,7 +116,15 @@ def homography_to_similarity(h, size):
|
|
|
104
116
|
size: (W, H) of the image.
|
|
105
117
|
|
|
106
118
|
Returns:
|
|
107
|
-
(sim_params dict with log_s/theta/t tensors, residual_px float
|
|
119
|
+
(sim_params dict with log_s/theta/t tensors, residual_px float,
|
|
120
|
+
corners float64 (4, 2), mapped float64 (4, 2)).
|
|
121
|
+
|
|
122
|
+
The two corner arrays are the integer source corners and their
|
|
123
|
+
true homography-mapped destinations, both in *pixel* coordinates
|
|
124
|
+
(vo/README.md s11.4 convention). They are what the mean corner error
|
|
125
|
+
of s6 is defined against, so they are returned rather than discarded;
|
|
126
|
+
previously only the fitted parameters survived, which left the loss
|
|
127
|
+
and the metric with no ground-truth corners to compare against.
|
|
108
128
|
"""
|
|
109
129
|
width, height = size
|
|
110
130
|
corners = np.array([
|
|
@@ -128,7 +148,7 @@ def homography_to_similarity(h, size):
|
|
|
128
148
|
"theta": fit.theta[0].item(),
|
|
129
149
|
"t": fit.t[0].tolist(),
|
|
130
150
|
}
|
|
131
|
-
return sim, residual
|
|
151
|
+
return sim, residual, src[0], dst[0]
|
|
132
152
|
|
|
133
153
|
|
|
134
154
|
def params_to_matrix_np(log_s, theta, t):
|
|
@@ -312,7 +332,8 @@ class VOPairDataset:
|
|
|
312
332
|
h_b = look_at_ground_h(cam_b, rpy_b, self._cam, self._size)
|
|
313
333
|
frame_a = self._render(base, h_a)
|
|
314
334
|
frame_b = self._render(base, h_b)
|
|
315
|
-
sim, residual = homography_to_similarity(
|
|
335
|
+
sim, residual, corners_src, corners_dst = homography_to_similarity(
|
|
336
|
+
h_b @ np.linalg.inv(h_a), self._size)
|
|
316
337
|
gt = {
|
|
317
338
|
"log_s": torch.tensor(sim["log_s"]),
|
|
318
339
|
"theta": torch.tensor(sim["theta"]),
|
|
@@ -325,5 +346,7 @@ class VOPairDataset:
|
|
|
325
346
|
terrain=self._terrain,
|
|
326
347
|
range_bin=int(
|
|
327
348
|
np.clip(np.searchsorted([0.1, 0.2, 0.3], abs(motion["dx"])),
|
|
328
|
-
0, 3))
|
|
349
|
+
0, 3)),
|
|
350
|
+
corners_src=corners_src.to(torch.float32),
|
|
351
|
+
corners_dst=corners_dst.to(torch.float32))
|
|
329
352
|
return {"image_a": image_a, "image_b": image_b, "meta": meta}
|
|
@@ -6,10 +6,12 @@ Both scripts import these functions rather than maintaining separate copies.
|
|
|
6
6
|
|
|
7
7
|
import io
|
|
8
8
|
import logging
|
|
9
|
+
import ntpath
|
|
9
10
|
import os
|
|
10
11
|
import re
|
|
11
12
|
import tarfile
|
|
12
13
|
import tempfile
|
|
14
|
+
from pathlib import PurePosixPath
|
|
13
15
|
|
|
14
16
|
import torch
|
|
15
17
|
|
|
@@ -86,6 +88,27 @@ def serialize_lora_state(model):
|
|
|
86
88
|
return blob
|
|
87
89
|
|
|
88
90
|
|
|
91
|
+
def _extract_lora_archive(tar, destination):
|
|
92
|
+
"""Extract a LoRA archive safely by validating members manually.
|
|
93
|
+
|
|
94
|
+
Does not rely on ``tarfile``'s ``filter=`` argument (only available on
|
|
95
|
+
Python >= 3.12): we always iterate the members ourselves, which gives the
|
|
96
|
+
same safety guarantees on every supported Python version, and is also
|
|
97
|
+
compatible with code that wraps ``TarFile.extractall``.
|
|
98
|
+
"""
|
|
99
|
+
members = tar.getmembers()
|
|
100
|
+
for member in members:
|
|
101
|
+
name = member.name
|
|
102
|
+
path = PurePosixPath(name)
|
|
103
|
+
if (path.is_absolute() or ".." in path.parts or "\\" in name or
|
|
104
|
+
ntpath.splitdrive(name)[0]):
|
|
105
|
+
raise ValueError(f"Unsafe path in LoRA archive: {name!r}")
|
|
106
|
+
if not (member.isdir() or member.isfile()):
|
|
107
|
+
raise ValueError(f"Unsupported LoRA archive member: {name!r}")
|
|
108
|
+
|
|
109
|
+
tar.extractall(destination, members)
|
|
110
|
+
|
|
111
|
+
|
|
89
112
|
def deserialize_lora_state(model, blob):
|
|
90
113
|
"""Restore PEFT adapter state from a tar blob.
|
|
91
114
|
|
|
@@ -101,7 +124,7 @@ def deserialize_lora_state(model, blob):
|
|
|
101
124
|
with tempfile.TemporaryDirectory() as tmpdir:
|
|
102
125
|
buf = io.BytesIO(blob)
|
|
103
126
|
with tarfile.open(fileobj=buf, mode="r") as tar:
|
|
104
|
-
tar
|
|
127
|
+
_extract_lora_archive(tar, tmpdir)
|
|
105
128
|
# When the model is already a PeftModel (e.g. resume_checkpoint called
|
|
106
129
|
# after apply_lora), PeftModel.from_pretrained would double-wrap it,
|
|
107
130
|
# producing mangled keys like "base_model.model.base_model.model.*".
|
|
@@ -25,7 +25,7 @@ import logging
|
|
|
25
25
|
|
|
26
26
|
from genml_kit.utils.logging import fatal
|
|
27
27
|
from genml_kit.utils.registry import Registry
|
|
28
|
-
from genml_kit.utils.script import load_extern
|
|
28
|
+
from genml_kit.utils.script import load_extern, is_script_spec
|
|
29
29
|
|
|
30
30
|
LOSSES = Registry("loss")
|
|
31
31
|
|
|
@@ -33,12 +33,6 @@ LOSSES = Registry("loss")
|
|
|
33
33
|
DEFAULT_LOSS = "focal"
|
|
34
34
|
|
|
35
35
|
|
|
36
|
-
def _is_script_spec(spec):
|
|
37
|
-
"""Return *True* if *spec* looks like a script path/URL rather than a name."""
|
|
38
|
-
return bool(spec) and (spec.endswith(".py") or spec.startswith(
|
|
39
|
-
("http://", "https://")))
|
|
40
|
-
|
|
41
|
-
|
|
42
36
|
def _load_loss_script(spec, **kwargs):
|
|
43
37
|
"""Load a loss from an external script defining ``build_loss(**kwargs)``."""
|
|
44
38
|
build_loss = load_extern(spec, "build_loss")
|
|
@@ -70,7 +64,7 @@ def load_loss(spec=DEFAULT_LOSS, *, protocol=None, **kwargs):
|
|
|
70
64
|
"""
|
|
71
65
|
if spec is None:
|
|
72
66
|
spec = DEFAULT_LOSS
|
|
73
|
-
if
|
|
67
|
+
if is_script_spec(spec):
|
|
74
68
|
built = _load_loss_script(spec, **kwargs)
|
|
75
69
|
return _merge_loss_bundle(built, protocol, kwargs)
|
|
76
70
|
if LOSSES.contains(spec):
|
|
@@ -94,11 +94,11 @@ def td_loss(pred_q, target, reduction="mean"):
|
|
|
94
94
|
def gae(rewards, values, next_values, dones, gamma, lam, terminated=None):
|
|
95
95
|
"""Compute Generalized Advantage Estimation (GAE).
|
|
96
96
|
|
|
97
|
-
Implements the backward recurrence from ``rl/README.md`` §10.3
|
|
97
|
+
Implements the backward recurrence from ``rl/README.md`` §10.3, using
|
|
98
|
+
separate masks for bootstrapping and trace continuation:
|
|
98
99
|
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
where δ_t = r_t + γ (1 − d_t) V(s_{t+1}) − V(s_t).
|
|
100
|
+
δ_t = r_t + γ (1 − terminated_t) V(s_{t+1}) − V(s_t)
|
|
101
|
+
A_t = δ_t + γ λ (1 − done_t) A_{t+1}
|
|
102
102
|
|
|
103
103
|
All inputs are (T,) or (T, 1) tensors over a single rollout
|
|
104
104
|
(T = rollout length).
|
|
@@ -106,15 +106,13 @@ def gae(rewards, values, next_values, dones, gamma, lam, terminated=None):
|
|
|
106
106
|
Args:
|
|
107
107
|
rewards: (T,) immediate rewards.
|
|
108
108
|
values: (T,) value estimates V(s_t).
|
|
109
|
-
next_values: (T,) value estimates V(s_{t+1})
|
|
110
|
-
the critic at the next state; zero-padded at episode
|
|
111
|
-
boundaries).
|
|
109
|
+
next_values: (T,) value estimates V(s_{t+1}) from the critic.
|
|
112
110
|
dones: (T,) float flags (0.0 / 1.0).
|
|
113
111
|
gamma: discount factor.
|
|
114
112
|
lam: GAE lambda (typically 0.95).
|
|
115
|
-
terminated: optional (T,) float tensor — true MDP end flags.
|
|
116
|
-
given, ``(1 − terminated)`` is used
|
|
117
|
-
|
|
113
|
+
terminated: optional (T,) float tensor — true MDP end flags. When
|
|
114
|
+
given, ``(1 − terminated)`` is used for bootstrapping while
|
|
115
|
+
``(1 − dones)`` always stops the trace at episode boundaries.
|
|
118
116
|
|
|
119
117
|
Returns:
|
|
120
118
|
advantages: (T,) GAE advantage estimates.
|
|
@@ -131,14 +129,17 @@ def gae(rewards, values, next_values, dones, gamma, lam, terminated=None):
|
|
|
131
129
|
advantages = torch.zeros_like(rewards)
|
|
132
130
|
last_gae = 0.0
|
|
133
131
|
|
|
134
|
-
#
|
|
135
|
-
#
|
|
136
|
-
|
|
132
|
+
# Use the true-termination flag for bootstrapping when available, so a
|
|
133
|
+
# truncated step still bootstraps its final observation. The GAE trace
|
|
134
|
+
# must stop at the episode boundary to avoid crossing into the next
|
|
135
|
+
# episode in the rollout. Legacy callers without ``terminated`` use done
|
|
136
|
+
# for both masks, preserving their historical behavior.
|
|
137
|
+
bootstrap_mask = 1.0 - dones if terminated is None else 1.0 - terminated
|
|
138
|
+
trace_mask = 1.0 - dones
|
|
137
139
|
|
|
138
140
|
for t in reversed(range(T)):
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
last_gae = delta + gamma * lam * non_terminal * last_gae
|
|
141
|
+
delta = (rewards[t] + gamma * next_values[t] * bootstrap_mask[t] - values[t])
|
|
142
|
+
last_gae = delta + gamma * lam * trace_mask[t] * last_gae
|
|
142
143
|
advantages[t] = last_gae
|
|
143
144
|
|
|
144
145
|
returns = advantages + values
|