opensportslib 0.3.1.dev14__tar.gz → 0.3.1.dev15__tar.gz

{opensportslib-0.3.1.dev14/opensportslib.egg-info → opensportslib-0.3.1.dev15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.3.1.dev14
+Version: 0.3.1.dev15
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/opensportslib/adaptation/spotta.py

@@ -59,7 +59,6 @@ class SpoTTAConfig:
     drift_scale: float = 10.0
     drift_threshold: float = 1.0
     drift_gamma: float = 0.2
-    augmentation: bool = True
 
     @classmethod
     def from_mapping(cls, value: Any) -> "SpoTTAConfig":
@@ -70,7 +69,6 @@ class SpoTTAConfig:
         memory = _mapping(root.get("memory"))
         optimizer = _mapping(root.get("optimizer"))
         teacher = _mapping(root.get("teacher"))
-        augmentation = _mapping(root.get("augmentation"))
 
         _require_supported_value(
             root,
@@ -78,34 +76,6 @@ class SpoTTAConfig:
             "adapt_then_predict",
             "adaptation",
         )
-        _require_supported_value(tether, "mode", "bayesian", "robust_bn.tether")
-        _require_supported_value(
-            gate,
-            "uncertainty",
-            "one_minus_max_probability",
-            "confidence_gate",
-        )
-        _require_supported_value(
-            gate,
-            "aggregation",
-            "min_over_predicted_action_frames",
-            "confidence_gate",
-        )
-        _require_supported_value(optimizer, "type", "Adam", "optimizer")
-        _require_supported_value(
-            optimizer,
-            "trainable_parameters",
-            "batch_norm_affine_only",
-            "optimizer",
-        )
-        _require_supported_value(teacher, "type", "ema", "teacher")
-        _require_supported_value(
-            teacher, "adaptive_from_bn_drift", True, "teacher"
-        )
-        _require_supported_value(
-            augmentation, "mode", "framewise_rotta_strong", "augmentation"
-        )
-
         config = cls(
             alpha=float(robust_bn.get("alpha", cls.alpha)),
             tether_cap=float(tether.get("cap", cls.tether_cap)),
@@ -133,7 +103,6 @@ class SpoTTAConfig:
                 teacher.get("drift_threshold", cls.drift_threshold)
             ),
             drift_gamma=float(teacher.get("drift_gamma", cls.drift_gamma)),
-            augmentation=bool(augmentation.get("enabled", cls.augmentation)),
         )
         config.validate()
         return config
@@ -577,13 +546,10 @@ class SpoTTA:
         with torch.no_grad():
             teacher_logits = _logits(self.teacher(clips))
 
-        if self.config.augmentation:
-            batch, time, channels, height, width = clips.shape
-            augmented = self.augmentation(
-                clips.reshape(batch * time, channels, height, width)
-            ).reshape(batch, time, channels, height, width)
-        else:
-            augmented = clips
+        batch, time, channels, height, width = clips.shape
+        augmented = self.augmentation(
+            clips.reshape(batch * time, channels, height, width)
+        ).reshape(batch, time, channels, height, width)
 
         self.student.train()
         student_logits = _logits(self.student(augmented))

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/opensportslib/configs/localization/e2e_spotta.yaml

@@ -39,6 +39,7 @@ DATA:
         resize:
           height: 224
           width: 398
+          preserve_aspect_ratio: true
       params:
         crop_dim: 224
         dilate_len: 0
@@ -69,13 +70,10 @@ MODEL:
       robust_bn:
         alpha: 0.05
         tether:
-          mode: bayesian
           cap: 0.5
       confidence_gate:
         action_class_index: 1
         min_action_frames: 1
-        uncertainty: one_minus_max_probability
-        aggregation: min_over_predicted_action_frames
         threshold: 0.3
       memory:
         capacity: 8
@@ -83,21 +81,14 @@ MODEL:
         lambda_t: 1.0
         lambda_u: 1.0
       optimizer:
-        type: Adam
         learning_rate: 0.001
         beta: 0.9
-        trainable_parameters: batch_norm_affine_only
       teacher:
-        type: ema
         base_nu: 0.001
-        adaptive_from_bn_drift: true
         max_nu: 0.02
         drift_scale: 10.0
         drift_threshold: 1.0
         drift_gamma: 0.2
-      augmentation:
-        enabled: true
-        mode: framewise_rotta_strong
 
 TRAIN:
   execution:

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/opensportslib/datasets/localization_dataset.py

@@ -161,6 +161,7 @@ class LocalizationDataset(Dataset):
             imagenet_std=normalization.get("std"),
             target_height=resize.get("height"),
             target_width=resize.get("width"),
+            preserve_aspect_ratio=resize.get("preserve_aspect_ratio", False),
         )
         annotation_path = annotations_path or getattr(
             split_cfg, "annotation_path", getattr(split_cfg, "path", None)
@@ -346,6 +347,7 @@ class LocalizationDataset(Dataset):
                 IMAGENET_STD=self.data_cfg.imagenet_std,
                 TARGET_HEIGHT=self.data_cfg.target_height,
                 TARGET_WIDTH=self.data_cfg.target_width,
+                preserve_aspect_ratio=self.data_cfg.preserve_aspect_ratio,
             )
         elif dataset_type == "VideoGameWithOpencvVideo":
             dataset = ActionSpotVideoDataset(
@@ -362,6 +364,7 @@ class LocalizationDataset(Dataset):
                 IMAGENET_STD=self.data_cfg.imagenet_std,
                 TARGET_HEIGHT=self.data_cfg.target_height,
                 TARGET_WIDTH=self.data_cfg.target_width,
+                preserve_aspect_ratio=self.data_cfg.preserve_aspect_ratio,
             )
         elif dataset_type == "VideoGameWithDali":
             if not DALI_AVAILABLE:
@@ -543,6 +546,7 @@ class FrameReader:
         sample_fps=2,
         TARGET_HEIGHT=224,
         TARGET_WIDTH=398,
+        preserve_aspect_ratio=False,
     ):
         self._is_flow = modality == "flow"
         self._crop_transform = crop_transform
@@ -551,6 +555,37 @@ class FrameReader:
         self._sample_fps = sample_fps
         self.TARGET_HEIGHT = TARGET_HEIGHT
         self.TARGET_WIDTH = TARGET_WIDTH
+        self.preserve_aspect_ratio = preserve_aspect_ratio
+
+    def _resize_frame_ocv(self, frame):
+        """Resize an OpenCV frame using the configured spatial policy.
+
+        With aspect preservation enabled, ``TARGET_HEIGHT`` fixes the output
+        height and the width is derived from the source aspect ratio. If no
+        target height is configured, ``TARGET_WIDTH`` fixes the output width.
+        """
+        import cv2
+
+        height, width = frame.shape[:2]
+        if self.preserve_aspect_ratio:
+            if self.TARGET_HEIGHT is not None and self.TARGET_HEIGHT > 0:
+                scale = self.TARGET_HEIGHT / height
+            elif self.TARGET_WIDTH is not None and self.TARGET_WIDTH > 0:
+                scale = self.TARGET_WIDTH / width
+            else:
+                return frame
+
+            output_height = max(1, round(height * scale))
+            output_width = max(1, round(width * scale))
+        else:
+            output_height = self.TARGET_HEIGHT
+            output_width = self.TARGET_WIDTH
+
+        if output_height is None or output_width is None:
+            return frame
+        if (height, width) == (output_height, output_width):
+            return frame
+        return cv2.resize(frame, (output_width, output_height))
 
     def adapt_frame_ocv(self, frame):
         """Apply some modifications to the frame to have the expected shape and format.
@@ -675,8 +710,7 @@ class FrameReader:
             ret, frame = vc.read()
             if ret:
                 if i % stride_extract == 0:
-                    if frame.shape[0] != oh or frame.shape[1] != ow:
-                        frame = cv2.resize(frame, (ow, oh))
+                    frame = self._resize_frame_ocv(frame)
                     img = self.adapt_frame_ocv(frame)
                     if self._crop_transform:
                         if self._same_transform:
@@ -781,6 +815,7 @@ class ActionSpotDataset(Dataset):
         IMAGENET_STD=[0.229, 0.224, 0.225],
         TARGET_HEIGHT=224,
         TARGET_WIDTH=398,
+        preserve_aspect_ratio=False,
     ):
         import random
         from opensportslib.core.utils.load_annotations import annotationstoe2eformat
@@ -850,7 +885,8 @@ class ActionSpotDataset(Dataset):
             same_transform,
             extract_fps,
             self.TARGET_HEIGHT,
-            self.TARGET_WIDTH
+            self.TARGET_WIDTH,
+            preserve_aspect_ratio,
         )
 
     def load_frame_gpu(self, batch, device):
@@ -1097,6 +1133,7 @@ class ActionSpotVideoDataset(Dataset, DatasetVideoSharedMethods):
         IMAGENET_STD=[0.229, 0.224, 0.225],
         TARGET_HEIGHT=224,
         TARGET_WIDTH=398,
+        preserve_aspect_ratio=False,
     ):
         from opensportslib.core.utils.load_annotations import annotationstoe2eformat, construct_labels
         from opensportslib.core.utils.video_processing import _get_img_transforms
@@ -1138,7 +1175,8 @@ class ActionSpotVideoDataset(Dataset, DatasetVideoSharedMethods):
             False,
             extract_fps,
             self.TARGET_HEIGHT,
-            self.TARGET_WIDTH
+            self.TARGET_WIDTH,
+            preserve_aspect_ratio,
         )
 
         self._flip = flip

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15/opensportslib.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.3.1.dev14
+Version: 0.3.1.dev15
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/opensportslib.egg-info/SOURCES.txt

@@ -169,7 +169,6 @@ tests/test_release_versions.py
 tests/test_remote_model_registry.py
 tests/test_server_model_registry.py
 tests/test_setup_cli.py
-tests/test_spotta_e2e.py
 tests/test_subset_train_infer_integration.py
 tests/test_task_model_api_contract.py
 tests/test_vqa_api.py

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opensportslib"
-version = "0.3.1.dev14"
+version = "0.3.1.dev15"
 description = "OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data."
 readme = "README.md"
 requires-python = ">=3.12"

{opensportslib-0.3.1.dev14 → opensportslib-0.3.1.dev15}/tests/test_localization_intervals.py

@@ -209,6 +209,26 @@ def test_frame_reader_seeks_and_stops_inside_interval(monkeypatch):
     assert max(capture.read_positions) < 50
 
 
+def test_frame_reader_can_preserve_aspect_ratio_before_crop():
+    reader = FrameReader(
+        "rgb",
+        crop_transform=None,
+        img_transform=lambda image: image,
+        same_transform=False,
+        TARGET_HEIGHT=224,
+        TARGET_WIDTH=398,
+        preserve_aspect_ratio=True,
+    )
+
+    four_by_three = np.zeros((224, 298, 3), dtype=np.uint8)
+    sixteen_by_nine = np.zeros((224, 398, 3), dtype=np.uint8)
+    scaled_four_by_three = np.zeros((448, 596, 3), dtype=np.uint8)
+
+    assert reader._resize_frame_ocv(four_by_three).shape == (224, 298, 3)
+    assert reader._resize_frame_ocv(sixteen_by_nine).shape == (224, 398, 3)
+    assert reader._resize_frame_ocv(scaled_four_by_three).shape == (224, 298, 3)
+
+
 def test_v2_evaluator_scores_only_verified_logical_intervals(tmp_path):
     verified = _record()
     unlabeled = _record(status="unlabeled")

opensportslib-0.3.1.dev14/tests/test_spotta_e2e.py

@@ -1,285 +0,0 @@
-from __future__ import annotations
-
-from types import SimpleNamespace
-from unittest.mock import patch
-
-import torch
-import pytest
-from torch import nn
-
-from opensportslib.core.config.accessors import get_loader_backend
-from opensportslib.models.base.e2e import E2EModel
-from opensportslib.adaptation.spotta import (
-    FramewiseStrongAugmentation,
-    RobustBatchNorm,
-    SpoTTA,
-    SpoTTAConfig,
-    replace_batch_norm,
-)
-from opensportslib.core.config import load_config
-
-
-class _TinyE2ESpot(nn.Module):
-    def __init__(self, num_classes=2, predicted_class=1):
-        super().__init__()
-        self.bn = nn.BatchNorm2d(3)
-        self.classifier = nn.Linear(3, num_classes)
-        with torch.no_grad():
-            self.classifier.weight.zero_()
-            self.classifier.bias.fill_(-2.0)
-            self.classifier.bias[predicted_class] = 2.0
-
-    def forward(self, clips):
-        batch, time, channels, height, width = clips.shape
-        frames = self.bn(clips.reshape(batch * time, channels, height, width))
-        features = frames.mean(dim=(-1, -2))
-        return self.classifier(features).reshape(batch, time, -1)
-
-
-def _spotta_config(**overrides):
-    config = {
-        "enabled": True,
-        "name": "spotta",
-        "prediction_timing": "adapt_then_predict",
-        "robust_bn": {"alpha": 0.05, "tether": {"cap": 0.5}},
-        "confidence_gate": {
-            "action_class_index": 1,
-            "min_action_frames": 1,
-            "threshold": 0.3,
-        },
-        "memory": {
-            "capacity": 8,
-            "update_frequency": 2,
-            "lambda_t": 1.0,
-            "lambda_u": 1.0,
-        },
-        "optimizer": {"learning_rate": 0.001, "beta": 0.9},
-        "teacher": {
-            "base_nu": 0.001,
-            "max_nu": 0.02,
-            "drift_scale": 10.0,
-            "drift_threshold": 1.0,
-            "drift_gamma": 0.2,
-        },
-        "augmentation": {"enabled": False},
-    }
-    config.update(overrides)
-    return config
-
-
-def test_robust_batch_norm_preserves_eval_output_and_frozen_anchor():
-    torch.manual_seed(7)
-    source = nn.Sequential(nn.BatchNorm2d(3)).eval()
-    with torch.no_grad():
-        source[0].running_mean.copy_(torch.tensor([0.2, -0.1, 0.4]))
-        source[0].running_var.copy_(torch.tensor([0.7, 1.3, 2.0]))
-        source[0].weight.copy_(torch.tensor([0.8, 1.1, 0.9]))
-        source[0].bias.copy_(torch.tensor([-0.2, 0.3, 0.1]))
-    inputs = torch.randn(4, 3, 5, 5)
-    expected = source(inputs)
-
-    layers = replace_batch_norm(source, alpha=0.05)
-    source.eval()
-    actual = source(inputs)
-
-    assert len(layers) == 1
-    assert isinstance(layers[0], RobustBatchNorm)
-    assert torch.allclose(actual, expected, atol=1e-6, rtol=1e-5)
-
-    anchor_mean = layers[0].anchor_mean.clone()
-    source.train()
-    source(inputs + 2)
-    assert torch.equal(layers[0].anchor_mean, anchor_mean)
-    assert not torch.equal(layers[0].source_mean, anchor_mean)
-
-
-def test_spotta_gates_actions_and_updates_every_second_gated_clip():
-    spotta = SpoTTA(_TinyE2ESpot(), _spotta_config())
-    clips = torch.randn(2, 4, 3, 5, 5)
-
-    predicted, probabilities = spotta.predict(clips, use_amp=False)
-
-    assert predicted.shape == (2, 4)
-    assert probabilities.shape == (2, 4, 2)
-    assert spotta.stats["clips_seen"] == 2
-    assert spotta.stats["clips_gated"] == 2
-    assert spotta.stats["memory_insertions"] == 2
-    assert spotta.stats["memory_occupancy"] == 2
-    assert spotta.stats["update_attempts"] == 1
-    assert spotta.stats["updates_completed"] == 1
-    assert {layer.tether_rho for layer in spotta.tether_layers} == {0.25}
-
-
-def test_framewise_strong_augmentation_preserves_clip_frame_shape_and_range():
-    torch.manual_seed(11)
-    frames = torch.randn(3, 3, 16, 20)
-
-    augmented = FramewiseStrongAugmentation(image_size=16)(frames)
-
-    assert augmented.shape == frames.shape
-    assert float(augmented.min()) >= 0.0
-    assert float(augmented.max()) <= 1.0
-
-
-def test_e2e_wrapper_starts_fresh_spotta_without_mutating_source_model():
-    wrapper = E2EModel.__new__(E2EModel)
-    wrapper._model = _TinyE2ESpot()
-    wrapper._num_classes = 2
-    wrapper._multi_gpu = False
-    wrapper._test_time_adapter = None
-    source_state = {
-        name: tensor.clone() for name, tensor in wrapper._model.state_dict().items()
-    }
-
-    wrapper.configure_test_time_adaptation(_spotta_config())
-    wrapper.predict(torch.randn(2, 4, 3, 5, 5), use_amp=False)
-    first_adapter = wrapper._test_time_adapter
-    wrapper.configure_test_time_adaptation(_spotta_config())
-
-    assert wrapper._test_time_adapter is not first_adapter
-    for name, tensor in wrapper._model.state_dict().items():
-        assert torch.equal(tensor, source_state[name])
-
-
-def test_e2e_wrapper_keeps_one_spotta_instance_across_session_batches():
-    wrapper = E2EModel.__new__(E2EModel)
-    wrapper._model = _TinyE2ESpot()
-    wrapper._num_classes = 2
-    wrapper._multi_gpu = False
-    wrapper._test_time_adapter = None
-    wrapper.configure_test_time_adaptation(_spotta_config())
-    session_adapter = wrapper._test_time_adapter
-
-    wrapper.predict(torch.randn(1, 4, 3, 5, 5), use_amp=False)
-    wrapper.predict(torch.randn(1, 4, 3, 5, 5), use_amp=False)
-
-    assert wrapper._test_time_adapter is session_adapter
-    assert wrapper.test_time_adaptation_stats["clips_seen"] == 2
-    assert wrapper.test_time_adaptation_stats["memory_occupancy"] == 2
-    assert wrapper.test_time_adaptation_stats["updates_completed"] == 1
-
-
-def test_disabled_spotta_uses_ordinary_e2e_prediction_path():
-    wrapper = E2EModel.__new__(E2EModel)
-    wrapper._model = _TinyE2ESpot()
-    wrapper._num_classes = 2
-    wrapper._multi_gpu = False
-    wrapper._test_time_adapter = None
-    wrapper.device = torch.device("cpu")
-    clips = torch.randn(2, 4, 3, 5, 5)
-    wrapper._model.eval()
-    with torch.no_grad():
-        expected_probabilities = wrapper._model(clips).softmax(dim=2)
-        expected_classes = expected_probabilities.argmax(dim=2)
-
-    wrapper.configure_test_time_adaptation({"enabled": False, "name": "spotta"})
-    predicted_classes, probabilities = wrapper.predict(clips, use_amp=False)
-
-    assert wrapper._test_time_adapter is None
-    assert torch.equal(torch.from_numpy(predicted_classes), expected_classes)
-    assert torch.allclose(
-        torch.from_numpy(probabilities), expected_probabilities, atol=1e-7
-    )
-
-
-def test_spotta_supports_configured_action_class_in_multiclass_output():
-    config = _spotta_config()
-    config["confidence_gate"]["action_class_index"] = 2
-    spotta = SpoTTA(
-        _TinyE2ESpot(num_classes=4, predicted_class=2), config
-    )
-
-    predicted, probabilities = spotta.predict(
-        torch.randn(2, 4, 3, 5, 5), use_amp=False
-    )
-
-    assert predicted.shape == (2, 4)
-    assert probabilities.shape == (2, 4, 4)
-    assert torch.from_numpy(predicted).eq(2).all()
-    assert spotta.stats["clips_gated"] == 2
-    assert spotta.stats["updates_completed"] == 1
-
-
-def test_spotta_config_rejects_negative_action_class_index():
-    config = _spotta_config()
-    config["confidence_gate"]["action_class_index"] = -1
-
-    with pytest.raises(ValueError, match="must be non-negative"):
-        SpoTTAConfig.from_mapping(config)
-
-
-def test_spotta_rejects_action_class_outside_model_outputs():
-    config = _spotta_config()
-    config["confidence_gate"]["action_class_index"] = 4
-    spotta = SpoTTA(_TinyE2ESpot(num_classes=4), config)
-
-    with pytest.raises(ValueError, match="outside the model's 4 output classes"):
-        spotta.predict(torch.randn(1, 4, 3, 5, 5), use_amp=False)
-
-
-def test_spotta_config_rejects_unsupported_semantic_changes():
-    config = _spotta_config()
-    config["confidence_gate"]["aggregation"] = "mean"
-
-    try:
-        SpoTTAConfig.from_mapping(config)
-    except ValueError as exc:
-        assert "min_over_predicted_action_frames" in str(exc)
-    else:
-        raise AssertionError("Expected a changed confidence aggregation to fail.")
-
-
-def test_spotta_config_contains_only_supported_options():
-    with patch(
-        "opensportslib.core.config.loader._dali_available", return_value=False
-    ):
-        config = load_config(
-            "opensportslib/configs/localization/e2e_spotta.yaml",
-            as_namespace=False,
-        )
-    spotta_config = config["MODEL"]["policies"]["test_time_adaptation"]
-
-    assert spotta_config["enabled"] is True
-    assert spotta_config["prediction_timing"] == "adapt_then_predict"
-    assert spotta_config["confidence_gate"]["threshold"] == 0.3
-    assert spotta_config["memory"]["capacity"] == 8
-    assert spotta_config["memory"]["update_frequency"] == 2
-    assert spotta_config["robust_bn"]["tether"]["mode"] == "bayesian"
-    assert "class_policy" not in spotta_config["memory"]
-    assert "frame_filter" not in spotta_config
-    assert "steps" not in spotta_config
-    assert "reset_frequency" not in spotta_config
-    assert "action_frame_weight" not in spotta_config
-
-
-def test_localization_starts_fresh_spotta_session_and_forces_opencv_runtime():
-    from opensportslib.apis.localization import LocalizationModel
-
-    configured = []
-
-    class _Model:
-        def configure_test_time_adaptation(self, adaptation):
-            configured.append(adaptation)
-
-    adaptation = SimpleNamespace(enabled=True, name="spotta")
-    api = LocalizationModel.__new__(LocalizationModel)
-    api.model = _Model()
-    api.config = SimpleNamespace(
-        MODEL=SimpleNamespace(
-            metadata=SimpleNamespace(family="E2E"),
-            policies=SimpleNamespace(test_time_adaptation=adaptation),
-        ),
-        DATA=SimpleNamespace(
-            common=SimpleNamespace(
-                runtime=SimpleNamespace(loader_backend="dali"),
-                splits=SimpleNamespace(
-                    test=SimpleNamespace(type="VideoGameWithOpencvVideo")
-                ),
-            )
-        ),
-    )
-
-    api._configure_test_time_adaptation()
-
-    assert configured == [adaptation]
-    assert get_loader_backend(api.config) == "opencv"