opensportslib 0.2.0.dev4__tar.gz → 0.2.0.dev6__tar.gz

{opensportslib-0.2.0.dev4/opensportslib.egg-info → opensportslib-0.2.0.dev6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.2.0.dev4
+Version: 0.2.0.dev6
 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
@@ -26,6 +26,7 @@ Requires-Dist: tabulate
 Requires-Dist: pytorch-lightning
 Requires-Dist: pandas
 Requires-Dist: pyarrow
+Requires-Dist: h5py
 Requires-Dist: huggingface_hub
 Requires-Dist: easydict
 Provides-Extra: test

{opensportslib-0.2.0.dev4 → opensportslib-0.2.0.dev6}/opensportslib/apis/classification.py

@@ -17,6 +17,17 @@ from opensportslib.core.config.accessors import (
 )
 from opensportslib.core.utils.config import expand
 
+
+def _is_tracking_graph_modality(modality):
+    return str(modality).lower() in {
+        "tracking",
+        "tracking_parquet",
+        "tracking_h5",
+        "player_centroids_h5",
+        "player_joints_h5",
+    }
+
+
 class ClassificationModel(BaseTaskModel):
     """Top-level task wrapper for classification."""
 
@@ -100,7 +111,7 @@ class ClassificationModel(BaseTaskModel):
         trainer.model = model
 
         modality = get_data_modality(config)
-        use_tracking_collate = modality in {"tracking", "tracking_parquet"}
+        use_tracking_collate = _is_tracking_graph_modality(modality)
         logging.info(
             "Worker setup | mode=%s | modality=%s | tracking_collate=%s",
             mode,

{opensportslib-0.2.0.dev4 → opensportslib-0.2.0.dev6}/opensportslib/apis/localization.py

@@ -14,6 +14,7 @@ from opensportslib.core.config.accessors import (
     get_split_annotation_path,
     get_split_cfg,
     set_split_annotation_path,
+    get_model_family,
 )
 from opensportslib.core.utils.config import expand
 
@@ -170,6 +171,11 @@ class LocalizationModel(BaseTaskModel):
 
         del kwargs
 
+        if str(get_model_family(self.config)).lower() == "rulebased":
+            raise NotImplementedError(
+                "RuleBased localization models are inference-only; call infer() instead of train()."
+            )
+
         train_set = self._resolve_split_path("train", train_set)
         valid_set = self._resolve_split_path("valid", valid_set)
         self._set_split_path("train", train_set)

{opensportslib-0.2.0.dev4 → opensportslib-0.2.0.dev6}/opensportslib/apis/vqa.py

@@ -51,6 +51,7 @@ class VQAModel(BaseTaskModel):
         return_queue=None,
         train_set=None,
         valid_set=None,
+        resume_from_checkpoint=None,
         use_wandb=False,
     ):
         import torch
@@ -97,14 +98,10 @@ class VQAModel(BaseTaskModel):
             train_data = build_dataset(config, train_set, None, split="train")
             valid_data = build_dataset(config, valid_set, None, split="valid")
             trainer = Trainer_VQA(config)
-            ckpt = trainer.train(
-                None,
-                train_data,
-                valid_data,
-                rank=rank,
-                world_size=world_size,
-                use_wandb=use_wandb,
-            )
+            train_kwargs = {"rank": rank, "world_size": world_size, "use_wandb": use_wandb}
+            if resume_from_checkpoint is not None:
+                train_kwargs["resume_from_checkpoint"] = resume_from_checkpoint
+            ckpt = trainer.train(None, train_data, valid_data, **train_kwargs)
             if rank == 0 and return_queue is not None:
                 return_queue.put(ckpt)
         finally:
@@ -139,37 +136,73 @@ class VQAModel(BaseTaskModel):
         train_set: str | None = None,
         valid_set: str | None = None,
         weights: str | None = None,
+        resume_from_checkpoint: str | None = None,
+        use_ddp: bool | None = None,
         use_wandb: bool = True,
         **kwargs,
     ) -> str | None:
         del kwargs
+
+        self.config = resolve_config_omega(self.config, weights=weights)
+        execution = get_train_execution(self.config)
+        backend = str(execution.get("training_backend", "placeholder")).lower()
+        vqa_backend = get_vqa_backend(self.config)
+        if vqa_backend == "qwen_xvars_infer" and backend != "qwen_xvars_lora":
+            raise ValueError(
+                "The 'qwen_xvars_infer' backend requires TRAIN.execution.training_backend='qwen_xvars_lora' for train()."
+            )
+        if vqa_backend == "qwen_vl_native_infer" and backend != "qwen_vl_native_lora":
+            raise ValueError(
+                "The 'qwen_vl_native_infer' backend requires TRAIN.execution.training_backend='qwen_vl_native_lora' for train()."
+            )
         import torch
         import torch.multiprocessing as mp
 
-        self.config = resolve_config_omega(self.config, weights=weights)
-        if get_vqa_backend(self.config) == "qwen_xvars_infer":
-            raise ValueError("The 'qwen_xvars_infer' backend is inference-only and does not support train().")
         train_set = self._resolve_split_path("train", train_set)
         valid_set = self._resolve_split_path("valid", valid_set)
-        execution = get_train_execution(self.config)
-        backend = str(execution.get("training_backend", "placeholder")).lower()
-        if backend == "xvars_videochatgpt_lora":
-            world_size = torch.cuda.device_count() or get_system_gpu_count(self.config)
+        if backend in {"xvars_videochatgpt_lora", "qwen_xvars_lora", "qwen_vl_native_lora"}:
             requested_gpus = get_system_gpu_count(self.config)
-            use_ddp = world_size > 1 and int(requested_gpus) > 1
+            available_gpus = int(torch.cuda.device_count() or 0)
+            if requested_gpus > 0:
+                world_size = min(available_gpus, int(requested_gpus)) if available_gpus > 0 else int(requested_gpus)
+            else:
+                world_size = available_gpus
+            effective_world_size = max(int(world_size), 1)
+            if use_ddp is None:
+                launch_ddp = effective_world_size > 1
+            else:
+                launch_ddp = bool(use_ddp) and effective_world_size > 1
+            if requested_gpus > 0 and available_gpus > 0 and requested_gpus != available_gpus:
+                logging.info(
+                    "VQA train GPU selection | requested=%s | available=%s | launching=%s",
+                    requested_gpus,
+                    available_gpus,
+                    effective_world_size if launch_ddp else 1,
+                )
             logging.info(
-                "VQA train launch | mode=%s | world_size=%s",
-                "ddp" if use_ddp else "single",
-                world_size if use_ddp else 1,
+                "VQA train launch | mode=%s | requested_gpus=%s | available_gpus=%s | world_size=%s",
+                "ddp" if launch_ddp else "single",
+                requested_gpus,
+                available_gpus,
+                effective_world_size if launch_ddp else 1,
             )
 
             ctx = mp.get_context("spawn")
             queue = ctx.SimpleQueue()
-            if use_ddp:
+            if launch_ddp:
                 mp.spawn(
                     VQAModel._worker_ddp,
-                    args=(world_size, self.config_path, self.config, queue, train_set, valid_set, use_wandb),
-                    nprocs=world_size,
+                    args=(
+                        effective_world_size,
+                        self.config_path,
+                        self.config,
+                        queue,
+                        train_set,
+                        valid_set,
+                        resume_from_checkpoint,
+                        use_wandb,
+                    ),
+                    nprocs=effective_world_size,
                 )
             else:
                 VQAModel._worker_ddp(
@@ -180,6 +213,7 @@ class VQAModel(BaseTaskModel):
                     return_queue=queue,
                     train_set=train_set,
                     valid_set=valid_set,
+                    resume_from_checkpoint=resume_from_checkpoint,
                     use_wandb=use_wandb,
                 )
             ckpt = queue.get()
@@ -189,7 +223,7 @@ class VQAModel(BaseTaskModel):
 
         raise ValueError(
             f"Unsupported VQA training backend '{backend}'. "
-            "Only 'xvars_videochatgpt_lora' is supported."
+            "Only 'xvars_videochatgpt_lora' and 'qwen_xvars_lora' are supported."
         )
 
     def infer(
@@ -216,8 +250,6 @@ class VQAModel(BaseTaskModel):
         self.config = resolve_config_omega(self.config, weights=weights)
         backend = get_vqa_backend(self.config)
         effective_weights = weights if weights is not None else self.last_loaded_weights
-        if backend == "qwen_xvars_infer" and effective_weights is not None:
-            raise ValueError("The 'qwen_xvars_infer' backend does not support adapter weights for infer().")
         _set_model_checkpoint_path(self.config, effective_weights)
         self.trainer = Trainer_VQA(self.config)
         if effective_weights is not None:
@@ -237,6 +269,8 @@ class VQAModel(BaseTaskModel):
                     "question": str(question).strip(),
                     "references": [],
                     "video_path": resolved_video_path,
+                    "frame_paths": [],
+                    "video_frames": [],
                     "video_spatio_temporal_features": None,
                     "prior_prediction_text": "",
                     "labels": {},

opensportslib-0.2.0.dev6/opensportslib/configs/localization/h5_header_distance.yaml

@@ -0,0 +1,93 @@
+TASK: localization
+VERSION: 2
+
+SYSTEM:
+  paths:
+    work_dir: ./outputs/header_spotting_distance
+  device: cpu
+  gpu:
+    count: 0
+
+DATA:
+  common:
+    dataset_name: h5_headers
+    data_root: ./data
+    classes:
+      - header
+    splits:
+      test:
+        type: H5OSLJsonSpotting
+        annotation_path: ./data/h5.json
+        source_path: ./data
+        dataloader:
+          batch_size: 1
+          shuffle: false
+          num_workers: 0
+          pin_memory: false
+  inputs:
+    tracking:
+      modality: player_joints_h5
+      representation: raw
+      source:
+        format: h5
+      sampling: {}
+      transform: {}
+      augmentations: {}
+      params: {}
+
+MODEL:
+  metadata:
+    family: RuleBased
+    runner:
+      type: runner_h5_header_rule
+  components:
+    rule:
+      kind: algorithm
+      source:
+        provider: opensportslib
+        registry: rule_based
+        name: h5_header_distance
+      params:
+        label: header
+        head_name: action
+        distance_threshold_m: 0.5
+        min_confidence: 0.5
+        confidence_mode: linear_inverse_distance
+        confidence_power: 1.0
+        nms_window_ms: 1000
+        nms_scope: sample
+        ball_tolerance_ms: 60
+        chunk_size: 100000
+        head_joints: [nose, neck, l_eye, r_eye, l_ear, r_ear]
+        required_input_type: player_joints_h5
+        ball_path_field: ball_path
+        timestamp_field: timestamp_utc
+        output_task: action_spotting
+        include_diagnostics: true
+        position_ms_origin: joint_h5_start
+        metadata_start_field: start_utc
+        metadata_end_field: end_utc
+        ball_coordinate_fields: [x, y, z]
+        joint_coordinate_suffixes: [x, y, z]
+        identity_fields: [player_id, jersey_number, team_id, is_home]
+        invalid_coordinate_values: [-200.0]
+        sideline_filter_enabled: true
+        pitch_half_width_m: 50.0
+        sideline_exclusion_m: 1.0
+        sideline_reference: ball_y
+        trajectory_filter_enabled: false
+        trajectory_change_mode: either_angle_or_speed
+        trajectory_pre_window_ms: 200
+        trajectory_post_window_ms: 200
+        trajectory_min_angle_deg: 25.0
+        trajectory_min_speed_delta_ratio: 0.25
+        trajectory_min_vector_norm_m: 0.05
+        trajectory_use_xy_only: false
+        confidence_output_key: confidence_score
+  topology: []
+
+TRAIN:
+  trainer:
+    type: trainer_rule_based
+  execution:
+    enabled: false

opensportslib-0.2.0.dev6/opensportslib/configs/localization/h5_header_distance_angle.yaml

@@ -0,0 +1,93 @@
+TASK: localization
+VERSION: 2
+
+SYSTEM:
+  paths:
+    work_dir: ./outputs/header_spotting_distance_angle
+  device: cpu
+  gpu:
+    count: 0
+
+DATA:
+  common:
+    dataset_name: h5_headers
+    data_root: ./data
+    classes:
+      - header
+    splits:
+      test:
+        type: H5OSLJsonSpotting
+        annotation_path: ./data/h5.json
+        source_path: ./data
+        dataloader:
+          batch_size: 1
+          shuffle: false
+          num_workers: 0
+          pin_memory: false
+  inputs:
+    tracking:
+      modality: player_joints_h5
+      representation: raw
+      source:
+        format: h5
+      sampling: {}
+      transform: {}
+      augmentations: {}
+      params: {}
+
+MODEL:
+  metadata:
+    family: RuleBased
+    runner:
+      type: runner_h5_header_rule
+  components:
+    rule:
+      kind: algorithm
+      source:
+        provider: opensportslib
+        registry: rule_based
+        name: h5_header_distance_angle
+      params:
+        label: header
+        head_name: action
+        distance_threshold_m: 0.5
+        min_confidence: 0.5
+        confidence_mode: linear_inverse_distance
+        confidence_power: 1.0
+        nms_window_ms: 1000
+        nms_scope: sample
+        ball_tolerance_ms: 60
+        chunk_size: 100000
+        head_joints: [nose, neck, l_eye, r_eye, l_ear, r_ear]
+        required_input_type: player_joints_h5
+        ball_path_field: ball_path
+        timestamp_field: timestamp_utc
+        output_task: action_spotting
+        include_diagnostics: true
+        position_ms_origin: joint_h5_start
+        metadata_start_field: start_utc
+        metadata_end_field: end_utc
+        ball_coordinate_fields: [x, y, z]
+        joint_coordinate_suffixes: [x, y, z]
+        identity_fields: [player_id, jersey_number, team_id, is_home]
+        invalid_coordinate_values: [-200.0]
+        sideline_filter_enabled: true
+        pitch_half_width_m: 50.0
+        sideline_exclusion_m: 1.0
+        sideline_reference: ball_y
+        trajectory_filter_enabled: true
+        trajectory_change_mode: angle
+        trajectory_pre_window_ms: 200
+        trajectory_post_window_ms: 200
+        trajectory_min_angle_deg: 25.0
+        trajectory_min_speed_delta_ratio: 0.25
+        trajectory_min_vector_norm_m: 0.05
+        trajectory_use_xy_only: false
+        confidence_output_key: confidence_score
+  topology: []
+
+TRAIN:
+  trainer:
+    type: trainer_rule_based
+  execution:
+    enabled: false

opensportslib-0.2.0.dev6/opensportslib/configs/localization/h5_header_distance_speed.yaml

@@ -0,0 +1,93 @@
+TASK: localization
+VERSION: 2
+
+SYSTEM:
+  paths:
+    work_dir: ./outputs/header_spotting_distance_speed
+  device: cpu
+  gpu:
+    count: 0
+
+DATA:
+  common:
+    dataset_name: h5_headers
+    data_root: ./data
+    classes:
+      - header
+    splits:
+      test:
+        type: H5OSLJsonSpotting
+        annotation_path: ./data/h5.json
+        source_path: ./data
+        dataloader:
+          batch_size: 1
+          shuffle: false
+          num_workers: 0
+          pin_memory: false
+  inputs:
+    tracking:
+      modality: player_joints_h5
+      representation: raw
+      source:
+        format: h5
+      sampling: {}
+      transform: {}
+      augmentations: {}
+      params: {}
+
+MODEL:
+  metadata:
+    family: RuleBased
+    runner:
+      type: runner_h5_header_rule
+  components:
+    rule:
+      kind: algorithm
+      source:
+        provider: opensportslib
+        registry: rule_based
+        name: h5_header_distance_speed
+      params:
+        label: header
+        head_name: action
+        distance_threshold_m: 0.5
+        min_confidence: 0.5
+        confidence_mode: linear_inverse_distance
+        confidence_power: 1.0
+        nms_window_ms: 1000
+        nms_scope: sample
+        ball_tolerance_ms: 60
+        chunk_size: 100000
+        head_joints: [nose, neck, l_eye, r_eye, l_ear, r_ear]
+        required_input_type: player_joints_h5
+        ball_path_field: ball_path
+        timestamp_field: timestamp_utc
+        output_task: action_spotting
+        include_diagnostics: true
+        position_ms_origin: joint_h5_start
+        metadata_start_field: start_utc
+        metadata_end_field: end_utc
+        ball_coordinate_fields: [x, y, z]
+        joint_coordinate_suffixes: [x, y, z]
+        identity_fields: [player_id, jersey_number, team_id, is_home]
+        invalid_coordinate_values: [-200.0]
+        sideline_filter_enabled: true
+        pitch_half_width_m: 50.0
+        sideline_exclusion_m: 1.0
+        sideline_reference: ball_y
+        trajectory_filter_enabled: true
+        trajectory_change_mode: speed
+        trajectory_pre_window_ms: 200
+        trajectory_post_window_ms: 200
+        trajectory_min_angle_deg: 25.0
+        trajectory_min_speed_delta_ratio: 0.25
+        trajectory_min_vector_norm_m: 0.05
+        trajectory_use_xy_only: false
+        confidence_output_key: confidence_score
+  topology: []
+
+TRAIN:
+  trainer:
+    type: trainer_rule_based
+  execution:
+    enabled: false

opensportslib-0.2.0.dev6/opensportslib/configs/localization/h5_header_distance_speed_angle.yaml

@@ -0,0 +1,93 @@
+TASK: localization
+VERSION: 2
+
+SYSTEM:
+  paths:
+    work_dir: ./outputs/header_spotting_distance_speed_angle
+  device: cpu
+  gpu:
+    count: 0
+
+DATA:
+  common:
+    dataset_name: h5_headers
+    data_root: /home/vorajv/dataset/silvio-h5/
+    classes:
+      - header
+    splits:
+      test:
+        type: H5OSLJsonSpotting
+        annotation_path: /home/vorajv/dataset/silvio-h5/h5_tiny.json
+        source_path: /home/vorajv/dataset/silvio-h5/
+        dataloader:
+          batch_size: 1
+          shuffle: false
+          num_workers: 0
+          pin_memory: false
+  inputs:
+    tracking:
+      modality: player_joints_h5
+      representation: raw
+      source:
+        format: h5
+      sampling: {}
+      transform: {}
+      augmentations: {}
+      params: {}
+
+MODEL:
+  metadata:
+    family: RuleBased
+    runner:
+      type: runner_h5_header_rule
+  components:
+    rule:
+      kind: algorithm
+      source:
+        provider: opensportslib
+        registry: rule_based
+        name: h5_header_distance_speed_angle
+      params:
+        label: header
+        head_name: action
+        distance_threshold_m: 0.5
+        min_confidence: 0.5
+        confidence_mode: linear_inverse_distance
+        confidence_power: 1.0
+        nms_window_ms: 1000
+        nms_scope: sample
+        ball_tolerance_ms: 60
+        chunk_size: 100000
+        head_joints: [nose, neck, l_eye, r_eye, l_ear, r_ear]
+        required_input_type: player_joints_h5
+        ball_path_field: ball_path
+        timestamp_field: timestamp_utc
+        output_task: action_spotting
+        include_diagnostics: true
+        position_ms_origin: joint_h5_start
+        metadata_start_field: start_utc
+        metadata_end_field: end_utc
+        ball_coordinate_fields: [x, y, z]
+        joint_coordinate_suffixes: [x, y, z]
+        identity_fields: [player_id, jersey_number, team_id, is_home]
+        invalid_coordinate_values: [-200.0]
+        sideline_filter_enabled: true
+        pitch_half_width_m: 50.0
+        sideline_exclusion_m: 1.0
+        sideline_reference: ball_y
+        trajectory_filter_enabled: true
+        trajectory_change_mode: both_angle_and_speed
+        trajectory_pre_window_ms: 200
+        trajectory_post_window_ms: 200
+        trajectory_min_angle_deg: 25.0
+        trajectory_min_speed_delta_ratio: 0.25
+        trajectory_min_vector_norm_m: 0.05
+        trajectory_use_xy_only: false
+        confidence_output_key: confidence_score
+  topology: []
+
+TRAIN:
+  trainer:
+    type: trainer_rule_based
+  execution:
+    enabled: false

opensportslib-0.2.0.dev6/opensportslib/configs/localization/h5_header_rule.yaml

@@ -0,0 +1,99 @@
+TASK: localization
+VERSION: 2
+
+SYSTEM:
+  paths:
+    work_dir: ./outputs/header_spotting
+  device: cpu
+  gpu:
+    count: 0
+
+DATA:
+  common:
+    dataset_name: h5_headers
+    data_root: ./data
+    classes:
+      - header
+    splits:
+      test:
+        type: H5OSLJsonSpotting
+        annotation_path: ./data/h5.json
+        source_path: ./data
+        dataloader:
+          batch_size: 1
+          shuffle: false
+          num_workers: 0
+          pin_memory: false
+  inputs:
+    tracking:
+      modality: player_joints_h5
+      representation: raw
+      source:
+        format: h5
+      sampling: {}
+      transform: {}
+      augmentations: {}
+      params: {}
+
+MODEL:
+  metadata:
+    family: RuleBased
+    runner:
+      type: runner_h5_header_rule
+  components:
+    rule:
+      kind: algorithm
+      source:
+        provider: opensportslib
+        registry: rule_based
+        # Available models:
+        # - h5_header_distance
+        # - h5_header_distance_speed
+        # - h5_header_distance_angle
+        # - h5_header_distance_speed_angle
+        name: h5_header_distance_speed_angle
+      params:
+        label: header
+        head_name: action
+        distance_threshold_m: 0.20
+        min_confidence: 0.5
+        confidence_mode: linear_inverse_distance
+        confidence_power: 1.0
+        nms_window_ms: 1000
+        nms_scope: sample
+        ball_tolerance_ms: 60
+        chunk_size: 100000
+        head_joints: [nose, neck, l_eye, r_eye, l_ear, r_ear]
+        required_input_type: player_joints_h5
+        ball_path_field: ball_path
+        timestamp_field: timestamp_utc
+        output_task: action_spotting
+        include_diagnostics: true
+        created_by: h5_header_distance_speed_angle_rule
+        position_ms_origin: joint_h5_start
+        metadata_start_field: start_utc
+        metadata_end_field: end_utc
+        ball_coordinate_fields: [x, y, z]
+        joint_coordinate_suffixes: [x, y, z]
+        identity_fields: [player_id, jersey_number, team_id, is_home]
+        invalid_coordinate_values: [-200.0]
+        sideline_filter_enabled: true
+        pitch_half_width_m: 50.0
+        sideline_exclusion_m: 1.0
+        sideline_reference: ball_y
+        trajectory_filter_enabled: true
+        trajectory_change_mode: both_angle_and_speed
+        trajectory_pre_window_ms: 200
+        trajectory_post_window_ms: 200
+        trajectory_min_angle_deg: 25.0
+        trajectory_min_speed_delta_ratio: 0.25
+        trajectory_min_vector_norm_m: 0.05
+        trajectory_use_xy_only: false
+        confidence_output_key: confidence_score
+  topology: []
+
+TRAIN:
+  trainer:
+    type: trainer_rule_based
+  execution:
+    enabled: false