opensportslib 0.0.1.dev17__tar.gz → 0.0.1.dev19__tar.gz

{opensportslib-0.0.1.dev17/opensportslib.egg-info → opensportslib-0.0.1.dev19}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.0.1.dev17
+Version: 0.0.1.dev19
 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-0.0.1.dev19/opensportslib/config/localization-json_calf_resnetpca512.yaml

@@ -0,0 +1,169 @@
+TASK: localization
+
+dali: false
+
+DATA:
+  dataset_name: SoccerNet
+  data_dir: /home/vorajv/opensportslib/SoccerNet/
+  classes:
+    - Penalty
+    - Kick-off
+    - Goal
+    - Substitution
+    - Offside
+    - Shots on target
+    - Shots off target
+    - Clearance
+    - Ball out of play
+    - Throw-in
+    - Foul
+    - Indirect free-kick
+    - Direct free-kick
+    - Corner
+    - Yellow card
+    - Red card
+    - Yellow->red card
+    
+  epoch_num_frames: 500000
+  mixup: true
+  modality: rgb
+  crop_dim: -1
+  dilate_len: 0        # Dilate ground truth labels
+  clip_len: 100
+  input_fps: 25
+  extract_fps: 2
+  imagenet_mean: [0.485, 0.456, 0.406]
+  imagenet_std: [0.229, 0.224, 0.225]
+  target_height: 224
+  target_width: 398
+
+  train:
+    type: FeatureClipChunksfromJson
+    classes: ${DATA.classes}
+    output_map: [data, label]
+    video_path: ${DATA.data_dir}
+    path: ${DATA.train.video_path}/annotations-2024-224p-train.json
+    framerate: 2
+    chunk_size: 120
+    receptive_field: 40
+    chunks_per_epoch: 6000
+    dataloader:
+      batch_size: 256
+      shuffle: true
+      num_workers: 4
+      pin_memory: true
+
+  valid:
+    type: FeatureClipChunksfromJson
+    classes: ${DATA.classes}
+    output_map: [data, label]
+    video_path: ${DATA.data_dir}
+    path: ${DATA.valid.video_path}/annotations-2024-224p-valid.json
+    framerate: 2
+    chunk_size: 120
+    receptive_field: 40
+    chunks_per_epoch: 6000
+    dataloader:
+      batch_size: 256
+      shuffle: true
+      num_workers: 4
+      pin_memory: true
+
+  test:
+    type: FeatureVideosChunksfromJson
+    classes: ${DATA.classes}
+    output_map: [data, label]
+    video_path: ${DATA.data_dir}
+    path: ${DATA.test.video_path}/annotations-2024-224p-test.json
+    results: results_spotting_test_calf_resnetpca512
+    framerate: 2
+    chunk_size: 120
+    receptive_field: 40
+    chunks_per_epoch: 6000
+    metric: loose
+    dataloader:
+      batch_size: 1
+      shuffle: false
+      num_workers: 1
+      pin_memory: true
+
+MODEL:
+    type: ContextAware
+    runner:
+      type: runner_JSON
+    backbone:
+      type: PreExtactedFeatures
+      encoder: ResNET_TF2_PCA512
+      feature_dim: 512
+      output_dim: 512
+      framerate: 2
+    neck:
+      type: CNN++
+      input_size: 512
+      num_classes: 17
+      chunk_size: 120
+      dim_capsule: 16
+      receptive_field: 40
+      num_detections: 15
+      framerate: 2
+    head:
+      type: SpottingCALF
+      dim_capsule: 16
+      num_detections: 15
+      chunk_size: 120
+      num_classes: 17
+    load_weights: null
+
+TRAIN:
+  type: trainer_CALF
+  max_epochs: 200
+  evaluation_frequency: 200
+  framerate: 2
+  batch_size: 32
+  monitor: valid_loss
+  mode: min
+
+  criterion:
+    type: Combined2x
+    w_1: 0.000367
+    loss_1:
+      type: ContextAwareLoss
+      K: [[-100, -98, -20, -40, -96, -5, -8, -93, -99, -31, -75, -10, -97, -75, -20, -84, -18], 
+            [-50, -49, -10, -20, -48, -3, -4, -46, -50, -15, -37, -5, -49, -38, -10, -42, -9], 
+            [50, 49, 60, 10, 48, 3, 4, 46, 50, 15, 37, 5, 49, 38, 10, 42, 9], 
+            [100, 98, 90, 20, 96, 5, 8, 93, 99, 31, 75, 10, 97, 75, 20, 84, 18]]
+      framerate: 2
+      hit_radius: 0.1
+      miss_radius: 0.9
+    w_2: 1.0
+    loss_2:
+      type: SpottingLoss
+      lambda_coord: 5.0
+      lambda_noobj: 0.5
+
+  optimizer:
+    type: Adam
+    lr: 0.001
+    betas: [0.9, 0.999]
+    eps: 1e-07
+    weight_decay: 0
+    amsgrad: false
+
+  scheduler:
+    type: ReduceLROnPlateau
+    mode: min
+    factor: 1e-03
+    min_lr: 1e-06
+    patience: 25
+    verbose: true
+
+
+SYSTEM:
+  log_dir: ./logs
+  save_dir: ./checkpoints
+  work_dir: ${SYSTEM.save_dir}
+  seed: 42
+  GPU: 4        # number of gpus to use
+  device: cuda   # auto | cuda | cpu
+  gpu_id: 0      # device id for single gpu training
+

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/config/localization-json_netvlad++_resnetpca512.yaml

@@ -74,7 +74,7 @@ DATA:
     results: results_spotting_test_netvlad++_resnetpca512
     framerate: 2
     window_size: 20
-    metric: tight
+    metric: loose
     dataloader:
       batch_size: 1
       shuffle: false
@@ -109,10 +109,12 @@ MODEL:
 
 TRAIN:
   type: trainer_pooling
-  max_epochs: 1000
-  evaluation_frequency: 1000
+  max_epochs: 200
+  evaluation_frequency: 200
   framerate: 2
   batch_size: 256
+  monitor: valid_loss
+  mode: min
 
   criterion:
     type: NLLLoss

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/core/loss/calf.py

@@ -26,7 +26,7 @@ class ContextAwareLoss(torch.nn.Module):
 
         super(ContextAwareLoss, self).__init__()
 
-        self.K = torch.FloatTensor(K * framerate).cuda()
+        self.K = torch.FloatTensor(K * framerate)
         self.hit_radius = float(hit_radius)
         self.miss_radius = float(miss_radius)
 
@@ -40,7 +40,7 @@ class ContextAwareLoss(torch.nn.Module):
         Returns:
             torch.Tensor: The returned loss.
         """
-        K = self.K
+        K = self.K.to(output.device)
         hit_radius = self.hit_radius
         miss_radius = self.miss_radius
 

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/core/trainer/localization_trainer.py

@@ -42,6 +42,7 @@ import tqdm
 import numpy as np
 from opensportslib.core.utils.config import load_gz_json, load_json
 from abc import ABC, abstractmethod
+
 import logging
 logger = logging.getLogger(__name__)
 
@@ -159,14 +160,29 @@ class Trainer_pl(Trainer):
 
     def __init__(self, cfg, work_dir):
         from opensportslib.core.utils.lightning import CustomProgressBar, MyCallback
+        from pytorch_lightning.callbacks import ModelCheckpoint
         import pytorch_lightning as pl
+        from pytorch_lightning.loggers import WandbLogger
+
+        wandb_logger = None
+        if wandb.run is not None:  # means init_wandb already ran
+            wandb_logger = WandbLogger(experiment=wandb.run)
 
         self.work_dir = work_dir
         call = MyCallback()
+        checkpoint_callback = ModelCheckpoint(
+            dirpath=self.work_dir,
+            filename="best",
+            monitor=getattr(cfg.TRAIN, 'monitor', 'valid_loss'),   # or your metric
+            mode=getattr(cfg.TRAIN, 'mode', 'min'),
+            save_top_k=1,
+        )
         self.trainer = pl.Trainer(
-            max_epochs=cfg.TRAIN.max_epochs,
-            devices=cfg.SYSTEM.GPU,
-            callbacks=[call, CustomProgressBar(refresh_rate=1)],
+            logger=wandb_logger,
+            max_epochs=getattr(cfg.TRAIN, 'max_epochs', 200),
+            devices="auto",
+            accelerator="auto",
+            callbacks=[call, CustomProgressBar(refresh_rate=1), checkpoint_callback],
             num_sanity_val_steps=0,
         )
         self.best_checkpoint_path = None
@@ -174,16 +190,27 @@ class Trainer_pl(Trainer):
     def train(self, **kwargs):
         self.trainer.fit(**kwargs)
 
-        best_model = kwargs["model"].best_state
+        # best_model = kwargs["model"].best_state
+
+        # logging.info("Done training")
+        # logging.info("Best epoch: {}".format(best_model.get("epoch")))
+        # best_path = os.path.join(self.work_dir, "model.pth.tar")
+        # self.best_checkpoint_path = best_path
+        # torch.save(best_model, best_path)
+
+        # logging.info("Model saved")
+        # logging.info(best_path)
 
-        logging.info("Done training")
-        logging.info("Best epoch: {}".format(best_model.get("epoch")))
-        best_path = os.path.join(self.work_dir, "model.pth.tar")
-        self.best_checkpoint_path = best_path
-        torch.save(best_model, best_path)
+        self.best_checkpoint_path = self.trainer.checkpoint_callback.best_model_path
 
-        logging.info("Model saved")
-        logging.info(best_path)
+        from pytorch_lightning.utilities.rank_zero import rank_zero_only
+
+        @rank_zero_only
+        def log():
+            logging.info("Done training")
+            logging.info(f"Best model saved at: {self.best_checkpoint_path}")
+
+        log()
 
 
 class Trainer_e2e(Trainer):
@@ -531,12 +558,19 @@ class Inferer:
         # Run Inference on Dataset
         from opensportslib.core.utils.lightning import CustomProgressBar, MyCallback
         import pytorch_lightning as pl
+        from pytorch_lightning.loggers import WandbLogger
+
+        wandb_logger = None
+        if wandb.run is not None:  # means init_wandb already ran
+            wandb_logger = WandbLogger(experiment=wandb.run)
 
         if cfg.SYSTEM.work_dir is not None and dataloader is not None:
             
             evaluator = pl.Trainer(
+                logger=wandb_logger,
                 callbacks=[CustomProgressBar()],
-                devices=cfg.SYSTEM.GPU,
+                devices=1,
+                accelerator="auto",
                 num_sanity_val_steps=0,
             )
             evaluator.predict(model, dataloader)
@@ -770,7 +804,6 @@ class Evaluator:
 
         # detect v2 prediction
         pred_is_v2 = isinstance(pred_data, dict) and pred_data is not None and "data" in pred_data
-        print("PRED V2 :", pred_is_v2)
         # --------------------------------------------------
         # CLASSES
         # --------------------------------------------------
@@ -815,7 +848,6 @@ class Evaluator:
 
             # ---------------- GT ----------------
             if gt_is_v2:
-                print("Game: ", game)
                 video_path = game["inputs"][0]["path"]
                 labels = [{"label": e.get("label"),  
                            "gameTime": e.get("gameTime"),
@@ -1048,10 +1080,13 @@ class Evaluator:
                 "{:0.2f}".format(results["a_mAP_unshown"] * 100),
             )
         )
-
+        header = ["", "Any", "Visible", "Unseen"]
+        result = tabulate(rows, headers=header)
+        from opensportslib.core.utils.wandb import log_table_wandb
+        log_table_wandb(name=f"Final Scores (Metric : {metric})", rows=rows, headers=header)
         logging.info("Best Performance at end of training ")
         logging.info("Metric: " + metric)
-        print(tabulate(rows, headers=["", "Any", "Visible", "Unseen"]))
+        logging.info("\n" + result)
         # logging.info("a_mAP visibility all: " +  str(results["a_mAP"]))
         # logging.info("a_mAP visibility all per class: " +  str( results["a_mAP_per_class"]))
 

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/core/utils/checkpoint.py

@@ -184,16 +184,25 @@ def load_checkpoint(
     model_keys = list(model.state_dict().keys())
     ckpt_keys  = list(state_dict.keys())
 
-    ckpt_has_module  = ckpt_keys[0].startswith("module.")
-    model_has_module = model_keys[0].startswith("module.")
+    # ckpt_has_module  = ckpt_keys[0].startswith("module.")
+    # model_has_module = model_keys[0].startswith("module.")
 
-    # Case 1: checkpoint has module., model doesn't
-    if ckpt_has_module and not model_has_module:
-        state_dict = {k.replace("module.", "", 1): v for k, v in state_dict.items()}
+    # # Case 1: checkpoint has module., model doesn't
+    # if ckpt_has_module and not model_has_module:
+    #     state_dict = {k.replace("module.", "", 1): v for k, v in state_dict.items()}
 
-    # Case 2: checkpoint doesn't have module., model does
-    elif not ckpt_has_module and model_has_module:
-        state_dict = {f"module.{k}": v for k, v in state_dict.items()}
+    # # Case 2: checkpoint doesn't have module., model does
+    # elif not ckpt_has_module and model_has_module:
+    #     state_dict = {f"module.{k}": v for k, v in state_dict.items()}
+
+    def strip_prefix(state_dict, prefix):
+        return {
+            k[len(prefix):] if k.startswith(prefix) else k: v
+            for k, v in state_dict.items()
+        }
+
+    state_dict = strip_prefix(state_dict, "module.")
+    state_dict = strip_prefix(state_dict, "model.")
 
     # Optional custom remap
     if key_remap_fn:

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/metrics/localization_metric.py

@@ -880,9 +880,9 @@ def label2vector(
 
         else:
             time = annotation["gameTime"]
-            print(time)
-            minutes = int(time[-5:-3])
-            seconds = int(time[-2::])
+            if time is not None:
+                minutes = int(time[-5:-3])
+                seconds = int(time[-2::])
             # annotation at millisecond precision
             if "position" in annotation:
                 frame = int(framerate * (int(annotation["position"]) / 1000))
@@ -931,7 +931,7 @@ def predictions2vector(
 
         event = annotation["label"]
 
-        if "frame" in annotation:
+        if "frame" in annotation and annotation["frame"] is not None:
             frame = int(annotation["frame"])
         else:
             time = int(annotation["position"])

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/models/base/contextaware.py

@@ -249,10 +249,10 @@ class LiteContextAwareModel(LiteBaseModel):
                     self.json_data = json_data
             elif self.runner == "runner_JSON":
                 list_videos = self.trainer.predict_dataloaders.dataset.data_json[0][
-                    "videos"
+                    "data"
                 ]
                 for index in np.arange(len(list_videos)):
-                    video = list_videos[index]["path"]
+                    video = list_videos[index]["inputs"][0]["path"]
 
                     if self.infer_split:
                         video = os.path.splitext(video)[0]

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib/models/utils/utils.py

@@ -275,7 +275,7 @@ def predictions2json_runnerjson(
 
         confidence = predictions_video[frame_index, class_index]
 
-        json_data["predictions"].append(
+        json_data["data"][0]["events"].append(
             get_prediction_data(
                 True,
                 frame_index,

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19/opensportslib.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.0.1.dev17
+Version: 0.0.1.dev19
 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-0.0.1.dev17 → opensportslib-0.0.1.dev19}/opensportslib.egg-info/SOURCES.txt

@@ -16,6 +16,7 @@ opensportslib/apis/classification.py
 opensportslib/apis/localization.py
 opensportslib/config/classification.yaml
 opensportslib/config/localization-e2e-ocv.yaml
+opensportslib/config/localization-json_calf_resnetpca512.yaml
 opensportslib/config/localization-json_netvlad++_resnetpca512.yaml
 opensportslib/config/localization.yaml
 opensportslib/config/sngar-frames.yaml

{opensportslib-0.0.1.dev17 → opensportslib-0.0.1.dev19}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opensportslib"
-version = "0.0.1.dev17"
+version = "0.0.1.dev19"
 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"