dry run

kprokofi · kprokofi · commit 440f815629ec · 2025-11-27T08:00:22.000+09:00
diff --git a/library/src/otx/backend/native/callbacks/ema.py b/library/src/otx/backend/native/callbacks/ema.py
@@ -371,7 +371,7 @@ def __init__(
         self,
         device: Optional[Union[torch.device, str, int]] = None,
         use_buffers: bool = True,
-        decay: float = 0.999,
+        decay: float = 0.9999,
         update_every_n_steps: int = 1,
         update_starting_at_step: Optional[int] = None,
         update_starting_at_epoch: Optional[int] = None,
diff --git a/library/src/otx/backend/native/models/base.py b/library/src/otx/backend/native/models/base.py
@@ -167,6 +167,7 @@ def __init__(
 
         self._label_info = self._dispatch_label_info(label_info)
         self.model_name = model_name
+        self.log_all_losses = False
         if isinstance(data_input_params, dict):
             data_input_params = DataInputParams(**data_input_params)
         elif data_input_params is None:
@@ -212,14 +213,15 @@ def training_step(self, batch: OTXDataBatch, batch_idx: int) -> Tensor:
             )
             return train_loss
         if isinstance(train_loss, dict):
-            for k, v in train_loss.items():
-                self.log(
-                    f"train/{k}",
-                    v,
-                    on_step=True,
-                    on_epoch=False,
-                    prog_bar=True,
-                )
+            if self.log_all_losses:
+                for k, v in train_loss.items():
+                    self.log(
+                        f"train/{k}",
+                        v,
+                        on_step=True,
+                        on_epoch=False,
+                        prog_bar=True,
+                    )
 
             total_train_loss = train_loss.get("total_loss", sum(train_loss.values()))
             self.log(
diff --git a/library/src/otx/backend/native/models/detection/deimv2.py b/library/src/otx/backend/native/models/detection/deimv2.py
@@ -106,6 +106,7 @@ def _create_model(self, num_classes: int | None = None) -> DETR:
             num_classes=num_classes,
             eval_spatial_size=self.data_input_params.input_size,
         )
+
         criterion = DEIMCriterion(
             weight_dict={
                 "loss_vfl": 1,
diff --git a/library/src/otx/data/transform_libs/torchvision.py b/library/src/otx/data/transform_libs/torchvision.py
@@ -18,6 +18,7 @@
 import numpy as np
 import PIL.Image
 import torch
+import uuid
 import torchvision.transforms.v2 as tvt_v2
 import typeguard
 from lightning.pytorch.cli import instantiate_class
@@ -2124,8 +2125,56 @@ def forward(self, *_inputs: OTXDataItem) -> OTXDataItem | None:
 
                 inputs.polygons = [mixup_gt_polygons[i] for i in np.where(inside_inds)[0]]
 
+        # self.visualize(inputs, output_path=f"/home/kprokofi/debug_images/{str(uuid.uuid4())}.jpg")
         return self.convert(inputs)
 
+    def visualize(
+        self,
+        inputs: OTXDataItem,
+        output_path: str | None = None,
+        show_blended: bool = True,
+    ) -> np.ndarray:
+        """Visualize CopyBlend augmentation for debugging.
+
+        Args:
+            inputs: OTXDataItem to visualize.
+            output_path: Optional path to save visualization.
+            show_blended: Whether to show blended boxes in different color.
+
+        Returns:
+            Visualization as numpy array.
+        """
+        import cv2
+
+        img = to_np_image(inputs.image).copy()
+        bboxes = inputs.bboxes
+        labels = inputs.label
+
+        # Draw bboxes
+        for idx, bbox in enumerate(bboxes):
+            x1, y1, x2, y2 = bbox.int().tolist()
+            label = labels[idx].item() if hasattr(labels[idx], "item") else labels[idx]
+
+            # Use different colors for original vs blended
+            # Assume last N boxes are blended (where N = num_objects)
+            color = (0, 255, 0)
+
+            cv2.rectangle(img, (x1, y1), (x2, y2), color, 2)
+            cv2.putText(
+                img,
+                f"{label}",
+                (x1, y1 - 5),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.5,
+                color,
+                2,
+            )
+
+        if output_path:
+            cv2.imwrite(output_path, img)
+
+        return img
+
     def __repr__(self):
         repr_str = self.__class__.__name__
         repr_str += f"(dynamic_scale={self.dynamic_scale}, "
@@ -2403,7 +2452,6 @@ def forward(self, *_inputs: OTXDataItem) -> OTXDataItem | None:
                 canvas_size=(img_h, img_w),
             )
             inputs.label = combined_labels
-
         return self.convert(inputs)
 
     def visualize(
diff --git a/library/src/otx/recipe/detection/deimv2_l.yaml b/library/src/otx/recipe/detection/deimv2_l.yaml
@@ -35,6 +35,10 @@ callbacks:
     init_args:
       max_interval: 1
       min_lrschedule_patience: 3
+  # - class_path: otx.backend.native.callbacks.EMAWeightAveraging
+  #   init_args:
+  #     update_starting_at_epoch: 10
+  #     decay: 0.999
   - class_path: otx.backend.native.callbacks.adaptive_early_stopping.EarlyStoppingWithWarmup
     init_args:
       mode: max
@@ -117,7 +121,7 @@ callbacks:
               transforms:
                 - class_path: otx.data.transform_libs.torchvision.CachedMosaic
                   init_args:
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 20
                     img_scale: [640, 640]
                 - class_path: otx.data.transform_libs.torchvision.CachedMixUp
@@ -127,7 +131,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
@@ -160,6 +164,7 @@ callbacks:
                   init_args:
                     fill: 0
                     p: 0.5
+                    side_range: [1.0, 2.0]
                 - class_path: otx.data.transform_libs.torchvision.RandomIoUCrop # Can't be used when using CachedMosaic
                   init_args:
                     probability: 0.8
@@ -176,7 +181,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
diff --git a/library/src/otx/recipe/detection/deimv2_m.yaml b/library/src/otx/recipe/detection/deimv2_m.yaml
@@ -117,7 +117,7 @@ callbacks:
               transforms:
                 - class_path: otx.data.transform_libs.torchvision.CachedMosaic
                   init_args:
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 20
                     img_scale: [640, 640]
                 - class_path: otx.data.transform_libs.torchvision.CachedMixUp
@@ -127,7 +127,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
@@ -176,7 +176,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
diff --git a/library/src/otx/recipe/detection/deimv2_s.yaml b/library/src/otx/recipe/detection/deimv2_s.yaml
@@ -117,7 +117,7 @@ callbacks:
               transforms:
                 - class_path: otx.data.transform_libs.torchvision.CachedMosaic
                   init_args:
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 20
                     img_scale: [640, 640]
                 - class_path: otx.data.transform_libs.torchvision.CachedMixUp
@@ -127,7 +127,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
@@ -176,7 +176,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
diff --git a/library/src/otx/recipe/detection/deimv2_x.yaml b/library/src/otx/recipe/detection/deimv2_x.yaml
@@ -117,7 +117,7 @@ callbacks:
               transforms:
                 - class_path: otx.data.transform_libs.torchvision.CachedMosaic
                   init_args:
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 20
                     img_scale: [640, 640]
                 - class_path: otx.data.transform_libs.torchvision.CachedMixUp
@@ -127,7 +127,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
@@ -176,7 +176,7 @@ callbacks:
                       - 1.0
                       - 1.0
                     probability: 0.5
-                    random_pop: false
+                    random_pop: true
                     max_cached_images: 10
                 - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
                   init_args:
diff --git a/library/tests/perf_v2/benchmark.py b/library/tests/perf_v2/benchmark.py
@@ -97,7 +97,6 @@ def __init__(
         num_epoch: int = 0,
         eval_upto: str = "train",
         tags: dict[str, str] | None = None,
-        dry_run: bool = False,
         deterministic: bool = False,
         accelerator: str = "gpu",
         reference_results: pd.DataFrame | None = None,
@@ -109,7 +108,6 @@ def __init__(
         self.num_epoch = num_epoch
         self.eval_upto = eval_upto
         self.tags = tags or {}
-        self.dry_run = dry_run
         self.deterministic = deterministic
         self.accelerator = accelerator
         self.reference_results = reference_results
@@ -641,7 +639,6 @@ def check(self, result: pd.DataFrame, criteria: list[Criterion]):
         num_epoch=args.num_epoch,
         eval_upto=args.eval_upto,
         tags=tags,
-        dry_run=args.dry_run,
         deterministic=(
             False if args.deterministic is None else {"true": True, "false": False, "warn": "warn"}[args.deterministic]
         ),
diff --git a/library/tests/perf_v2/run.py b/library/tests/perf_v2/run.py
@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+from email.policy import default
 import json
 import logging
 import subprocess
@@ -99,7 +100,6 @@ def load_failed_jobs(file_path: Path) -> list[list[str]]:
                     if (output_root / model.name / dataset.name / str(seed)).exists():
                         logger.info(f"Skipping existing job for {model.name} on {dataset.name} with seed {seed}")
                         continue
-
                     cmd = [
                         "python",
                         "-m",
@@ -116,8 +116,12 @@ def load_failed_jobs(file_path: Path) -> list[list[str]]:
                         str(output_root),
                         "--seed",
                         str(seed),
+                        "--eval-upto",
+                        str(args.eval_upto),
                         "--num-epoch",
                         str(args.num_epoch),
+                        "--deterministic",
+                        str(args.deterministic),
                         "--device",
                         args.device,
                         "--user-name",
diff --git a/library/tests/perf_v2/tasks/detection.py b/library/tests/perf_v2/tasks/detection.py
@@ -61,11 +61,11 @@
         #     path=Path("detection/skindetect-roboflow"),
         #     group="small",
         # ),
-        DatasetInfo(
-            name="diopsis",
-            path=Path("detection/diopsis_coco"),
-            group="medium",
-        ),
+        # DatasetInfo(
+        #     name="diopsis",
+        #     path=Path("detection/diopsis_coco"),
+        #     group="medium",
+        # ),
         DatasetInfo(
             name="bdd_medium",
             path=Path("detection/bdd_medium"),
@@ -76,11 +76,11 @@
         #     path=Path("detection/Vitens-Aeromonas-coco"),
         #     group="medium",
         # ),
-        DatasetInfo(
-            name="visdrone",
-            path=Path("detection/visdrone_coco_custom_split"),
-            group="large",
-        ),
+        # DatasetInfo(
+        #     name="visdrone",
+        #     path=Path("detection/visdrone_coco_custom_split"),
+        #     group="large",
+        # ),
     ]
 )
 

Original file line number	Diff line number	Diff line change
`@@ -106,6 +106,7 @@ def _create_model(self, num_classes: int \| None = None) -> DETR:`
`106`	`106`	`num_classes=num_classes,`
`107`	`107`	`eval_spatial_size=self.data_input_params.input_size,`
`108`	`108`	`)`
	`109`	`+`
`109`	`110`	`criterion = DEIMCriterion(`
`110`	`111`	`weight_dict={`
`111`	`112`	`"loss_vfl": 1,`