[ar-api] Support Action Recognition in the SDK

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to this project will be documented in this file.
 
+## 1.5.2
+
+### Added
+
+- Action recognition support:
+  - `roboflow project create --type action-recognition` creates an action
+    recognition project. `roboflow.config.TYPE_ACTION_RECOGNITION` names the type.
+  - `Version.train(model_type="cosmos3-edge")` now exports `video-coco`, the
+    format the server trains Cosmos on. `cosmos3-edge-vlm` exports `jsonl`.
+    Both previously fell back to `yolov5pytorch`.
+
 ## 1.5.1
 
 ### Added

roboflow/__init__.py

@@ -21,7 +21,7 @@
     CLIPModel = None  # type: ignore[assignment,misc]
     GazeModel = None  # type: ignore[assignment,misc]
 
-__version__ = "1.5.1"
+__version__ = "1.5.2"
 
 
 def check_key(api_key, model, notebook, num_retries=0):

roboflow/cli/handlers/project.py

@@ -19,6 +19,7 @@ class ProjectType(str, Enum):
     instance_segmentation = "instance-segmentation"
     semantic_segmentation = "semantic-segmentation"
     keypoint_detection = "keypoint-detection"
+    action_recognition = "action-recognition"
 
 
 project_app = typer.Typer(cls=SortedGroup, help="Manage projects", no_args_is_help=True)

roboflow/config.py

@@ -74,6 +74,7 @@ def get_conditional_configuration_variable(key, default):
 TYPE_SEMANTIC_SEGMENTATION = "semantic-segmentation"
 TYPE_KEYPOINT_DETECTION = "keypoint-detection"
 TYPE_TEXT_IMAGE_PAIRS = "text-image-pairs"
+TYPE_ACTION_RECOGNITION = "action-recognition"
 
 TASK_DET = "det"
 TASK_SEG = "seg"

roboflow/util/versions.py

@@ -126,6 +126,9 @@ def get_model_format(model_type: str) -> str:
         "rfdetr": "coco",
         "rf-detr": "coco",
         "deep": "png-mask-semantic",
+        # Substring match: the image VLM sibling must win before the video model's prefix.
+        "cosmos3-edge-vlm": "jsonl",
+        "cosmos3-edge": "video-coco",
     }
 
     for prefix, format in model_formats.items():

tests/util/test_versions.py

@@ -36,6 +36,8 @@ def test_get_model_format_with_various_ids(self):
             ("resnet14", "folder"),
             ("resenet38", "yolov5pytorch"),
             ("invlid-type", "yolov5pytorch"),
+            ("cosmos3-edge", "video-coco"),
+            ("cosmos3-edge-vlm", "jsonl"),
         ]
 
         for model_type, expected_format in cases: