multilabel classification upload with the CLI 😎

roboflow/__init__.py

@@ -15,7 +15,7 @@
 from roboflow.models import CLIPModel, GazeModel  # noqa: F401
 from roboflow.util.general import write_line
 
-__version__ = "1.1.65"
+__version__ = "1.1.66"
 
 
 def check_key(api_key, model, notebook, num_retries=0):

roboflow/core/workspace.py

@@ -364,6 +364,8 @@ def _save_annotation(image_id, imagedesc):
             if isinstance(annotationdesc, dict):
                 if annotationdesc.get("type") == "classification_folder":
                     annotation_path = annotationdesc.get("classification_label")
+                elif annotationdesc.get("type") == "classification_multilabel":
+                    annotation_path = json.dumps(annotationdesc.get("labels", []))
                 elif annotationdesc.get("rawText"):
                     annotation_path = annotationdesc
                 elif annotationdesc.get("file"):

roboflow/util/folderparser.py

@@ -194,6 +194,13 @@ def _filterIndividualAnnotations(image, annotation, format, imgRefMap, annotatio
             return _annotation
         else:
             return None
+    elif format == "multilabel_csv":
+        rows = [r for r in parsed["rows"] if r["file_name"] == image["name"]]
+        if rows:
+            labels = rows[0]["labels"]
+            return {"type": "classification_multilabel", "labels": labels}
+        else:
+            return None
     elif format == "jsonl":
         jsonlLines = [json.dumps(line) for line in parsed if line["image"] == image["name"]]
         if jsonlLines:
@@ -218,8 +225,9 @@ def _loadAnnotations(folder, annotations):
             ann["parsed"] = _read_jsonl(f"{folder}{ann['file']}")
             ann["parsedType"] = "jsonl"
         elif extension == ".csv":
-            ann["parsedType"] = "csv"
-            ann["parsed"] = _parseAnnotationCSV(f"{folder}{ann['file']}")
+            parsed = _parseAnnotationCSV(f"{folder}{ann['file']}")
+            ann["parsed"] = parsed
+            ann["parsedType"] = parsed.get("type", "csv")
     return annotations
 
 
@@ -241,10 +249,20 @@ def _parseAnnotationCSV(filename):
     # TODO: use a proper CSV library?
     with open(filename) as f:
         lines = f.readlines()
-    headers = lines[0]
+    headers = [h.strip() for h in lines[0].split(",")]
+    # Multi-label classification csv typically named _classes.csv
+    if os.path.basename(filename) == "_classes.csv":
+        parsed_lines = []
+        for line in lines[1:]:
+            parts = [p.strip() for p in line.split(",")]
+            file_name = parts[0]
+            labels = [headers[i] for i, v in enumerate(parts[1:], start=1) if v == "1"]
+            parsed_lines.append({"file_name": file_name, "labels": labels})
+        return {"type": "multilabel_csv", "rows": parsed_lines, "headers": headers}
+    header_line = lines[0]
     lines = [{"file_name": ld.split(",")[0].strip(), "line": ld} for ld in lines[1:]]
     return {
-        "headers": headers,
+        "headers": header_line,
         "lines": lines,
     }
 

tests/datasets/skinproblem-multilabel-classification/README.dataset.txt

@@ -0,0 +1,5 @@
+# Skin-Problem-MultiLabel > 2023-12-26 4:26pm
+https://universe.roboflow.com/parin-kittipongdaja-vwmn3/skin-problem-multilabel
+
+Provided by a Roboflow user
+License: CC BY 4.0

tests/test_project.py

@@ -1,3 +1,4 @@
+import json
 from unittest.mock import patch
 
 import requests
@@ -625,3 +626,43 @@ def capture_annotation_calls(annotation_path, **kwargs):
         finally:
             for mock in mocks.values():
                 mock.stop()
+
+    def test_multilabel_classification_dataset_upload(self):
+        from roboflow.util import folderparser
+
+        multilabel_folder = "tests/datasets/skinproblem-multilabel-classification"
+        parsed_dataset = folderparser.parsefolder(multilabel_folder, is_classification=True)
+
+        self.project.type = "classification"
+        self.project.multilabel = True
+        annotation_calls = []
+
+        def capture_annotation_calls(annotation_path, **kwargs):
+            annotation_calls.append(annotation_path)
+            return ({"success": True}, 0.1, 0)
+
+        mocks = {
+            "parser": patch("roboflow.core.workspace.folderparser.parsefolder", return_value=parsed_dataset),
+            "upload": patch(
+                "roboflow.core.workspace.Project.upload_image",
+                return_value=({"id": "test-id", "success": True}, 0.1, 0),
+            ),
+            "save_annotation": patch(
+                "roboflow.core.workspace.Project.save_annotation", side_effect=capture_annotation_calls
+            ),
+            "get_project": patch(
+                "roboflow.core.workspace.Workspace._get_or_create_project", return_value=(self.project, False)
+            ),
+        }
+        for mock in mocks.values():
+            mock.start()
+        try:
+            self.workspace.upload_dataset(dataset_path=multilabel_folder, project_name=PROJECT_NAME, num_workers=1)
+            self.assertEqual(len(annotation_calls), len(parsed_dataset["images"]))
+            for call in annotation_calls:
+                labels = json.loads(call)
+                self.assertIsInstance(labels, list)
+                self.assertGreater(len(labels), 0)
+        finally:
+            for mock in mocks.values():
+                mock.stop()

tests/util/test_folderparser.py

@@ -86,6 +86,15 @@ def test_parse_classification_folder_structure(self):
             self.assertEqual(img["annotationfile"]["type"], "classification_folder")
             self.assertEqual(img["annotationfile"]["classification_label"], "no-corrosion")
 
+    def test_parse_multilabel_classification_csv(self):
+        folder = f"{thisdir}/../datasets/skinproblem-multilabel-classification"
+        parsed = folderparser.parsefolder(folder, is_classification=True)
+        images = {img["name"]: img for img in parsed["images"]}
+        img1 = images.get("101_jpg.rf.ffb91e580c891eb04b715545274b2469.jpg")
+        self.assertIsNotNone(img1)
+        self.assertEqual(img1["annotationfile"]["type"], "classification_multilabel")
+        self.assertEqual(set(img1["annotationfile"]["labels"]), {"Blackheads"})
+
 
 def _assertJsonMatchesFile(actual, filename):
     with open(filename) as file: