what training input formats are supported?

[SV1367]

Hello. Very interesting. Do you support other training input formats apart from Coco? for example the text-based format used in other YOLO variants (e.g. yolov5)?

[LSH9832]

Not yet, but YOLO format is similar to Visdrone formats(both load labels from txt), and its dataloader locates in path edgeyolo/data/datasets/visdrone.py, I think just modify the following function and then you can use datasets with yolo format. You can have a try.

class VisDroneDataset(Dataset):

    def get_anno_by_name(self, name):
        msg = {
            "image": os.path.join(self.train_dir, f"{os.path.basename(name)[:-4]}.{self.suffix}"),
            "annotations": []
        }
        num_labels = 0
        with open(name, "r", encoding="utf8") as f:
            has_anno = False
            for line in f.read().split("\n"):

                line = line.replace(",", " ").split()
                if len(line) == 8:
                    x1, y1, w, h = [float(k) for k in line[:4]]
                    x2, y2 = x1 + w, y1 + h
                    score = int(line[4])
                    if score == 0:
                        continue

                    class_id = int(line[5])
                    if class_id == 0 or class_id > len(self.names):
                        # id=0: ignore regions  id=-1: others
                        continue
                    class_id -= 1
                    num_labels += 1
                    msg["annotations"].append([x1, y1, x2, y2, class_id])   # xywh
                    if not self.is_train:
                        if not has_anno:
                            self.coco_data.add_image(
                                image_id=self.idx,
                                file_name=os.path.basename(msg["image"]),
                                width=4000,
                                height=4000,
                            )
                            has_anno = True

                        self.coco_data.add_annotation(
                            image_id=self.idx,
                            anno_id=self.num_annos,
                            category_id=class_id,
                            bbox=[x1, y1, w, h],
                            iscrowd=0
                        )
                    self.num_annos += 1
            self.max_num_labels = max(self.max_num_labels, num_labels)
            msg["annotations"] = np.array(msg["annotations"])
            return msg