__getitem__.py

def __getitem__(self, idx):
    img_name = self.annotations.iloc[idx, 0]
    tg_height = self.annotations.iloc[idx, 1] if self.rect_training else 640
    tg_width = self.annotations.iloc[idx, 2] if self.rect_training else 640
    label_path = os.path.join(self.root_directory, "labels", self.annot_folder, img_name[:-4] + ".txt")
    # to avoid an annoying "UserWarning: loadtxt: Empty input file"
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        labels = np.loadtxt(fname=label_path, delimiter=" ", ndmin=2)
        # removing annotations with negative values
        labels = labels[np.all(labels >= 0, axis=1), :]
        # to avoid negative values
        labels[:, 3:5] = np.floor(labels[:, 3:5] * 1000) / 1000

    img = np.array(Image.open(os.path.join(self.root_directory, self.fname, img_name)).convert("RGB"))
    if self.bboxes_format == "coco":
        labels[:, -1] -= 1  # 0-indexing the classes of coco labels (1-80 --> 0-79)
        labels = np.roll(labels, axis=1, shift=1)
        # normalized coordinates are scale invariant, hence after resizing the img we don't resize labels
        labels[:, 1:] = coco_to_yolo_tensors(labels[:, 1:], w0=img.shape[1], h0=img.shape[0])

    img = resize_image(img, (int(tg_width), int(tg_height)))

    if self.transform:
        # albumentations requires bboxes to be (x,y,w,h,class_idx)
        batch_n = idx // self.bs
        if batch_n % 2 == 0:
            self.transform[1].p = 1
        else:
            self.transform[1].p = 0

        augmentations = self.transform(image=img, bboxes=np.roll(labels, axis=1, shift=4))
        img = augmentations["image"]
        labels = np.array(augmentations["bboxes"])
        
        # loss fx requires bboxes to be (class_idx,x,y,w,h)
        if len(labels):
            labels = np.roll(labels, axis=1, shift=1)

    if self.ultralytics_loss:
        labels = torch.from_numpy(labels)
        out_bboxes = torch.zeros((labels.shape[0], 6))
        if len(labels):
            out_bboxes[..., 1:] = labels

    img = img.transpose((2, 0, 1))
    img = np.ascontiguousarray(img)

    return torch.from_numpy(img), out_bboxes if self.ultralytics_loss else labels