AlessandroMondin

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

AlessandroMondin

if self.transform:
    batch_n = idx // self.bs
    if batch_n % 2 == 0:
        self.transform[1].p = 1
    else:
        self.transform[1].p = 0
        
    # albumentations requires bboxes to be (x,y,w,h,class_idx)
    augmentations = self.transform(image=img,
                                   bboxes=np.roll(labels, axis=1, shift=4)

AlessandroMondin

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
# img_name[:-4] to remove the .jpg or .png which are coco img formats
label_path = os.path.join(self.root_directory, "labels", self.annot_folder, img_name[:-4] + ".txt")

AlessandroMondin

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])

AlessandroMondin

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

AlessandroMondin

def adaptive_shape(self, annotations):

    name = "train" if self.train else "val"
    path = os.path.join(
        self.root_directory, "labels",
        "adaptive_ann_{}_{}_br_{}.csv".format(name, self.len_ann, int(self.batch_range))
    )

    if os.path.isfile(path):
        print(f"==> Loading cached annotations for rectangular training on {self.annot_folder}")

AlessandroMondin

try:
    self.annotations = pd.read_csv(os.path.join(root_directory, "labels", annot_file),
                                   header=None, index_col=0).sort_values(by=[0])
    self.annotations = self.annotations.head((len(self.annotations)-1))  # just removes last line
except FileNotFoundError:
    annotations = []
    for img_txt in os.listdir(os.path.join(self.root_directory, "labels", self.annot_folder)):
        img = img_txt.split(".txt")[0]
        try:
            w, h = imagesize.get(os.path.join(self.root_directory, "images", self.annot_folder, f"{img}.jpg"))

AlessandroMondin

class Training_Dataset(Dataset):

    def __init__(self,
                 root_directory=config.ROOT_DIR,
                 transform=None,
                 train=True,
                 rect_training=False,
                 default_size=640,
                 bs=64,
                 bboxes_format="coco",

AlessandroMondin

class Training_Dataset(Dataset):
    """COCO 2017 dataset constructed using the PyTorch built-in functionalities"""

    def __init__(self,
                 num_classes,
                 root_directory=config.ROOT_DIR,
                 transform=None,
                 train=True,
                 rect_training=False,
                 default_size=640,

AlessandroMondin

gist = "indentation error"