adaptive_shape.py

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}")
        parsed_annot = pd.read_csv(path, index_col=0)
    else:
        print("...Running adaptive_shape for 'rectangular training' on training set...")
        annotations["w_h_ratio"] = annotations.iloc[:, 2] / annotations.iloc[:, 1]
        annotations.sort_values(["w_h_ratio"], ascending=True, inplace=True)

        for i in range(0, len(annotations), self.batch_range):
            size = [annotations.iloc[i, 2], annotations.iloc[i, 1]]  # [width, height]
            max_dim = max(size)
            max_idx = size.index(max_dim)
            size[~max_idx] += 32
            sz = random.randrange(int(self.default_size * 0.9), int(self.default_size * 1.1)) // 32 * 32
            size[~max_idx] = ((sz/size[max_idx])*(size[~max_idx]) // 32) * 32
            size[max_idx] = sz
            if i + self.batch_range <= len(annotations):
                bs = self.batch_range
            else:
                bs = len(annotations) - i

            annotations.iloc[i:bs, 2] = size[0]
            annotations.iloc[i:bs, 1] = size[1]

            # sample annotation to avoid having pseudo-equal images in the same batch
            annotations.iloc[i:i+bs, :] = annotations.iloc[i:i+bs, :].sample(frac=1, axis=0)

        parsed_annot = pd.DataFrame(annotations.iloc[:,:3])
        parsed_annot.to_csv(path)

    return parsed_annot