cars 1.0.0rc1__cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl

cars/data_structures/cars_dataset.py

@@ -0,0 +1,1541 @@
+#!/usr/bin/env python
+# coding: utf8
+#
+# Copyright (c) 2020 Centre National d'Etudes Spatiales (CNES).
+#
+# This file is part of CARS
+# (see https://github.com/CNES/cars).
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# pylint: disable=too-many-lines
+"""
+cars_dataset module:
+
+"""
+
+
+import copy
+import logging
+import math
+
+# Standard imports
+import os
+import pickle
+from typing import Dict
+
+# Third party imports
+import numpy as np
+import pandas
+import pyproj
+import rasterio as rio
+import xarray as xr
+import yaml
+from rasterio.profiles import DefaultGTiffProfile
+from rasterio.windows import Window
+
+# CARS imports
+from cars.core import constants as cst
+from cars.core import outputs
+from cars.core.utils import safe_makedirs
+from cars.data_structures import cars_dict, dataframe_converter
+
+# cars dataset dtype
+CARS_DS_TYPE_ARRAY = "arrays"
+CARS_DS_TYPE_POINTS = "points"
+CARS_DS_TYPE_DICT = "dict"
+
+# cars_dataset names
+TILES_INFO_FILE = "tiles_info.yaml"
+OVERLAP_FILE = "overlaps.npy"
+GRID_FILE = "grid.npy"
+PROFILE_FILE = "profile.json"
+
+# single tile names
+ATTRIBUTE_FILE = "attributes.yaml"
+DATASET_FILE = "dataset"
+DATAFRAME_FILE = "dataframe.csv"
+CARSDICT_FILE = "cars_dict"
+
+PROFILE = "profile"
+WINDOW = "window"
+OVERLAPS = "overlaps"
+ATTRIBUTES = "attributes"
+SAVING_INFO = "saving_info"
+
+
+class CarsDataset:
+    """
+    CarsDataset.
+
+    Internal CARS structure for organazing tiles
+    (xr.Datasets or pd.DataFrames).
+    """
+
+    def __init__(self, dataset_type, load_from_disk=None, name="unknown"):
+        """
+        Init function of CarsDataset.
+        If a path is provided, restore CarsDataset saved on disk.
+
+        :param dataset_type: type of dataset : 'arrays' or 'points'
+        :type dataset_type: str
+        :param load_from_disk: path to saved CarsDataset
+        :type load_from_disk: str
+
+        """
+
+        self.dataset_type = dataset_type
+        if dataset_type not in [
+            CARS_DS_TYPE_ARRAY,
+            CARS_DS_TYPE_POINTS,
+            CARS_DS_TYPE_DICT,
+        ]:
+            raise ValueError("wrong dataset type")
+
+        self.tiles = None
+        self.tiles_info = {}
+        self._tiling_grid = None
+        self.overlaps = None
+        self.attributes = {}
+        self.name = name
+
+        # Method to apply at the reception of tile
+        self.final_function = None
+
+        if load_from_disk is not None:
+            self.load_cars_dataset_from_disk(load_from_disk)
+
+    def __repr__(self):
+        """
+        Repr function
+        :return: printable self CarsDataset
+        """
+        return self.custom_print()
+
+    def __str__(self):
+        """
+        Str function
+        :return: printable self CarsDataset
+        """
+        return self.custom_print()
+
+    def custom_print(self):
+        """
+        Return string of self
+        :return: printable self
+        """
+
+        res = str(self.__class__) + ":  \n" "dataset_type: " + str(
+            self.dataset_type
+        ) + "\n" + "shape: " + str(self.shape) + "\n" + "tiling_grid: " + str(
+            self._tiling_grid
+        ) + "\n" + "overlaps: " + str(
+            self.overlaps
+        ) + "\n" + "tiles_info: " + str(
+            self.tiles_info
+        ) + "\n" + "attributes: " + str(
+            self.attributes
+        ) + "\n" + "tiles:" + str(
+            self.tiles
+        )
+        return res
+
+    @property
+    def shape(self):
+        """
+        Return the shape of tiling grid (nb_row, nb_col)
+        :return: shape of grid
+        """
+        return self.tiling_grid.shape[0], self.tiling_grid.shape[1]
+
+    @property
+    def tiling_grid(self):
+        """
+        Tiling grid, containing pixel windows of tiles
+
+        :return: tiling grid, of shape [N, M, 4],
+                 containing [row_min, row_max, col_min, col_max]
+        :rtype: np.ndarray
+        """
+        return self._tiling_grid
+
+    @tiling_grid.setter
+    def tiling_grid(self, new_grid):
+        """
+        Set tiling_grid
+
+        :param new_grid: new grid
+        :type new_grid: np.ndarray
+        """
+        self._tiling_grid = new_grid
+        # reset overlaps to zeros
+        self.overlaps = np.zeros(new_grid.shape)
+        # fill dataset grid with Nones
+        self.generate_none_tiles()
+
+    def __getitem__(self, key):
+        """
+        Get item : return the [row, col] dataset
+
+        :param key: tuple index
+
+        :return: tile
+        :rtype: xr.Dataset or pd.DataFrame
+        """
+
+        if isinstance(key, (tuple, list)):
+            if len(key) == 2:
+                res = self.tiles[key[0]][key[1]]
+            elif len(key) == 1:
+                res = self.tiles[key[0]]
+            else:
+                raise ValueError("Too many indexes, expected 1 or 2")
+        else:
+            if isinstance(key, int):
+                res = self.tiles[key]
+            else:
+                raise ValueError("Index type not supported")
+
+        return res
+
+    def __setitem__(self, key, newvalue):
+        """
+        Set new tile
+
+        :param key: tuple of row and col indexes
+        :type key: tuple(int, int)
+        :param newvalue: tile to set
+        """
+        if isinstance(key, (tuple, list)):
+            if len(key) == 2:
+                self.tiles[key[0]][key[1]] = newvalue
+            else:
+                raise ValueError("Too many indexes, expected 2")
+        else:
+            raise ValueError("Index type not supported")
+
+    def load_single_tile(self, tile_path_name: str):
+        """
+        Load a single tile
+
+        :param tile_path_name: Path of tile to load
+        :type tile_path_name: str
+
+        :return: single tile
+        :rtype: xarray Dataset or Panda dataframe to file
+
+        """
+
+        functions = {
+            CARS_DS_TYPE_ARRAY: load_single_tile_array,
+            CARS_DS_TYPE_POINTS: load_single_tile_points,
+            CARS_DS_TYPE_DICT: load_single_tile_dict,
+        }
+
+        return functions[self.dataset_type](tile_path_name)
+
+    def save_single_tile(self, tile, tile_path_name: str):
+        """
+        Save xarray Dataset or Panda dataframe to file
+
+        :param tile: tile to save
+        :type tile: xr.Dataset or pd.DataFrame
+        :param tile_path_name: Path of file to save in
+        """
+
+        functions = {
+            CARS_DS_TYPE_ARRAY: save_single_tile_array,
+            CARS_DS_TYPE_POINTS: save_single_tile_points,
+            CARS_DS_TYPE_DICT: save_single_tile_dict,
+        }
+
+        return functions[self.dataset_type](tile, tile_path_name)
+
+    def run_save(self, future_result, file_name: str, **kwargs):
+        """
+        Save future result when arrived
+
+        :param future_result: xarray.Dataset received
+        :param file_name: filename to save data to
+        """
+
+        functions = {
+            CARS_DS_TYPE_ARRAY: run_save_arrays,
+            CARS_DS_TYPE_POINTS: run_save_points,
+        }
+
+        return functions[self.dataset_type](future_result, file_name, **kwargs)
+
+    def get_window_as_dict(self, row, col, from_terrain=False, resolution=1):
+        """
+        Get window in pixels for rasterio. Set from_terrain if tiling grid
+        was defined in geographic coordinates.
+
+        :param row: row
+        :type row: int
+        :param col: col
+        :type col: int
+        :param from_terrain: true if in terrain coordinates
+        :type from_terrain: bool
+        :param resolution: resolution
+        :type resolution: float
+
+        :return: New window :  {
+            "row_min" : row_min ,
+            "row_max" : row_max
+            "col_min" : col_min
+            "col_max" : col_max
+            }
+        :rtype: Dict
+
+        """
+
+        row_min = np.min(self.tiling_grid[:, :, 0])
+        col_min = np.min(self.tiling_grid[:, :, 2])
+        col_max = np.max(self.tiling_grid[:, :, 3])
+
+        window_arr = np.copy(self.tiling_grid[row, col, :])
+
+        if from_terrain:
+            #  row -> y axis : reversed by convention
+            window = np.array(
+                [
+                    col_max - window_arr[3],
+                    col_max - window_arr[2],
+                    window_arr[0] - row_min,
+                    window_arr[1] - row_min,
+                ]
+            )
+
+        else:
+            window = np.array(
+                [
+                    window_arr[0] - row_min,
+                    window_arr[1] - row_min,
+                    window_arr[2] - col_min,
+                    window_arr[3] - col_min,
+                ]
+            )
+
+        # normalize with resolution
+        window = np.round(window / resolution)
+
+        new_window = {
+            "row_min": int(window[0]),
+            "row_max": int(window[1]),
+            "col_min": int(window[2]),
+            "col_max": int(window[3]),
+        }
+
+        return new_window
+
+    def create_grid(  # pylint: disable=too-many-positional-arguments
+        self,
+        nb_col: int,
+        nb_row: int,
+        row_split: int,
+        col_split: int,
+        row_overlap: int,
+        col_overlap: int,
+    ):
+        """
+        Generate grid of positions by splitting [0, nb_row]x[0, nb_col]
+        in splits of xsplit x ysplit size
+
+        :param nb_col : number of columns
+        :param nb_row : number of lines
+        :param col_split: width of splits
+        :param row_split: height of splits
+        :param col_overlap: overlap to apply on rows
+        :param row_overlap: overlap to apply on cols
+
+        """
+        nb_col_splits = math.ceil(nb_col / row_split)
+        nb_row_splits = math.ceil(nb_row / col_split)
+
+        row_min, row_max = 0, nb_row
+        col_min, col_max = 0, nb_col
+
+        out_grid = np.ndarray(
+            shape=(nb_row_splits, nb_col_splits, 4), dtype=int
+        )
+
+        out_overlap = np.ndarray(
+            shape=(nb_row_splits, nb_col_splits, 4), dtype=int
+        )
+
+        for i in range(0, nb_row_splits):
+            for j in range(0, nb_col_splits):
+                row_down = row_min + row_split * i
+                col_left = col_min + col_split * j
+                row_up = min(row_max, row_min + (i + 1) * row_split)
+                col_right = min(col_max, col_min + (j + 1) * col_split)
+
+                out_grid[i, j, 0] = row_down
+                out_grid[i, j, 1] = row_up
+                out_grid[i, j, 2] = col_left
+                out_grid[i, j, 3] = col_right
+
+                # fill overlap [OL_row_down, OL_row_up, OL_col_left,
+                #  OL_col_right]
+                out_overlap[i, j, 0] = row_down - max(
+                    row_min, row_down - row_overlap
+                )
+                out_overlap[i, j, 1] = (
+                    min(row_max, row_up + row_overlap) - row_up
+                )
+                out_overlap[i, j, 2] = col_left - max(
+                    col_min, col_left - col_overlap
+                )
+                out_overlap[i, j, 3] = (
+                    min(col_right, col_right + col_overlap) - col_right
+                )
+
+        self.tiling_grid = out_grid
+        self.overlaps = out_overlap
+
+    def generate_none_tiles(self):
+        """
+        Generate the structure of data tiles, with Nones, according
+            to grid shape.
+
+        """
+
+        self.tiles = create_none(
+            self.tiling_grid.shape[0], self.tiling_grid.shape[1]
+        )
+
+    def create_empty_copy(self, cars_ds):
+        """
+        Copy attributes, grid, overlaps, and create Nones.
+
+        :param cars_ds: CarsDataset to copy
+        :type cars_ds: CarsDataset
+
+        """
+
+        self.tiles_info = copy.deepcopy(cars_ds.tiles_info)
+        self.tiling_grid = copy.deepcopy(cars_ds.tiling_grid)
+        self.overlaps = copy.deepcopy(cars_ds.overlaps)
+
+        self.tiles = []
+        for _ in range(cars_ds.overlaps.shape[0]):
+            tiles_row = []
+            for _ in range(cars_ds.overlaps.shape[1]):
+                tiles_row.append(None)
+            self.tiles.append(tiles_row)
+
+    def generate_descriptor(  # pylint: disable=too-many-positional-arguments
+        self, future_result, file_name, tag=None, dtype=None, nodata=None
+    ):
+        """
+        Generate de rasterio descriptor for the given future result
+
+        Only works with pixelic tiling grid
+
+        :param future_result: Future result
+        :type future_result: xr.Dataset
+        :param file_name: file name to save futures to
+        :type file_name: str
+        :param tag: tag to save
+        :type tag: str
+        :param dtype: dtype
+        :type dtype: str
+        :param nodata: no data value
+        :type nodata: float
+        """
+
+        # Get profile from 1st finished future
+        new_profile = get_profile_for_tag_dataset(future_result, tag)
+
+        if "width" not in new_profile or "height" not in new_profile:
+            logging.debug(
+                "CarsDataset doesn't have a profile, default is given"
+            )
+            new_profile = DefaultGTiffProfile(count=new_profile["count"])
+            new_profile["height"] = np.max(self.tiling_grid[:, :, 1])
+            new_profile["width"] = np.max(self.tiling_grid[:, :, 3])
+
+        # Change dtype
+        new_profile["dtype"] = dtype
+        if nodata is not None:
+            new_profile["nodata"] = nodata
+
+        descriptor = rio.open(
+            file_name, "w+", **new_profile, BIGTIFF="IF_SAFER"
+        )
+
+        return descriptor
+
+    def save_cars_dataset(self, directory):
+        """
+        Save whole CarsDataset to given directory, including tiling grids,
+        attributes, overlaps, and all the xr.Dataset or pd.DataFrames.
+
+        :param directory: Path where to save  self CarsDataset
+        :type directory: str
+
+        """
+
+        # Create CarsDataset folder
+        safe_makedirs(directory)
+
+        if self.tiles is None:
+            logging.error("No tiles managed by CarsDatasets")
+            raise RuntimeError("No tiles managed by CarsDatasets")
+
+        # save tiles info
+        tiles_info_file = os.path.join(directory, TILES_INFO_FILE)
+        save_dict(self.tiles_info, tiles_info_file)
+
+        # save grid
+        grid_file = os.path.join(directory, GRID_FILE)
+        save_numpy_array(self.tiling_grid, grid_file)
+
+        # save overlap
+        overlap_file = os.path.join(directory, OVERLAP_FILE)
+        save_numpy_array(self.overlaps, overlap_file)
+
+        nb_rows, nb_cols = self.tiling_grid.shape[0], self.tiling_grid.shape[1]
+
+        # save each tile
+        for col in range(nb_cols):
+            for row in range(nb_rows):
+                # Get name
+                current_tile_path_name = create_tile_path(col, row, directory)
+
+                # save tile
+                self.save_single_tile(
+                    self.tiles[row][col], current_tile_path_name
+                )
+
+    def load_cars_dataset_from_disk(self, directory):
+        """
+        Load whole CarsDataset from given directory
+
+        :param directory: Path where is saved CarsDataset to load
+        :type directory: str
+
+        """
+
+        # get tiles info
+        tiles_info_file = os.path.join(directory, TILES_INFO_FILE)
+        self.tiles_info = load_dict(tiles_info_file)
+
+        # load grid
+        grid_file = os.path.join(directory, GRID_FILE)
+        self.tiling_grid = load_numpy_array(grid_file)
+
+        nb_rows, nb_cols = self.tiling_grid.shape[0], self.tiling_grid.shape[1]
+
+        # load overlap
+        overlap_file = os.path.join(directory, OVERLAP_FILE)
+        self.overlaps = load_numpy_array(overlap_file)
+
+        # load each tile
+        self.tiles = []
+        for row in range(nb_rows):
+            tiles_row = []
+            for col in range(nb_cols):
+                # Get name
+                current_tile_path_name = create_tile_path(col, row, directory)
+
+                # load tile
+                tiles_row.append(self.load_single_tile(current_tile_path_name))
+
+            self.tiles.append(tiles_row)
+
+
+def run_save_arrays(future_result, file_name, tag=None, descriptor=None):
+    """
+    Save future when arrived
+
+    :param future_result: xarray.Dataset received
+    :type future_result: xarray.Dataset
+    :param file_name: filename to save data to
+    :type file_name: str
+    :param tag: dataset tag to rasterize
+    :type tag: str
+    :param descriptor: rasterio descriptor
+    """
+    # write future result using saved window and overlaps
+
+    save_dataset(
+        future_result,
+        file_name,
+        tag,
+        use_windows_and_overlaps=True,
+        descriptor=descriptor,
+    )
+
+
+def run_save_points(  # pylint: disable=too-many-positional-arguments
+    future_result,
+    file_name,
+    overwrite=False,
+    save_by_pair=False,
+    point_cloud_format="csv",
+    overwrite_file_name=True,
+):
+    """
+    Save future result when arrived
+
+    :param future_result: pandas Dataframe received
+    :type future_result: pandas Dataframe
+    :param file_name: filename to save data to
+    :type file_name: str
+    :param overwrite: overwrite file
+    :type overwrite: bool
+    :param point_cloud_format: output point cloud format
+    :type point_cloud_format: str
+    :param overwrite_file_name: generate a new filename from input filename
+        matching input dataframe attributes
+    :type overwrite_file_name: bool
+    """
+
+    # Save
+    save_all_dataframe(
+        future_result,
+        file_name,
+        save_by_pair=save_by_pair,
+        overwrite=overwrite,
+        point_cloud_format=point_cloud_format,
+        overwrite_file_name=overwrite_file_name,
+    )
+
+
+def load_single_tile_array(tile_path_name: str) -> xr.Dataset:
+    """
+    Load a xarray tile
+
+    :param tile_path_name: Path of tile to load
+    :type tile_path_name: str
+
+    :return: tile dataset
+    :rtype: xr.Dataset
+
+    """
+
+    # get dataset
+    dataset_file_name = os.path.join(tile_path_name, DATASET_FILE)
+    if not os.path.exists(dataset_file_name):
+        logging.error("Tile {} does not exists".format(dataset_file_name))
+        return None
+    with open(dataset_file_name, "rb") as handle:
+        dataset = pickle.load(handle)
+
+    # get attributes
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    attributes = load_dict(attributes_file_name)
+
+    # Format transformation
+    if PROFILE in attributes:
+        attributes[PROFILE] = dict_profile_to_rio_profile(attributes[PROFILE])
+
+    # add to dataset
+    dataset.attrs.update(attributes)
+
+    return dataset
+
+
+def load_single_tile_points(tile_path_name: str):
+    """
+    Load a panda dataframe
+
+    :param tile_path_name: Path of tile to load
+    :type tile_path_name: str
+
+    :return: Tile dataframe
+    :rtype: Panda dataframe
+
+    """
+
+    # get dataframe
+    dataframe_file_name = os.path.join(tile_path_name, DATAFRAME_FILE)
+
+    if not os.path.exists(dataframe_file_name):
+        logging.error("Tile {} does not exists".format(dataframe_file_name))
+        return None
+
+    with open(dataframe_file_name, "rb") as handle:
+        dataframe = pickle.load(handle)
+
+    # get attributes
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    attributes = load_dict(attributes_file_name)
+
+    # Format transformation
+
+    # add to dataframe
+    dataframe.attrs.update(attributes)
+
+    return dataframe
+
+
+def load_single_tile_dict(tile_path_name: str):
+    """
+    Load a CarsDict
+
+    :param tile_path_name: Path of tile to load
+    :type tile_path_name: str
+
+    :return: Tile dataframe
+    :rtype: Panda dataframe
+
+    """
+
+    # get dataframe
+    dict_file_name = os.path.join(tile_path_name, CARSDICT_FILE)
+
+    if not os.path.exists(dict_file_name):
+        logging.error("Tile {} does not exists".format(dict_file_name))
+        return None
+
+    with open(dict_file_name, "rb") as handle:
+        dict_cars = pickle.load(handle)
+
+    # get attributes
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    attributes = load_dict(attributes_file_name)
+
+    # Format transformation
+
+    # add to dataframe
+    dict_cars.attrs.update(attributes)
+
+    return dict_cars
+
+
+def save_single_tile_array(dataset: xr.Dataset, tile_path_name: str):
+    """
+    Save xarray to directory, saving the data in a different file that
+    the attributes (saved in a .json next to it).
+
+    :param dataset: dataset to save
+    :type dataset: xr.Dataset
+    :param tile_path_name: Path of file to save in
+    :type tile_path_name: str
+    """
+
+    if dataset is None:
+        logging.debug("Tile is None: not saved")
+        return
+
+    # Create tile folder
+    safe_makedirs(tile_path_name)
+
+    # save attributes
+    saved_dataset_attrs = copy.copy(dataset.attrs)
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    if dataset.attrs is None:
+        attributes = {}
+    else:
+        attributes = dataset.attrs
+
+    # Format transformation
+    if PROFILE in attributes:
+        attributes[PROFILE] = rio_profile_to_dict_profile(attributes[PROFILE])
+
+    # dump
+    # separate attributes
+    dataset.attrs, custom_attributes = separate_dicts(
+        attributes, [PROFILE, WINDOW, OVERLAPS, SAVING_INFO, ATTRIBUTES]
+    )
+    # save
+    save_dict(custom_attributes, attributes_file_name)
+    dataset_file_name = os.path.join(tile_path_name, DATASET_FILE)
+    with open(dataset_file_name, "wb") as handle:
+        pickle.dump(dataset, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+    # Retrieve attrs
+    dataset.attrs = saved_dataset_attrs
+
+
+def save_single_tile_points(dataframe, tile_path_name: str):
+    """
+    Save dataFrame to directory, saving the data in a different file that
+    the attributes (saved in a .json next to it).
+
+    :param dataframe: dataframe to save
+    :type dataframe: pd.DataFrame
+    :param tile_path_name: Path of file to save in
+    :type tile_path_name: str
+    """
+    if dataframe is None:
+        logging.debug("Tile is None: not saved")
+        return
+    # Create tile folder
+    safe_makedirs(tile_path_name)
+
+    # save attributes
+    saved_dataframe_attrs = copy.copy(dataframe.attrs)
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    if dataframe.attrs is None:
+        attributes = {}
+    else:
+        attributes = dataframe.attrs
+
+    # Format transformation
+
+    # dump
+    # separate attributes
+    dataframe.attrs, custom_attributes = separate_dicts(
+        attributes, [SAVING_INFO, ATTRIBUTES]
+    )
+    # save
+    save_dict(custom_attributes, attributes_file_name)
+    dataframe_file_name = os.path.join(tile_path_name, DATAFRAME_FILE)
+    with open(dataframe_file_name, "wb") as handle:
+        pickle.dump(dataframe, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+    # Retrieve attrs
+    dataframe.attrs = saved_dataframe_attrs
+
+
+def save_single_tile_dict(dict_cars, tile_path_name: str):
+    """
+    Save cars_dict to directory, saving the data in a different file that
+    the attributes (saved in a .json next to it).
+
+    :param dict_cars: dataframe to save
+    :type dict_cars: pd.DataFrame
+    :param tile_path_name: Path of file to save in
+    :type tile_path_name: str
+    """
+    # Create tile folder
+    safe_makedirs(tile_path_name)
+
+    # save attributes
+    saved_dict_cars_attrs = copy.copy(dict_cars.attrs)
+    attributes_file_name = os.path.join(tile_path_name, ATTRIBUTE_FILE)
+    if dict_cars.attrs is None:
+        attributes = {}
+    else:
+        attributes = dict_cars.attrs
+
+    # Format transformation
+
+    # dump
+    # separate attributes
+    dict_cars.attrs, custom_attributes = separate_dicts(
+        attributes, [SAVING_INFO, ATTRIBUTES]
+    )
+    # save
+    save_dict(custom_attributes, attributes_file_name)
+    dict_cars_file_name = os.path.join(tile_path_name, CARSDICT_FILE)
+    with open(dict_cars_file_name, "wb") as handle:
+        pickle.dump(dict_cars, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+    # Retrieve attrs
+    dict_cars.attrs = saved_dict_cars_attrs
+
+
+def fill_dataset(  # pylint: disable=too-many-positional-arguments
+    dataset,
+    saving_info=None,
+    window=None,
+    profile=None,
+    attributes=None,
+    overlaps=None,
+):
+    """
+    From a full xarray dataset, fill info properly.
+    User can fill with saving information (containing CarsDataset id),
+    window of current tile and its overlaps,
+    rasterio profile of full data, and attributes associated to data
+
+    :param dataset: dataset to fill
+    :type dataset: xarray_dataset
+    :param saving_info: created by Orchestrator.get_saving_infos
+    :type saving_info: dict
+    :param window:
+    :type window: dict
+    :param profile:
+    :type profile: dict
+    :param attributes:
+    :type attributes: dict
+
+    """
+
+    if attributes is not None:
+        dataset.attrs[ATTRIBUTES] = attributes
+
+    if saving_info is not None:
+        dataset.attrs[SAVING_INFO] = saving_info
+
+    if window is not None:
+        dataset.attrs[WINDOW] = window
+
+    if overlaps is not None:
+        dataset.attrs[OVERLAPS] = overlaps
+
+    if profile is not None:
+        dataset.attrs[PROFILE] = profile
+
+
+def fill_dataframe(dataframe, saving_info=None, attributes=None):
+    """
+    From a full pandas dataframe, fill info properly.
+    User can fill with saving information (containing CarsDataset id),
+    and attributes associated to data
+
+
+    :param dataframe: dataframe to fill
+    :type dataframe: pandas dataframe
+    :param saving_info: created by Orchestrator.get_saving_infos
+    :type saving_info: dict
+    :param attributes:
+    :type attributes: dict
+
+    """
+
+    if attributes is not None:
+        dataframe.attrs[ATTRIBUTES] = attributes
+
+    if saving_info is not None:
+        dataframe.attrs[SAVING_INFO] = saving_info
+
+
+def fill_dict(data_dict, saving_info=None, attributes=None):
+    """
+    From a fulldict, fill info properly.
+    User can fill with saving information (containing CarsDataset id),
+    and attributes associated to data
+
+
+    :param data_dict: dictionnary to fill
+    :type data_dict: Dict
+    :param saving_info: created by Orchestrator.get_saving_infos
+    :type saving_info: dict
+    :param attributes: attributes associated to data
+    :type attributes: dict
+
+    """
+
+    # TODO only use CarsDict
+
+    if isinstance(data_dict, dict):
+        if attributes is not None:
+            data_dict[ATTRIBUTES] = attributes
+
+        if saving_info is not None:
+            data_dict[SAVING_INFO] = saving_info
+
+    elif isinstance(data_dict, cars_dict.CarsDict):
+        if attributes is not None:
+            data_dict.attrs[ATTRIBUTES] = attributes
+
+        if saving_info is not None:
+            data_dict.attrs[SAVING_INFO] = saving_info
+
+
+def save_all_dataframe(  # pylint: disable=too-many-positional-arguments
+    dataframe,
+    file_name,
+    save_by_pair=False,
+    overwrite=True,
+    point_cloud_format="csv",
+    overwrite_file_name=True,
+):
+    """
+    Save DataFrame to csv and laz format. The content of dataframe is merged to
+    the content of existing saved Dataframe, if overwrite==False
+    The option save_by_pair separate the dataframe by pair (one folder by pair)
+
+    :param file_name: file name to save data to
+    :type file_name: str
+    :param overwrite: overwrite file if exists
+    :type overwrite: bool
+    :param point_cloud_format: point cloud format (csv or laz)
+    :type point_cloud_format: str
+    :param overwrite_file_name: generate a new filename from input filename
+        matching input dataframe attributes, using only directory from input
+        filename
+    :type overwrite_file_name: bool
+    """
+
+    # generate filename if attributes have xstart and ystart settings
+    if overwrite_file_name:
+        if (
+            "attributes" in dataframe.attrs
+            and "xmin" in dataframe.attrs["attributes"]
+        ):
+            file_name = os.path.dirname(file_name)
+            file_name = os.path.join(
+                file_name,
+                (
+                    str(dataframe.attrs["attributes"]["xmin"])
+                    + "_"
+                    + str(dataframe.attrs["attributes"]["ymax"])
+                ),
+            )
+        elif "saving_info" in dataframe.attrs:
+            file_name = os.path.dirname(file_name)
+            file_name = os.path.join(
+                file_name,
+                (
+                    str(dataframe.attrs["saving_info"]["cars_ds_col"])
+                    + "_"
+                    + str(dataframe.attrs["saving_info"]["cars_ds_row"])
+                ),
+            )
+    if not save_by_pair:
+        save_dataframe(dataframe, file_name, overwrite, point_cloud_format)
+    else:
+        pairing_indexes = set(np.array(dataframe["global_id"]).flat)
+        source_pc_names = dataframe.attrs["attributes"]["source_pc_names"]
+        for pair_index in pairing_indexes:
+            dir_name = os.path.join(
+                os.path.dirname(file_name), source_pc_names[int(pair_index)]
+            )
+            safe_makedirs(dir_name)
+            base_name = os.path.basename(file_name)
+            points_indexes = dataframe["global_id"] == pair_index
+            file_name_by_pair = os.path.join(dir_name, base_name)
+            save_dataframe(
+                dataframe.loc[points_indexes],
+                file_name_by_pair,
+                overwrite,
+                point_cloud_format,
+            )
+
+
+def save_dataframe(
+    dataframe, file_name, overwrite=True, point_cloud_format="csv"
+):
+    """
+    Save dataframe (csv, laz, attr file)
+    """
+    # Save attributes
+    attributes_file_name = file_name + "_attrs.yaml"
+    save_dict(dataframe.attrs, attributes_file_name)
+
+    # Save point cloud to laz format
+
+    if point_cloud_format == "laz":
+        _, extension = os.path.splitext(file_name)
+        if "laz" not in extension:
+            file_name = file_name + ".laz"
+        dataframe_converter.convert_pcl_to_laz(dataframe, file_name)
+    elif point_cloud_format == "csv":
+        _, extension = os.path.splitext(file_name)
+        if "csv" not in extension:
+            file_name = file_name + ".csv"
+        if overwrite and os.path.exists(file_name):
+            dataframe.to_csv(file_name, index=False)
+        else:
+            if os.path.exists(file_name):
+                # merge files
+                existing_dataframe = pandas.read_csv(file_name)
+                merged_dataframe = pandas.concat(
+                    [existing_dataframe, dataframe],
+                    ignore_index=True,
+                    sort=False,
+                )
+                merged_dataframe.to_csv(file_name, index=False)
+            else:
+                dataframe.to_csv(file_name, index=False)
+    else:
+        raise RuntimeError(
+            "Invalid point cloud format {0}".format(point_cloud_format)
+        )
+
+
+def save_dataset(
+    dataset, file_name, tag, use_windows_and_overlaps=False, descriptor=None
+):
+    """
+    Reconstruct and save data.
+    In order to save properly the dataset to corresponding tiff file,
+    dataset must have been filled with saving info, profile, window,
+    overlaps (if not 0), and rasterio descriptor if already created.
+    See fill_dataset.
+
+    :param dataset: dataset to save
+    :type dataset: xr.Dataset
+    :param file_name: file name to save data to
+    :type file_name: str
+    :param tag: tag to reconstruct
+    :type tag: str
+    :param use_windows_and_overlaps: use saved window and overlaps
+    :type use_windows_and_overlaps: bool
+    :param descriptor: descriptor to use with rasterio
+    :type descriptor: rasterio dataset
+
+    """
+    if dataset is None:
+        logging.error("Tile is None: not saved ")
+        return
+
+    overlaps = get_overlaps_dataset(dataset)
+    window = get_window_dataset(dataset)
+
+    rio_window = None
+    overlap = [0, 0, 0, 0]
+    if use_windows_and_overlaps:
+        if window is None:
+            logging.debug("User wants to use window but none was set")
+
+        else:
+            rio_window = generate_rasterio_window(window)
+
+            if overlaps is not None:
+                overlap = [
+                    overlaps["up"],
+                    overlaps["down"],
+                    overlaps["left"],
+                    overlaps["right"],
+                ]
+    if len(dataset[tag].values.shape) > 2:
+        nb_rows, nb_cols = (
+            dataset[tag].values.shape[1],
+            dataset[tag].values.shape[2],
+        )
+
+        data = dataset[tag].values[
+            :,
+            overlap[0] : nb_rows - overlap[1],
+            overlap[2] : nb_cols - overlap[3],
+        ]
+    else:
+        nb_rows, nb_cols = (
+            dataset[tag].values.shape[0],
+            dataset[tag].values.shape[1],
+        )
+
+        data = dataset[tag].values[
+            overlap[0] : nb_rows - overlap[1],
+            overlap[2] : nb_cols - overlap[3],
+        ]
+
+    if tag == cst.EPI_TEXTURE and "int" in descriptor.dtypes[0]:
+        # Prepare color data for cast
+        data = np.nan_to_num(data, nan=descriptor.nodata)
+        data = np.round(data)
+
+    profile = get_profile_for_tag_dataset(dataset, tag)
+
+    new_profile = profile
+    if "width" not in new_profile or "height" not in new_profile:
+        logging.debug("CarsDataset doesn't have a profile, default is given")
+        new_profile = DefaultGTiffProfile(count=new_profile["count"])
+        new_profile["height"] = data.shape[0]
+        new_profile["width"] = data.shape[1]
+        new_profile["dtype"] = "float32"
+
+    bands_description = None
+    if tag in (cst.EPI_CLASSIFICATION, cst.RASTER_CLASSIF, cst.DSM_CLASSIF):
+        bands_description = dataset.coords[cst.BAND_CLASSIF].values
+    if tag in (cst.EPI_TEXTURE, cst.POINT_CLOUD_CLR_KEY_ROOT, cst.DSM_COLOR):
+        bands_description = dataset.coords[cst.BAND_IM].values
+    if tag in (cst.RASTER_SOURCE_PC, cst.DSM_SOURCE_PC):
+        bands_description = dataset.coords[cst.BAND_SOURCE_PC].values
+    if tag in (cst.EPI_FILLING, cst.RASTER_FILLING, cst.DSM_FILLING):
+        bands_description = dataset.coords[cst.BAND_FILLING].values
+    if tag in (
+        cst.RASTER_PERFORMANCE_MAP,
+        cst.RASTER_PERFORMANCE_MAP_RAW,
+        cst.DSM_PERFORMANCE_MAP,
+    ):
+        if cst.BAND_PERFORMANCE_MAP in dataset.coords:
+            bands_description = dataset.coords[cst.BAND_PERFORMANCE_MAP].values
+
+    classes_info_tag = None
+    if tag == cst.RASTER_PERFORMANCE_MAP:
+        classes_info_tag = dataset.attrs.get(
+            cst.RIO_TAG_PERFORMANCE_MAP_CLASSES, None
+        )
+
+    outputs.rasterio_write_georaster(
+        file_name,
+        data,
+        new_profile,
+        window=rio_window,
+        descriptor=descriptor,
+        bands_description=bands_description,
+        classes_info_tag=classes_info_tag,
+    )
+
+
+def create_tile_path(col: int, row: int, directory: str) -> str:
+    """
+    Create path of tile, according to its position in CarsDataset grid
+
+    :param col: numero of column
+    :type col: int
+    :param row: numero of row
+    :type row: int
+    :param directory: path where to save tile
+    :type directory: str
+
+    :return: full path
+    :rtype: str
+
+    """
+
+    tail = "col_" + repr(col) + "_row_" + repr(row)
+    name = os.path.join(directory, tail)
+
+    return name
+
+
+def save_numpy_array(array: np.ndarray, file_name: str):
+    """
+    Save numpy array to file
+
+    :param array: array to save
+    :type array: np.ndarray
+    :param file_name: numero of row
+    :type file_name: str
+
+    """
+
+    with open(file_name, "wb") as descriptor:
+        np.save(descriptor, array)
+
+
+def load_numpy_array(file_name: str) -> np.ndarray:
+    """
+    Load numpy array from file
+
+    :param file_name: numero of row
+    :type file_name: str
+
+    :return: array
+    :rtype: np.ndarray
+
+    """
+    with open(file_name, "rb") as descriptor:
+        return np.load(descriptor)
+
+
+def create_none(nb_row: int, nb_col: int):
+    """
+    Create a grid filled with None. The created grid is a 2D list :
+    ex: [[None, None], [None, None]]
+
+    :param nb_row: number of rows
+    :param nb_col: number of cols
+    :return: Grid filled with None
+    :rtype: list of list
+    """
+    grid = []
+    for _ in range(nb_row):
+        tmp = []
+        for _ in range(nb_col):
+            tmp.append(None)
+        grid.append(tmp)
+    return grid
+
+
+def overlap_array_to_dict(overlap):
+    """
+    Convert matrix of overlaps, to dict format used in CarsDatasets.
+    Input is : [o_up, o_down, o_left, o_right].
+    Output is : {"up": o_up, "down": o_down, "left": o_left, "right": o_right}
+
+    :param overlap: overlaps
+    :type overlap: List
+
+    :return: New overlaps
+    :rtype: Dict
+
+    """
+    new_overlap = {
+        "up": int(overlap[0]),
+        "down": int(overlap[1]),
+        "left": int(overlap[2]),
+        "right": int(overlap[3]),
+    }
+    return new_overlap
+
+
+def window_array_to_dict(window, overlap=None):
+    """
+    Convert matrix of windows, to dict format used in CarsDatasets.
+    Use overlaps if you want to get window with overlaps
+    inputs are :
+
+      - window : [row_min, row_max, col_min, col_max], with pixel format
+      - overlap (optional): [o_row_min, o_row_max, o_col_min, o_col_max]
+
+    outputs are :
+      {
+          "row_min" : row_min - o_row_min,
+          "row_max" : row_max + o_row_max,
+          "col_min" : col_min - o_col_min,
+          "col_max" : col_max - o_col_max,
+
+      }
+
+    :param window: window
+    :type window: List
+    :param overlap: overlaps
+    :type overlap: List
+
+    :return: New window
+    :rtype: Dict
+
+    """
+
+    new_window = {
+        "row_min": int(window[0]),
+        "row_max": int(window[1]),
+        "col_min": int(window[2]),
+        "col_max": int(window[3]),
+    }
+
+    if overlap is not None:
+        new_window["row_min"] -= int(overlap[0])
+        new_window["row_max"] += int(overlap[1])
+        new_window["col_min"] -= int(overlap[2])
+        new_window["col_max"] += int(overlap[3])
+
+    return new_window
+
+
+def dict_profile_to_rio_profile(dict_profile: Dict) -> Dict:
+    """
+    Transform a rasterio Profile transformed into serializable Dict,
+    into a rasterio profile.
+
+    :param profile: rasterio Profile transformed into serializable Dict
+    :type profile: Dict
+
+    :return: Profile
+    :rtype: Rasterio Profile
+
+    """
+
+    rio_profile = copy.copy(dict_profile)
+
+    transform = None
+    if "transform" in dict_profile:
+        if dict_profile["transform"] is not None:
+            transform = rio.Affine(
+                *dict_profile["transform"][0:6],
+            )
+    crs = None
+    if "crs" in dict_profile:
+        if dict_profile["crs"] is not None:
+            if (
+                isinstance(dict_profile["crs"], str)
+                and "EPSG:" in dict_profile["crs"]
+            ):
+                crs = pyproj.CRS(dict_profile["crs"].replace("EPSG:", ""))
+            else:
+                crs = pyproj.CRS(dict_profile["crs"])
+
+    rio_profile["crs"] = crs
+    rio_profile["transform"] = transform
+
+    return rio_profile
+
+
+def rio_profile_to_dict_profile(in_profile: Dict) -> Dict:
+    """
+    Transform a rasterio profile into a serializable Dict.
+
+    :param in_profile: rasterio Profile transformed into serializable Dict
+    :type in_profile: Dict
+
+    :return: Profile
+    :rtype: Dict
+
+    """
+
+    profile = copy.copy(in_profile)
+
+    profile = {**profile}
+    crs = None
+    if "crs" in profile:
+        if profile["crs"] is not None:
+            if isinstance(profile["crs"], str):
+                crs = profile["crs"]
+            else:
+                crs = profile["crs"].to_epsg()
+
+    transform = None
+    if "transform" in profile:
+        if profile["transform"] is not None:
+            transform = list(profile["transform"])[:6]
+
+    profile.update(crs=crs, transform=transform)
+
+    return profile
+
+
+def save_dict(dictionary, file_path: str):
+    """
+    Save dict to json file
+
+    :param dictionary: dictionary to save
+    :type dictionary: Dict
+    :param file_path: file path to use
+    :type file_path: str
+    """
+
+    dictionary_yaml = make_yaml_safe(dictionary)
+
+    with open(file_path, "w", encoding="utf8") as fstream:
+        yaml.safe_dump(
+            dictionary_yaml,
+            fstream,
+            allow_unicode=True,
+            sort_keys=False,
+        )
+
+
+def make_yaml_safe(obj):  # pylint: disable=too-many-return-statements
+    """
+    Converter
+    """
+    if isinstance(obj, dict):
+        return {make_yaml_safe(k): make_yaml_safe(v) for k, v in obj.items()}
+    if isinstance(obj, list):
+        return [make_yaml_safe(v) for v in obj]
+    if isinstance(obj, tuple):
+        return tuple(make_yaml_safe(v) for v in obj)
+    if isinstance(obj, (np.generic, np.number)):
+        return obj.item()
+
+    return obj
+
+
+def load_dict(file_path: str) -> Dict:
+    """
+    Load dict from json file
+
+    :param file_path: file path to use
+    :type file_path: str
+
+    """
+
+    with open(file_path, "r", encoding="utf8") as fstream:
+        dictionary = yaml.safe_load(fstream)
+
+    return dictionary
+
+
+def separate_dicts(dictionary, list_tags):
+    """
+    Separate a dict into two, the second one containing the given tags.
+
+    For example, {key1: val1, key2: val2, key3: val3}
+    with list_tags = [key2] will be split in :
+    {key1: val1, key3: val3} and {key2: val2}
+
+    """
+
+    dict1 = {}
+    dict2 = {}
+
+    for key in dictionary:
+        if key in list_tags:
+            dict2[key] = dictionary[key]
+        else:
+            dict1[key] = dictionary[key]
+
+    return dict1, dict2
+
+
+def get_attributes(obj):
+    """
+    Get attributes in .attrs of dataset or dataframe
+
+    :param obj: dataset or dataframe
+    :type obj: xr.Dataset or pandas.Dataframe
+    """
+
+    return obj.attrs.get(ATTRIBUTES, None)
+
+
+def get_window_dataset(dataset):
+    """
+    Get window in dataset
+
+    :param dataset: dataset
+    :type dataset: xr.Dataset
+    """
+
+    return dataset.attrs.get(WINDOW, None)
+
+
+def get_overlaps_dataset(dataset):
+    """
+    Get overlaps in dataset
+
+    :param dataset: dataset
+    :type dataset: xr.Dataset
+    """
+
+    return dataset.attrs.get(OVERLAPS, None)
+
+
+def get_profile_rasterio(dataset):
+    """
+    Get profile in dataset
+
+    :param dataset: dataset
+    :type dataset: xr.Dataset
+    """
+
+    return dataset.attrs.get(PROFILE, None)
+
+
+def get_profile_for_tag_dataset(dataset, tag: str) -> Dict:
+    """
+    Get profile according to layer to save.
+    This function modify current rasterio dataset to fix the number of
+    bands of the data associated to given tag.
+
+    :param tag: tag to use
+    :type tag: str
+
+    :return: Profile
+    :rtype: Rasterio Profile
+
+    """
+
+    new_profile = get_profile_rasterio(dataset)
+    if new_profile is None:
+        new_profile = {}
+
+    new_profile["count"] = 1
+    if len(dataset[tag].values.shape) > 2:
+        new_profile["count"] = dataset[tag].values.shape[0]
+
+    return new_profile
+
+
+def generate_rasterio_window(window: Dict) -> rio.windows.Window:
+    """
+    Generate rasterio window to use.
+
+    :param window: window to convert, containing 'row_min',
+                'row_max', 'col_min', 'col_max
+    :type window: dict
+
+    :return: rasterio window
+    :rtype: rio.windows.Window
+
+    """
+    returned_window = None
+
+    if window is not None:
+        return Window.from_slices(
+            (window["row_min"], window["row_max"]),
+            (window["col_min"], window["col_max"]),
+        )
+
+    return returned_window