# Copyright (C) 2023 Intel Corporation
#
# SPDX-License-Identifier: MIT
import json
import os
import os.path as osp
import re
import warnings
from typing import Dict, Optional, Type, TypeVar, Union
import numpy as np
import pandas as pd
from defusedxml import ElementTree
from datumaro.components.annotation import (
AnnotationType,
Bbox,
ExtractedMask,
Label,
LabelCategories,
MaskCategories,
)
from datumaro.components.dataset import DatasetItem
from datumaro.components.dataset_base import DEFAULT_SUBSET_NAME, DatasetBase, SubsetBase
from datumaro.components.errors import InvalidAnnotationError, InvalidFieldError, MissingFieldError
from datumaro.components.importer import ImportContext
from datumaro.components.media import Image, ImageFromFile
from datumaro.plugins.data_formats.coco.base import CocoInstancesBase
from datumaro.plugins.data_formats.coco.format import CocoTask
from datumaro.plugins.data_formats.coco.page_mapper import COCOPageMapper
from datumaro.util import parse_json_file
from datumaro.util.image import IMAGE_EXTENSIONS, lazy_image
T = TypeVar("T")
[docs]
class KaggleImageCsvBase(DatasetBase):
def __init__(
self,
path: str,
ann_file: str,
columns: Dict[str, str],
*,
subset: Optional[str] = DEFAULT_SUBSET_NAME,
ctx: Optional[ImportContext] = None,
):
super().__init__(ctx=ctx)
self._subset = subset
self._path = path
if "media" not in columns:
raise MissingFieldError("media")
self._label_cat = LabelCategories()
self._items = self._load_items(ann_file, columns)
self._categories = {AnnotationType.label: self._label_cat}
def _get_media_path(self, media_name: str):
media_path = osp.join(self._path, media_name)
if osp.exists(media_path):
return media_path
for ext in IMAGE_EXTENSIONS:
media_path_with_ext = media_path + ext
if osp.exists(media_path_with_ext):
return media_path_with_ext
return None
def _parse_bbox_coords(self, bbox_str):
coords = re.findall(r"[-+]?\d*\.\d+|\d+", bbox_str)
if len(coords) != 4:
raise ValueError("Bounding box coordinates must have exactly 4 values.")
# expected to output [x1, y1, x2, y2]
return [float(coord.strip()) for coord in coords]
def _load_annotations(
self, datas: list, indices: Dict[str, Union[int, Dict[str, int]]], bbox_flag: bool
):
if "label" in indices:
label_indices = indices["label"]
if isinstance(label_indices, dict):
labels = []
list_values = datas[1:]
index_to_label = {v: k for k, v in label_indices.items()}
present_labels = [
index_to_label[i + 1] for i, value in enumerate(list_values) if value == "1"
]
for label_name in present_labels:
label, cat = self._label_cat.find(label_name)
if not cat:
self._label_cat.add(label_name)
label, _ = self._label_cat.find(label_name)
labels.append(Label(label=label))
else:
label_name = str(datas[indices["label"]])
label, cat = self._label_cat.find(label_name)
if not cat:
self._label_cat.add(label_name)
label, _ = self._label_cat.find(label_name)
else:
_, cat = self._label_cat.find("object")
if not cat:
self._label_cat.add("object")
label = 0
if "label" in indices and not bbox_flag:
label_indices = indices["label"]
if isinstance(label_indices, dict):
return labels
return Label(label=label)
if bbox_flag:
if "bbox" in indices:
coords = self._parse_bbox_coords(datas[indices["bbox"]])
return Bbox(
label=label,
x=coords[0],
y=coords[1],
w=coords[2] - coords[0],
h=coords[3] - coords[1],
)
if "width" in indices and "height" in indices:
return Bbox(
label=label,
x=float(datas[indices["x1"]]),
y=float(datas[indices["y1"]]),
w=float(datas[indices["width"]]),
h=float(datas[indices["height"]]),
)
if "x2" in indices and "y2" in indices:
return Bbox(
label=label,
x=float(datas[indices["x1"]]),
y=float(datas[indices["y1"]]),
w=float(datas[indices["x2"]]) - float(datas[indices["x1"]]),
h=float(datas[indices["y2"]]) - float(datas[indices["y1"]]),
)
def _load_items(self, ann_file: str, columns: Dict[str, Union[str, list]]):
df = pd.read_csv(ann_file, header=None, on_bad_lines="skip")
df_fields = list(df.iloc[0])
indices = {"media": df_fields.index(columns["media"])}
if "label" in columns:
label_columns = columns["label"]
if isinstance(label_columns, list):
indices_label = {}
for label in label_columns:
indices_label[label] = df_fields.index(label)
indices.update({"label": indices_label})
else:
indices.update({"label": df_fields.index(label_columns)})
bbox_flag = False
bbox_index = columns.get("bbox")
if bbox_index:
bbox_flag = True
bbox_indices = {"x1", "x2", "y1", "y2", "width", "height"}
if isinstance(bbox_index, str):
indices["bbox"] = df_fields.index(bbox_index)
elif isinstance(bbox_index, dict):
indices.update(
{
key: df_fields.index(bbox_index[key])
for key in bbox_indices
if bbox_index.get(key)
}
)
if not (
{"x1", "x2", "y1", "y2"} <= bbox_indices
or {"x1", "y1", "width", "height"} <= bbox_indices
):
warnings.warn("Insufficient box coordinate is given for importing bounding boxes.")
bbox_flag = False
items = dict()
for _, row in df.iloc[1:].iterrows(): # Skip header row
data_info = list(row)
media_name = data_info[indices["media"]]
item_id = osp.splitext(media_name)[0]
media_path = self._get_media_path(media_name)
if not media_path or not osp.exists(media_path):
warnings.warn(
f"'{media_path}' is not existed in the directory, "
f"so we skip to create an dataset item according to {row}."
)
continue
ann = self._load_annotations(data_info, indices, bbox_flag)
if isinstance(ann, list):
for label in ann:
self._ann_types.add(label.type)
if item_id in items:
for label in ann:
items[item_id].annotations.append(label)
else:
items[item_id] = DatasetItem(
id=item_id,
subset=self._subset,
media=Image.from_file(path=media_path),
annotations=ann,
)
else:
self._ann_types.add(ann.type)
if item_id in items:
items[item_id].annotations.append(ann)
else:
items[item_id] = DatasetItem(
id=item_id,
subset=self._subset,
media=Image.from_file(path=media_path),
annotations=[ann],
)
return items.values()
[docs]
def categories(self):
return self._categories
def __iter__(self):
yield from self._items
[docs]
@classmethod
def build_cmdline_parser(cls, **kwargs):
parser = super().build_cmdline_parser(**kwargs)
parser.add_argument("--path", required=True)
parser.add_argument("--ann_file", required=True)
parser.add_argument("--columns", required=True, type=json.loads)
return parser
[docs]
class KaggleImageTxtBase(KaggleImageCsvBase):
def __init__(
self,
path: str,
ann_file: str,
columns: Dict[str, int],
*,
subset: Optional[str] = DEFAULT_SUBSET_NAME,
ctx: Optional[ImportContext] = None,
):
super().__init__(path=path, ann_file=ann_file, columns=columns, subset=subset, ctx=ctx)
def _load_items(self, ann_file: str, columns: Dict[str, Union[int, Dict]]):
bbox_flag = False
if "bbox" in columns:
bbox_flag = True
bbox_columns = columns.pop("bbox")
if isinstance(bbox_columns, dict):
if not (
all(item in bbox_columns for item in ["x1", "x2", "y1", "y2"])
or all(item in bbox_columns for item in ["x1", "y1", "width", "height"])
):
warnings.warn(
"Insufficient box coordinate is given for importing bounding boxes."
)
bbox_flag = False
columns.update(bbox_columns)
items = dict()
with open(ann_file, "r", encoding="utf-8") as f:
for line in f:
line = re.split(r"\s|,", line)
media_name = line[columns["media"]]
item_id = osp.splitext(media_name)[0]
media_path = self._get_media_path(media_name)
if not media_path or not osp.exists(media_path):
warnings.warn(
f"'{media_path}' is not existed in the directory, "
f"so we skip to create an dataset item according to {line}."
)
continue
ann = self._load_annotations(line, columns, bbox_flag)
self._ann_types.add(ann.type)
if item_id in items:
items[item_id].annotations.append(ann)
else:
items[item_id] = DatasetItem(
id=item_id,
subset=self._subset,
media=Image.from_file(path=media_path),
annotations=[ann],
)
return items.values()
[docs]
class KaggleImageMaskBase(DatasetBase):
def __init__(
self,
path: str,
mask_path: str,
labelmap_file: Optional[str] = None,
*,
subset: Optional[str] = DEFAULT_SUBSET_NAME,
ctx: Optional[ImportContext] = None,
):
super().__init__(ctx=ctx)
self._subset = subset
self._path = path
self._mask_path = mask_path
self._label_ids = []
self._categories = self._load_categories(labelmap_file)
self._items = self._load_items()
def _load_categories(self, label_map_file: Optional[str]):
label_map = dict()
if not label_map_file:
label_map["background"] = (0, 0, 0)
label_map["object"] = (255, 255, 255)
else:
df = pd.read_csv(label_map_file)
for _, row in df.iterrows():
name = row[0]
color = tuple([int(c) for c in row[1:]])
label_map[name] = color
label_categories = LabelCategories()
for label in label_map:
label_categories.add(label)
categories = {}
categories[AnnotationType.label] = label_categories
colormap = {}
for label_name, label_color in label_map.items():
label_id = label_categories.find(label_name)[0]
colormap[label_id] = label_color
self._label_ids.append(label_id)
categories[AnnotationType.mask] = MaskCategories(colormap)
return categories
def _load_items(self):
items = []
for media_name in sorted(os.listdir(self._path)):
id = osp.splitext(media_name)[0]
anns = []
for mask_name in os.listdir(self._mask_path):
if id in mask_name:
index_mask = lazy_image(
path=osp.join(self._mask_path, mask_name), dtype=np.int32
)
for label_id in self._label_ids:
anns.append(
ExtractedMask(
index_mask=index_mask,
index=label_id,
label=label_id,
)
)
self._ann_types.add(AnnotationType.mask)
items.append(
DatasetItem(
id=id,
subset=self._subset,
media=Image.from_file(path=osp.join(self._path, media_name)),
annotations=anns,
)
)
return items
[docs]
def categories(self):
return self._categories
def __iter__(self):
yield from self._items
[docs]
@classmethod
def build_cmdline_parser(cls, **kwargs):
parser = super().build_cmdline_parser(**kwargs)
parser.add_argument("--path", required=True)
parser.add_argument("--mask_path", required=True)
parser.add_argument("--labelmap_file")
return parser
[docs]
class KaggleVocBase(SubsetBase):
ann_extensions = ".xml"
def __init__(
self,
path: str,
ann_path: str,
*,
subset: Optional[str] = None,
ctx: Optional[ImportContext] = None,
):
super().__init__(subset=subset, ctx=ctx)
self._label_cat = LabelCategories()
self._items = []
self._size = None
for img_filename in sorted(os.listdir(path)):
if not img_filename.lower().endswith(tuple(IMAGE_EXTENSIONS)):
continue
item_id = osp.splitext(img_filename)[0]
img_file = osp.join(path, img_filename)
ann_file = osp.join(ann_path, item_id + self.ann_extensions)
annotations = (
self._parse_annotations(img_file, ann_file) if osp.isfile(ann_file) else []
)
for ann in annotations:
self._ann_types.add(ann.type)
media = Image.from_file(path=img_file, size=self._size)
self._items.append(
DatasetItem(
id=item_id,
subset=self._subset,
media=media,
annotations=annotations,
)
)
self._categories = {AnnotationType.label: self._label_cat}
def _parse_annotations(self, img_file: str, ann_file: str):
root_elem = ElementTree.parse(ann_file).getroot()
if root_elem.tag != "annotation":
raise MissingFieldError("annotation")
height = self._parse_field(root_elem, "size/height", int, required=False)
width = self._parse_field(root_elem, "size/width", int, required=False)
if height and width:
self._size = (height, width)
annotations = []
for obj_id, object_elem in enumerate(root_elem.iterfind("object")):
label_name = self._parse_field(object_elem, "name", str, required=True)
bbox_elem = object_elem.find("bndbox")
if not bbox_elem:
raise MissingFieldError("bndbox")
xmin = self._parse_field(bbox_elem, "xmin", float)
xmax = self._parse_field(bbox_elem, "xmax", float)
ymin = self._parse_field(bbox_elem, "ymin", float)
ymax = self._parse_field(bbox_elem, "ymax", float)
label_id, cat = self._label_cat.find(label_name)
if not cat:
self._label_cat.add(label_name)
label_id, _ = self._label_cat.find(label_name)
annotations.append(
Bbox(id=obj_id, label=label_id, x=xmin, y=ymin, w=xmax - xmin, h=ymax - ymin)
)
return annotations
@staticmethod
def _parse_field(root, xpath: str, cls: Type[T] = str, required: bool = True) -> Optional[T]:
elem = root.find(xpath)
if elem is None:
if required:
raise MissingFieldError(xpath)
else:
return None
if cls is str:
return elem.text
try:
return cls(elem.text)
except Exception as e:
raise InvalidFieldError(xpath) from e
[docs]
@classmethod
def build_cmdline_parser(cls, **kwargs):
parser = super().build_cmdline_parser(**kwargs)
parser.add_argument("--path", required=True)
parser.add_argument("--ann_path", required=True)
parser.add_argument("--subset")
return parser
[docs]
class KaggleYoloBase(KaggleVocBase, SubsetBase):
ann_extensions = ".txt"
def __init__(
self,
path: str,
ann_path: str,
*,
subset: Optional[str] = None,
ctx: Optional[ImportContext] = None,
):
super().__init__(path=path, ann_path=ann_path, subset=subset, ctx=ctx)
def _parse_annotations(self, img_file: str, ann_file: str):
image = ImageFromFile(path=img_file)
image_height, image_width = image.size
lines = []
with open(ann_file, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if line:
lines.append(line)
annotations = []
for obj_id, line in enumerate(lines):
parts = line.split()
if len(parts) != 5:
raise InvalidAnnotationError(
f"Unexpected field count {len(parts)} in the bbox description. "
"Expected 5 fields (label, xc, yc, w, h)."
)
label_name, xc, yc, w, h = parts
xc = float(xc)
yc = float(yc)
w = float(w)
h = float(h)
x = (xc - w * 0.5) * image_width
y = (yc - h * 0.5) * image_height
w *= image_width
h *= image_height
label_id, cat = self._label_cat.find(label_name)
if not cat:
self._label_cat.add(label_name)
label_id, _ = self._label_cat.find(label_name)
label_id, _ = self._label_cat.find(label_name)
annotations.append(Bbox(id=obj_id, label=label_id, x=x, y=y, w=w, h=h))
return annotations
[docs]
class KaggleCocoBase(CocoInstancesBase, SubsetBase):
def __init__(
self,
path: str,
ann_file: str,
*,
subset: Optional[str] = None,
ctx: Optional[ImportContext] = None,
stream: bool = False,
):
SubsetBase.__init__(self, subset=subset, ctx=ctx)
self._rootpath = path
self._images_dir = path
self._path = ann_file
self._task = CocoTask.instances
self._merge_instance_polygons = False
keep_original_category_ids = False
self._stream = stream
if not stream:
self._page_mapper = None # No use in case of stream = False
json_data = parse_json_file(ann_file)
self._load_categories(
json_data,
keep_original_ids=keep_original_category_ids,
)
self._items = self._load_items(json_data)
del json_data
else:
self._page_mapper = COCOPageMapper(ann_file)
categories_data = self._page_mapper.stream_parse_categories_data()
self._load_categories(
{"categories": categories_data},
keep_original_ids=keep_original_category_ids,
)
self._length = None
[docs]
@classmethod
def build_cmdline_parser(cls, **kwargs):
parser = super().build_cmdline_parser(**kwargs)
parser.add_argument("--path", required=True)
parser.add_argument("--ann_file", required=True)
parser.add_argument("--subset")
return parser
