From 0ce1a86525e5447009b5c1859c714f901dc84f6b Mon Sep 17 00:00:00 2001
From: Mathieu Beligon <mathieu@feedly.com>
Date: Sun, 13 Sep 2020 22:55:03 +0200
Subject: [PATCH] [common] (dataset) add names
---
common/polystar/common/models/image.py | 4 +-
common/polystar/common/models/object.py | 6 +++
.../polystar/common/utils/iterable_utils.py | 10 ++++
common/polystar/common/utils/tqdm.py | 11 ++++
.../common/dataset/improvement/zoom.py | 12 ++---
.../common/dataset/tensorflow_record.py | 15 ++++--
common/research/common/datasets/dataset.py | 53 ++++++++++++-------
.../research/common/datasets/image_dataset.py | 6 +--
.../datasets/roco/directory_roco_dataset.py | 7 ++-
.../common/datasets/roco/roco_annotation.py | 7 +--
.../common/datasets/roco/roco_dataset.py | 7 +--
.../roco}/roco_dataset_descriptor.py | 0
.../common/datasets/roco/zoo/twitch.py | 20 +++----
.../common/datasets/simple_dataset.py | 11 ++--
.../research/common/datasets/union_dataset.py | 11 ++--
.../image_pipeline_evaluation_reporter.py | 16 +++---
.../image_pipeline_evaluator.py | 44 ++++++++-------
...t_twith_datasets_from_manual_annotation.py | 21 ++++----
.../scripts/create_tensorflow_records.py | 13 +++--
.../common/scripts/improve_roco_by_zooming.py | 8 +--
.../common/scripts/visualize_dataset.py | 2 +-
.../common/unittests/datasets/test_dataset.py | 36 +++++++++----
22 files changed, 195 insertions(+), 125 deletions(-)
create mode 100644 common/polystar/common/utils/iterable_utils.py
create mode 100644 common/polystar/common/utils/tqdm.py
rename common/research/common/{dataset => datasets/roco}/roco_dataset_descriptor.py (100%)
diff --git a/common/polystar/common/models/image.py b/common/polystar/common/models/image.py
index fcc4faa..2cc6c34 100644
--- a/common/polystar/common/models/image.py
+++ b/common/polystar/common/models/image.py
@@ -2,9 +2,9 @@ from pathlib import Path
from typing import Iterable
import cv2
-from nptyping import Array
+import numpy as np
-Image = Array[int, ..., ..., 3]
+Image = np.ndarray
def load_image(image_path: Path, conversion: int = cv2.COLOR_BGR2RGB) -> Image:
diff --git a/common/polystar/common/models/object.py b/common/polystar/common/models/object.py
index e14c055..96b3f45 100644
--- a/common/polystar/common/models/object.py
+++ b/common/polystar/common/models/object.py
@@ -16,6 +16,9 @@ class ArmorColor(NoCaseEnum):
Red = auto()
Unknown = auto()
+ def __str__(self):
+ return self.name.lower()
+
ORDERED_ARMOR_COLORS = [ArmorColor.Blue, ArmorColor.Grey, ArmorColor.Red]
@@ -42,6 +45,9 @@ class Armor(Object):
number: ArmorNumber
color: ArmorColor
+ def __repr__(self):
+ return f"<{self} {self.color} {self.number}>"
+
class ObjectFactory:
@staticmethod
diff --git a/common/polystar/common/utils/iterable_utils.py b/common/polystar/common/utils/iterable_utils.py
new file mode 100644
index 0000000..a004688
--- /dev/null
+++ b/common/polystar/common/utils/iterable_utils.py
@@ -0,0 +1,10 @@
+from typing import Iterable
+
+from more_itertools import ilen
+
+
+def smart_len(it: Iterable) -> int:
+ try:
+ return len(it)
+ except AttributeError:
+ return ilen(it)
diff --git a/common/polystar/common/utils/tqdm.py b/common/polystar/common/utils/tqdm.py
new file mode 100644
index 0000000..9bf329d
--- /dev/null
+++ b/common/polystar/common/utils/tqdm.py
@@ -0,0 +1,11 @@
+from typing import Iterable, TypeVar
+
+from tqdm import tqdm
+
+T = TypeVar("T")
+
+
+def smart_tqdm(
+ it: Iterable[T], *args, desc: str = None, total: int = None, unit: str = "it", leave: bool = True, **kwargs
+) -> Iterable[T]:
+ return tqdm(it, *args, desc=desc, total=total, unit=unit, leave=leave, **kwargs)
diff --git a/common/research/common/dataset/improvement/zoom.py b/common/research/common/dataset/improvement/zoom.py
index 785ea34..9404b43 100644
--- a/common/research/common/dataset/improvement/zoom.py
+++ b/common/research/common/dataset/improvement/zoom.py
@@ -4,8 +4,7 @@ from typing import Iterable, List, Tuple
from polystar.common.models.box import Box
from polystar.common.models.image import Image
-from polystar.common.target_pipeline.objects_validators.in_box_validator import \
- InBoxValidator
+from polystar.common.target_pipeline.objects_validators.in_box_validator import InBoxValidator
from polystar.common.view.plt_results_viewer import PltResultViewer
from research.common.datasets.roco.roco_annotation import ROCOAnnotation
from research.common.datasets.roco.zoo.roco_datasets_zoo import ROCODatasetsZoo
@@ -13,7 +12,7 @@ from research.common.datasets.roco.zoo.roco_datasets_zoo import ROCODatasetsZoo
def crop_image_annotation(
image: Image, annotation: ROCOAnnotation, box: Box, min_coverage: float, name: str
-) -> Tuple[Image, ROCOAnnotation]:
+) -> Tuple[Image, ROCOAnnotation, str]:
objects = InBoxValidator(box, min_coverage).filter(annotation.objects, image)
objects = [copy(o) for o in objects]
for obj in objects:
@@ -25,7 +24,8 @@ def crop_image_annotation(
)
return (
image[box.y1 : box.y2, box.x1 : box.x2],
- ROCOAnnotation(w=box.w, h=box.h, objects=objects, has_rune=False, name=name),
+ ROCOAnnotation(w=box.w, h=box.h, objects=objects, has_rune=False),
+ name,
)
@@ -36,12 +36,12 @@ class Zoomer:
max_overlap: float
min_coverage: float
- def zoom(self, image: Image, annotation: ROCOAnnotation) -> Iterable[Tuple[Image, ROCOAnnotation]]:
+ def zoom(self, image: Image, annotation: ROCOAnnotation, name: str) -> Iterable[Tuple[Image, ROCOAnnotation, str]]:
boxes = [obj.box for obj in annotation.objects]
boxes = self._create_views_covering(boxes, annotation)
boxes = self._remove_overlapping_boxes(boxes)
return (
- crop_image_annotation(image, annotation, box, self.min_coverage, name=f"{annotation.name}_zoom_{i}")
+ crop_image_annotation(image, annotation, box, self.min_coverage, name=f"{name}_zoom_{i}")
for (i, box) in enumerate(boxes, 1)
)
diff --git a/common/research/common/dataset/tensorflow_record.py b/common/research/common/dataset/tensorflow_record.py
index 7c1e40c..9fb7067 100644
--- a/common/research/common/dataset/tensorflow_record.py
+++ b/common/research/common/dataset/tensorflow_record.py
@@ -16,15 +16,20 @@ from tqdm import tqdm
class TensorflowRecordFactory:
@staticmethod
def from_datasets(datasets: List[DirectoryROCODataset], prefix: str = ""):
- name = prefix + "_".join(d.name for d in datasets)
- writer = python_io.TFRecordWriter(str(TENSORFLOW_RECORDS_DIR / f"{name}.record"))
+ record_name = prefix + "_".join(d.name for d in datasets)
+ writer = python_io.TFRecordWriter(str(TENSORFLOW_RECORDS_DIR / f"{record_name}.record"))
c = 0
- for dataset in tqdm(datasets, desc=name, total=len(datasets), unit="dataset"):
- for image_path, annotation in tqdm(dataset, desc=dataset.name, total=len(dataset), unit="img", leave=False):
+ for dataset in tqdm(datasets, desc=record_name, total=len(datasets), unit="dataset"):
+ for image_path, annotation, _ in tqdm(
+ dataset, desc=dataset.name, total=len(dataset), unit="img", leave=False
+ ):
writer.write(_example_from_image_annotation(image_path, annotation).SerializeToString())
c += 1
writer.close()
- move(str(TENSORFLOW_RECORDS_DIR / f"{name}.record"), str(TENSORFLOW_RECORDS_DIR / f"{name}_{c}_imgs.record"))
+ move(
+ str(TENSORFLOW_RECORDS_DIR / f"{record_name}.record"),
+ str(TENSORFLOW_RECORDS_DIR / f"{record_name}_{c}_imgs.record"),
+ )
@staticmethod
def from_dataset(dataset: DirectoryROCODataset, prefix: str = ""):
diff --git a/common/research/common/datasets/dataset.py b/common/research/common/datasets/dataset.py
index eeb9a6e..7437687 100644
--- a/common/research/common/datasets/dataset.py
+++ b/common/research/common/datasets/dataset.py
@@ -3,6 +3,7 @@ from collections import deque
from typing import Callable, Generic, Iterable, Iterator, Tuple, TypeVar
from more_itertools import ilen
+from polystar.common.utils.iterable_utils import smart_len
from polystar.common.utils.misc import identity
ExampleT = TypeVar("ExampleT")
@@ -11,7 +12,7 @@ ExampleU = TypeVar("ExampleU")
TargetU = TypeVar("TargetU")
-class Dataset(Generic[ExampleT, TargetT], Iterable[Tuple[ExampleT, TargetT]], ABC):
+class Dataset(Generic[ExampleT, TargetT], Iterable[Tuple[ExampleT, TargetT, str]], ABC):
def __init__(self, name: str):
self.name = name
@@ -25,8 +26,13 @@ class Dataset(Generic[ExampleT, TargetT], Iterable[Tuple[ExampleT, TargetT]], AB
def targets(self) -> Iterable[TargetT]:
pass
+ @property
+ @abstractmethod
+ def names(self) -> Iterable[TargetT]:
+ pass
+
@abstractmethod
- def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT]]:
+ def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT, str]]:
pass
@abstractmethod
@@ -45,7 +51,10 @@ class Dataset(Generic[ExampleT, TargetT], Iterable[Tuple[ExampleT, TargetT]], AB
self, example_transformer: Callable[[ExampleT], ExampleU], target_transformer: Callable[[TargetT], TargetU]
) -> "Dataset[ExampleU, TargetU]":
return GeneratorDataset(
- self.name, lambda: ((example_transformer(example), target_transformer(target)) for example, target in self)
+ self.name,
+ lambda: (
+ (example_transformer(example), target_transformer(target), name) for example, target, name in self
+ ),
)
def __str__(self):
@@ -54,16 +63,13 @@ class Dataset(Generic[ExampleT, TargetT], Iterable[Tuple[ExampleT, TargetT]], AB
__repr__ = __str__
def check_consistency(self):
- targets, examples = self.targets, self.examples
- if isinstance(targets, list) and isinstance(examples, list):
- assert len(targets) == len(examples)
- assert ilen(targets) == ilen(examples)
+ assert smart_len(self.targets) == smart_len(self.examples) == smart_len(self.names)
class LazyUnzipper:
- def __init__(self, iterator: Iterator[Tuple]):
+ def __init__(self, iterator: Iterator[Tuple], n: int):
self._iterator = iterator
- self._memory = [deque(), deque()]
+ self._memory = [deque() for _ in range(n)]
def empty(self, i: int):
return self._iterator is None and not self._memory[i]
@@ -77,7 +83,9 @@ class LazyUnzipper:
else:
try:
elements = next(self._iterator)
- self._memory[1 - i].append(elements[1 - i])
+ for k in range(len(elements)):
+ if k != i:
+ self._memory[k].append(elements[k])
yield elements[i]
except StopIteration:
self._iterator = None
@@ -87,28 +95,33 @@ class LazyUnzipper:
class LazyDataset(Dataset[ExampleT, TargetT], ABC):
def __init__(self, name: str):
super().__init__(name)
- self._unzipper = LazyUnzipper(iter(self))
+ self._unzipper = None
@property
def examples(self) -> Iterable[ExampleT]:
- if self._unzipper.empty(0):
- self._unzipper = LazyUnzipper(iter(self))
- return self._unzipper.elements(0)
+ return self._elements(0)
@property
- def targets(self) -> Iterable[ExampleT]:
- if self._unzipper.empty(1):
- self._unzipper = LazyUnzipper(iter(self))
- return self._unzipper.elements(1)
+ def targets(self) -> Iterable[TargetT]:
+ return self._elements(1)
+
+ @property
+ def names(self) -> Iterable[str]:
+ return self._elements(2)
def __len__(self):
return ilen(self)
+ def _elements(self, i: int) -> Iterable:
+ if self._unzipper is None or self._unzipper.empty(i):
+ self._unzipper = LazyUnzipper(iter(self), 3)
+ return self._unzipper.elements(i)
+
class GeneratorDataset(LazyDataset[ExampleT, TargetT]):
- def __init__(self, name: str, generator: Callable[[], Iterator[Tuple[ExampleT, TargetT]]]):
+ def __init__(self, name: str, generator: Callable[[], Iterator[Tuple[ExampleT, TargetT, str]]]):
self.generator = generator
super().__init__(name)
- def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT]]:
+ def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT, str]]:
return self.generator()
diff --git a/common/research/common/datasets/image_dataset.py b/common/research/common/datasets/image_dataset.py
index 0476336..de58c6c 100644
--- a/common/research/common/datasets/image_dataset.py
+++ b/common/research/common/datasets/image_dataset.py
@@ -4,16 +4,16 @@ from typing import Iterator, List, Tuple
from memoized_property import memoized_property
from more_itertools import ilen
-from polystar.common.models.image import Image
+from polystar.common.models.image import Image, load_image
from research.common.datasets.dataset import Dataset, LazyDataset, TargetT
ImageDataset = Dataset[Image, TargetT]
class ImageFileDataset(LazyDataset[Path, TargetT], ABC):
- def __iter__(self) -> Iterator[Tuple[Path, TargetT]]:
+ def __iter__(self) -> Iterator[Tuple[Path, TargetT, str]]:
for image_file in self.image_files:
- yield image_file, self.target_from_image_file(image_file)
+ yield image_file, self.target_from_image_file(image_file), image_file.stem
@abstractmethod
def target_from_image_file(self, image_file: Path) -> TargetT:
diff --git a/common/research/common/datasets/roco/directory_roco_dataset.py b/common/research/common/datasets/roco/directory_roco_dataset.py
index 5df4021..1b84b8f 100644
--- a/common/research/common/datasets/roco/directory_roco_dataset.py
+++ b/common/research/common/datasets/roco/directory_roco_dataset.py
@@ -10,7 +10,6 @@ class DirectoryROCODataset(ImageDirectoryDataset[ROCOAnnotation]):
super().__init__(dataset_path / "image", name)
self.main_dir = dataset_path
self.annotations_dir: Path = self.main_dir / "image_annotation"
- self.annotations_dir: Path = self.main_dir / "image_annotation"
def target_from_image_file(self, image_file: Path) -> ROCOAnnotation:
return ROCOAnnotation.from_xml_file(self.annotations_dir / f"{image_file.stem}.xml")
@@ -20,6 +19,6 @@ class DirectoryROCODataset(ImageDirectoryDataset[ROCOAnnotation]):
self.images_dir.mkdir()
self.annotations_dir.mkdir()
- def add(self, image: Image, annotation: ROCOAnnotation):
- save_image(image, self.images_dir / f"{annotation.name}.jpg")
- (self.annotations_dir / f"{annotation.name}.xml").write_text(annotation.to_xml())
+ def add(self, image: Image, annotation: ROCOAnnotation, name: str):
+ save_image(image, self.images_dir / f"{name}.jpg")
+ (self.annotations_dir / f"{name}.xml").write_text(annotation.to_xml())
diff --git a/common/research/common/datasets/roco/roco_annotation.py b/common/research/common/datasets/roco/roco_annotation.py
index 4378679..0317bdd 100644
--- a/common/research/common/datasets/roco/roco_annotation.py
+++ b/common/research/common/datasets/roco/roco_annotation.py
@@ -11,8 +11,6 @@ from polystar.common.models.object import Object, ObjectFactory
@dataclass
class ROCOAnnotation:
- name: str
-
objects: List[Object]
has_rune: bool
@@ -23,13 +21,13 @@ class ROCOAnnotation:
@staticmethod
def from_xml_file(xml_file: Path) -> "ROCOAnnotation":
try:
- return ROCOAnnotation.from_xml_dict(xmltodict.parse(xml_file.read_text())["annotation"], xml_file.stem)
+ return ROCOAnnotation.from_xml_dict(xmltodict.parse(xml_file.read_text())["annotation"])
except Exception as e:
logging.exception(f"Error parsing annotation file {xml_file}")
raise e
@staticmethod
- def from_xml_dict(xml_dict: Dict, name: str) -> "ROCOAnnotation":
+ def from_xml_dict(xml_dict: Dict) -> "ROCOAnnotation":
json_objects = xml_dict.get("object", []) or []
json_objects = json_objects if isinstance(json_objects, list) else [json_objects]
roco_json_objects = [obj_json for obj_json in json_objects if not obj_json["name"].startswith("rune")]
@@ -40,7 +38,6 @@ class ROCOAnnotation:
has_rune=len(roco_json_objects) != len(json_objects),
w=int(xml_dict["size"]["width"]),
h=int(xml_dict["size"]["height"]),
- name=name,
)
def to_xml(self) -> str:
diff --git a/common/research/common/datasets/roco/roco_dataset.py b/common/research/common/datasets/roco/roco_dataset.py
index f2fb206..0d58f5d 100644
--- a/common/research/common/datasets/roco/roco_dataset.py
+++ b/common/research/common/datasets/roco/roco_dataset.py
@@ -1,6 +1,7 @@
-from research.common.datasets.image_dataset import (ImageDataset,
- ImageFileDataset)
+from polystar.common.models.image import Image
+from research.common.datasets.dataset import Dataset
+from research.common.datasets.image_dataset import ImageFileDataset
from research.common.datasets.roco.roco_annotation import ROCOAnnotation
-ROCODataset = ImageDataset[ROCOAnnotation]
+ROCODataset = Dataset[Image, ROCOAnnotation]
ROCOFileDataset = ImageFileDataset[ROCOAnnotation]
diff --git a/common/research/common/dataset/roco_dataset_descriptor.py b/common/research/common/datasets/roco/roco_dataset_descriptor.py
similarity index 100%
rename from common/research/common/dataset/roco_dataset_descriptor.py
rename to common/research/common/datasets/roco/roco_dataset_descriptor.py
diff --git a/common/research/common/datasets/roco/zoo/twitch.py b/common/research/common/datasets/roco/zoo/twitch.py
index 68ca46a..12eac54 100644
--- a/common/research/common/datasets/roco/zoo/twitch.py
+++ b/common/research/common/datasets/roco/zoo/twitch.py
@@ -9,16 +9,16 @@ from research.common.datasets.roco.roco_datasets import ROCODatasets
class TwitchROCODatasets(ROCODatasets):
directory = TWITCH_DSET_DIR / "v1"
- TWITCH_470149568 = ()
- TWITCH_470150052 = ()
- TWITCH_470151286 = ()
- TWITCH_470152289 = ()
- TWITCH_470152730 = ()
- TWITCH_470152838 = ()
- TWITCH_470153081 = ()
- TWITCH_470158483 = ()
+ T470149568 = ()
+ T470150052 = ()
+ T470151286 = ()
+ T470152289 = ()
+ T470152730 = ()
+ T470152838 = ()
+ T470153081 = ()
+ T470158483 = ()
@classmethod
def make_dataset(cls, dataset_name: str, *args: Any) -> DirectoryROCODataset:
- twitch_id = dataset_name[len("TWITCH_") :]
- return DirectoryROCODataset(cls.directory / twitch_id, f"T{twitch_id}")
+ twitch_id = dataset_name[len("T") :]
+ return DirectoryROCODataset(cls.directory / twitch_id, dataset_name)
diff --git a/common/research/common/datasets/simple_dataset.py b/common/research/common/datasets/simple_dataset.py
index 1209ae9..8e6ee1e 100644
--- a/common/research/common/datasets/simple_dataset.py
+++ b/common/research/common/datasets/simple_dataset.py
@@ -4,10 +4,11 @@ from research.common.datasets.dataset import Dataset, ExampleT, TargetT
class SimpleDataset(Dataset[ExampleT, TargetT]):
- def __init__(self, examples: Iterable[ExampleT], targets: Iterable[TargetT], name: str):
+ def __init__(self, examples: Iterable[ExampleT], targets: Iterable[TargetT], names: Iterable[str], name: str):
super().__init__(name)
self._examples = list(examples)
self._targets = list(targets)
+ self._names = list(names)
self.check_consistency()
@property
@@ -18,8 +19,12 @@ class SimpleDataset(Dataset[ExampleT, TargetT]):
def targets(self) -> List[TargetT]:
return self._targets
- def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT]]:
- return zip(self.examples, self.targets)
+ @property
+ def names(self) -> List[TargetT]:
+ return self._names
+
+ def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT, str]]:
+ return zip(self.examples, self.targets, self.names)
def __len__(self):
return len(self.examples)
diff --git a/common/research/common/datasets/union_dataset.py b/common/research/common/datasets/union_dataset.py
index 24f314d..4967631 100644
--- a/common/research/common/datasets/union_dataset.py
+++ b/common/research/common/datasets/union_dataset.py
@@ -1,16 +1,15 @@
from typing import Iterable, Iterator, Tuple
-from polystar.common.models.image import Image
-from research.common.datasets.dataset import ExampleT, LazyDataset, TargetT
-from research.common.datasets.image_dataset import ImageDataset
+from research.common.datasets.dataset import (Dataset, ExampleT, LazyDataset,
+ TargetT)
class UnionDataset(LazyDataset[ExampleT, TargetT]):
- def __init__(self, datasets: Iterable[ImageDataset[TargetT]], name: str):
- super().__init__(name)
+ def __init__(self, datasets: Iterable[Dataset[ExampleT, TargetT]], name: str = None):
self.datasets = list(datasets)
+ super().__init__(name or "_".join(d.name for d in self.datasets))
- def __iter__(self) -> Iterator[Tuple[Image, TargetT]]:
+ def __iter__(self) -> Iterator[Tuple[ExampleT, TargetT, str]]:
for dataset in self.datasets:
yield from dataset
diff --git a/common/research/common/image_pipeline_evaluation/image_pipeline_evaluation_reporter.py b/common/research/common/image_pipeline_evaluation/image_pipeline_evaluation_reporter.py
index 47cacc2..4e672b2 100644
--- a/common/research/common/image_pipeline_evaluation/image_pipeline_evaluation_reporter.py
+++ b/common/research/common/image_pipeline_evaluation/image_pipeline_evaluation_reporter.py
@@ -12,10 +12,12 @@ from polystar.common.utils.dataframe import format_df_column, format_df_row, for
from polystar.common.utils.markdown import MarkdownFile
from polystar.common.utils.time import create_time_id
from research.common.constants import DSET_DIR, EVALUATION_DIR
-from research.common.dataset.roco_dataset import ROCODataset
-from research.common.image_pipeline_evaluation.image_pipeline_evaluator import (ClassificationResults,
- ImagePipelineEvaluator,
- SetClassificationResults)
+from research.common.datasets.roco.roco_dataset import ROCOFileDataset
+from research.common.image_pipeline_evaluation.image_pipeline_evaluator import (
+ ClassificationResults,
+ ImagePipelineEvaluator,
+ SetClassificationResults,
+)
@dataclass
@@ -54,17 +56,17 @@ class ImagePipelineEvaluationReporter:
@staticmethod
def _report_dataset(
- mf: MarkdownFile, roco_datasets: List[ROCODataset], dataset_sizes: List[int], labels: List[Any]
+ mf: MarkdownFile, roco_datasets: List[ROCOFileDataset], dataset_sizes: List[int], labels: List[Any]
):
total = len(labels)
mf.paragraph(f"{total} images")
df = DataFrame(
{
- dataset.dataset_name: Counter(labels[start:end])
+ dataset.name: Counter(labels[start:end])
for dataset, start, end in zip(roco_datasets, np.cumsum([0] + dataset_sizes), np.cumsum(dataset_sizes))
}
).fillna(0)
- df["Total"] = sum([df[d.dataset_name] for d in roco_datasets])
+ df["Total"] = sum([df[d.name] for d in roco_datasets])
df["Repartition"] = (df["Total"] / total).map("{:.1%}".format)
mf.table(df)
diff --git a/common/research/common/image_pipeline_evaluation/image_pipeline_evaluator.py b/common/research/common/image_pipeline_evaluation/image_pipeline_evaluator.py
index e47782e..8c530b8 100644
--- a/common/research/common/image_pipeline_evaluation/image_pipeline_evaluator.py
+++ b/common/research/common/image_pipeline_evaluation/image_pipeline_evaluator.py
@@ -1,16 +1,18 @@
import logging
from dataclasses import dataclass
+from pathlib import Path
from time import time
-from typing import Any, Dict, Iterable, List, Sequence
+from typing import Any, Dict, Iterable, List, Sequence, Tuple
import numpy as np
-from sklearn.metrics import classification_report, confusion_matrix
-
from memoized_property import memoized_property
from polystar.common.image_pipeline.image_pipeline import ImagePipeline
-from polystar.common.models.image import Image
-from research.common.dataset.directory_roco_dataset import DirectoryROCODataset
-from research.common.image_pipeline_evaluation.image_dataset_generator import ImageDatasetGenerator
+from polystar.common.models.image import Image, load_image
+from research.common.datasets.roco.directory_roco_dataset import \
+ DirectoryROCODataset
+from research.robots_at_robots.dataset.armor_value_dataset import \
+ ArmorValueDatasetGenerator
+from sklearn.metrics import classification_report, confusion_matrix
@dataclass
@@ -48,23 +50,17 @@ class ImagePipelineEvaluator:
self,
train_roco_datasets: List[DirectoryROCODataset],
test_roco_datasets: List[DirectoryROCODataset],
- image_dataset_generator: ImageDatasetGenerator,
+ image_dataset_generator: ArmorValueDatasetGenerator,
):
logging.info("Loading data")
self.train_roco_datasets = train_roco_datasets
self.test_roco_datasets = test_roco_datasets
- (
- self.train_images_paths,
- self.train_images,
- self.train_labels,
- self.train_dataset_sizes,
- ) = image_dataset_generator.from_roco_datasets(train_roco_datasets)
- (
- self.test_images_paths,
- self.test_images,
- self.test_labels,
- self.test_dataset_sizes,
- ) = image_dataset_generator.from_roco_datasets(test_roco_datasets)
+ (self.train_images_paths, self.train_images, self.train_labels, self.train_dataset_sizes) = load_datasets(
+ train_roco_datasets, image_dataset_generator
+ )
+ (self.test_images_paths, self.test_images, self.test_labels, self.test_dataset_sizes) = load_datasets(
+ test_roco_datasets, image_dataset_generator
+ )
def evaluate_pipelines(self, pipelines: Iterable[ImagePipeline]) -> Dict[str, ClassificationResults]:
return {str(pipeline): self.evaluate(pipeline) for pipeline in pipelines}
@@ -87,3 +83,13 @@ class ImagePipelineEvaluator:
preds = pipeline.predict(images)
mean_time = (time() - t) / len(images)
return SetClassificationResults(np.asarray(labels), np.asarray(preds), mean_time)
+
+
+def load_datasets(
+ datasets: List[DirectoryROCODataset], image_dataset_generator: ArmorValueDatasetGenerator
+) -> Tuple[List[Path], List[Image], List[Any], List[int]]:
+ dataset_sizes = [len(d) for d in datasets]
+ dataset = image_dataset_generator.from_roco_datasets(datasets)
+ paths, targets = list(dataset.examples), list(dataset.targets)
+ images = list(map(load_image, paths))
+ return paths, images, targets, dataset_sizes
diff --git a/common/research/common/scripts/construct_twith_datasets_from_manual_annotation.py b/common/research/common/scripts/construct_twith_datasets_from_manual_annotation.py
index 767b0ca..459ae3d 100644
--- a/common/research/common/scripts/construct_twith_datasets_from_manual_annotation.py
+++ b/common/research/common/scripts/construct_twith_datasets_from_manual_annotation.py
@@ -4,10 +4,10 @@ from shutil import copy, make_archive, move, rmtree
from research.common.constants import (TWITCH_DSET_DIR,
TWITCH_DSET_ROBOTS_VIEWS_DIR,
TWITCH_ROBOTS_VIEWS_DIR)
-from research.common.dataset.roco_dataset_descriptor import \
- make_markdown_dataset_report
from research.common.datasets.roco.directory_roco_dataset import \
DirectoryROCODataset
+from research.common.datasets.roco.roco_dataset_descriptor import \
+ make_markdown_dataset_report
from research.common.scripts.construct_dataset_from_manual_annotation import \
construct_dataset_from_manual_annotations
from research.common.scripts.correct_annotations import AnnotationFileCorrector
@@ -26,22 +26,23 @@ def _correct_manual_annotations():
def _extract_runes_images():
all_twitch_dataset = _get_mixed_dataset()
- for annotation in all_twitch_dataset.image_annotations:
+ for image_file, annotation, _ in all_twitch_dataset:
if annotation.has_rune:
- copy(str(annotation.image_path), str(TWITCH_DSET_DIR / "runes" / annotation.image_path.name))
+ copy(str(image_file), str(TWITCH_DSET_DIR / "runes" / image_file.name))
def _separate_twitch_videos():
all_twitch_dataset = _get_mixed_dataset()
- for annotation in all_twitch_dataset.image_annotations:
- video_name = annotation.image_path.name.split("-")[0]
+ for image_file, annotation, _ in all_twitch_dataset:
+ video_name = image_file.name.split("-")[0]
dset_path = TWITCH_DSET_ROBOTS_VIEWS_DIR / video_name
images_path = dset_path / "image"
annotations_path = dset_path / "image_annotation"
images_path.mkdir(exist_ok=True, parents=True)
annotations_path.mkdir(exist_ok=True, parents=True)
- move(str(annotation.image_path), str(images_path / annotation.image_path.name))
- move(str(annotation.xml_path), str(annotations_path / annotation.xml_path.name))
+ move(str(image_file), str(images_path / image_file.name))
+ xml_name = f"{image_file.stem}.xml"
+ move(str(all_twitch_dataset.annotations_dir / xml_name), str(annotations_path / xml_name))
if list((TWITCH_DSET_ROBOTS_VIEWS_DIR / "image").glob("*")):
raise Exception(f"Some images remains unmoved")
for remaining_file in (TWITCH_DSET_ROBOTS_VIEWS_DIR / "image_annotation").glob("*"):
@@ -53,7 +54,7 @@ def _separate_twitch_videos():
def _make_global_report():
all_twitch_dataset = _get_mixed_dataset()
- make_markdown_dataset_report(all_twitch_dataset, all_twitch_dataset.dataset_path)
+ make_markdown_dataset_report(all_twitch_dataset, all_twitch_dataset.main_dir)
def _get_mixed_dataset() -> DirectoryROCODataset:
@@ -64,7 +65,7 @@ def _make_separate_reports():
for video_dset_path in TWITCH_DSET_ROBOTS_VIEWS_DIR.glob("*"):
if video_dset_path.is_dir():
twitch_dset = DirectoryROCODataset(video_dset_path, f"TWITCH_{video_dset_path.name}")
- make_markdown_dataset_report(twitch_dset, twitch_dset.dataset_path)
+ make_markdown_dataset_report(twitch_dset, twitch_dset.main_dir)
if __name__ == "__main__":
diff --git a/common/research/common/scripts/create_tensorflow_records.py b/common/research/common/scripts/create_tensorflow_records.py
index 6983ff3..7563a0a 100644
--- a/common/research/common/scripts/create_tensorflow_records.py
+++ b/common/research/common/scripts/create_tensorflow_records.py
@@ -15,17 +15,16 @@ def create_one_record_per_roco_dset():
def create_twitch_records():
TensorflowRecordFactory.from_datasets(
[
- TwitchROCODatasets.TWITCH_470149568,
- TwitchROCODatasets.TWITCH_470150052,
- TwitchROCODatasets.TWITCH_470151286,
- TwitchROCODatasets.TWITCH_470152289,
- TwitchROCODatasets.TWITCH_470152730,
+ TwitchROCODatasets.T470149568,
+ TwitchROCODatasets.T470150052,
+ TwitchROCODatasets.T470151286,
+ TwitchROCODatasets.T470152289,
+ TwitchROCODatasets.T470152730,
],
"Twitch_Train_",
)
TensorflowRecordFactory.from_datasets(
- [TwitchROCODatasets.TWITCH_470152838, TwitchROCODatasets.TWITCH_470153081, TwitchROCODatasets.TWITCH_470158483],
- "Twitch_Test_",
+ [TwitchROCODatasets.T470152838, TwitchROCODatasets.T470153081, TwitchROCODatasets.T470158483], "Twitch_Test_",
)
diff --git a/common/research/common/scripts/improve_roco_by_zooming.py b/common/research/common/scripts/improve_roco_by_zooming.py
index d38ddf1..ed7d40a 100644
--- a/common/research/common/scripts/improve_roco_by_zooming.py
+++ b/common/research/common/scripts/improve_roco_by_zooming.py
@@ -1,3 +1,4 @@
+from polystar.common.utils.tqdm import smart_tqdm
from research.common.dataset.improvement.zoom import Zoomer
from research.common.dataset.perturbations.image_modifiers.brightness import \
BrightnessModifier
@@ -9,7 +10,6 @@ from research.common.dataset.perturbations.perturbator import ImagePerturbator
from research.common.datasets.roco.roco_dataset import ROCOFileDataset
from research.common.datasets.roco.zoo.dji_zoomed import DJIROCOZoomedDatasets
from research.common.datasets.roco.zoo.roco_datasets_zoo import ROCODatasetsZoo
-from tqdm import tqdm
def improve_dji_roco_dataset_by_zooming_and_perturbating(
@@ -18,10 +18,10 @@ def improve_dji_roco_dataset_by_zooming_and_perturbating(
zoomed_dset = DJIROCOZoomedDatasets.make_dataset(dset.name)
zoomed_dset.create()
- for img, annotation in tqdm(dset.open(), desc=f"Processing {dset}", unit="image", total=len(dset)):
- for zoomed_image, zoomed_annotation in zoomer.zoom(img, annotation):
+ for img, annotation, name in smart_tqdm(dset.open(), desc=f"Processing {dset}", unit="image", total=len(dset)):
+ for zoomed_image, zoomed_annotation, zoomed_name in zoomer.zoom(img, annotation, name):
zoomed_image = perturbator.perturbate(zoomed_image)
- zoomed_dset.add(zoomed_image, zoomed_annotation)
+ zoomed_dset.add(zoomed_image, zoomed_annotation, zoomed_name)
def improve_all_dji_datasets_by_zooming_and_perturbating(zoomer: Zoomer, perturbator: ImagePerturbator):
diff --git a/common/research/common/scripts/visualize_dataset.py b/common/research/common/scripts/visualize_dataset.py
index 1bdb43a..eda5a19 100644
--- a/common/research/common/scripts/visualize_dataset.py
+++ b/common/research/common/scripts/visualize_dataset.py
@@ -6,7 +6,7 @@ from research.common.datasets.roco.zoo.roco_datasets_zoo import ROCODatasetsZoo
def visualize_dataset(dataset: ROCODataset, n_images: int):
viewer = PltResultViewer(dataset.name)
- for i, (image, annotation) in enumerate(dataset, 1):
+ for i, (image, annotation, name) in enumerate(dataset, 1):
viewer.display_image_with_objects(image, annotation.objects)
if i == n_images:
diff --git a/common/tests/common/unittests/datasets/test_dataset.py b/common/tests/common/unittests/datasets/test_dataset.py
index fff885d..cf34bba 100644
--- a/common/tests/common/unittests/datasets/test_dataset.py
+++ b/common/tests/common/unittests/datasets/test_dataset.py
@@ -11,23 +11,30 @@ class TestDataset(TestCase):
str_str_dataset: Dataset[str, str] = dataset.transform(str, str)
- self.assertEqual([("0", "8"), ("1", "9"), ("2", "10"), ("3", "11")], list(str_str_dataset))
+ self.assertEqual(
+ [("0", "8", "data_1"), ("1", "9", "data_2"), ("2", "10", "data_3"), ("3", "11", "data_4")],
+ list(str_str_dataset),
+ )
def test_transform_examples(self):
dataset = _make_fake_dataset()
str_int_dataset: Dataset[str, int] = dataset.transform_examples(str)
- self.assertEqual([("0", 8), ("1", 9), ("2", 10), ("3", 11)], list(str_int_dataset))
+ self.assertEqual(
+ [("0", 8, "data_1"), ("1", 9, "data_2"), ("2", 10, "data_3"), ("3", 11, "data_4")], list(str_int_dataset)
+ )
def test_transform_not_exhaustible(self):
dataset = _make_fake_dataset()
str_int_dataset: Dataset[str, float] = dataset.transform_examples(str)
- self.assertEqual([("0", 8), ("1", 9), ("2", 10), ("3", 11)], list(str_int_dataset))
- self.assertEqual([("0", 8), ("1", 9), ("2", 10), ("3", 11)], list(str_int_dataset))
- self.assertEqual([("0", 8), ("1", 9), ("2", 10), ("3", 11)], list(str_int_dataset))
+ items = [("0", 8, "data_1"), ("1", 9, "data_2"), ("2", 10, "data_3"), ("3", 11, "data_4")]
+
+ self.assertEqual(items, list(str_int_dataset))
+ self.assertEqual(items, list(str_int_dataset))
+ self.assertEqual(items, list(str_int_dataset))
class TestSimpleDataset(TestCase):
@@ -36,11 +43,12 @@ class TestSimpleDataset(TestCase):
self.assertEqual([0, 1, 2, 3], dataset.examples)
self.assertEqual([8, 9, 10, 11], dataset.targets)
+ self.assertEqual(["data_1", "data_2", "data_3", "data_4"], dataset.names)
def test_iter(self):
dataset = _make_fake_dataset()
- self.assertEqual([(0, 8), (1, 9), (2, 10), (3, 11)], list(dataset))
+ self.assertEqual([(0, 8, "data_1"), (1, 9, "data_2"), (2, 10, "data_3"), (3, 11, "data_4")], list(dataset))
def test_len(self):
dataset = _make_fake_dataset()
@@ -49,14 +57,14 @@ class TestSimpleDataset(TestCase):
def test_consistency(self):
with self.assertRaises(AssertionError):
- SimpleDataset([0, 1], [8, 9, 10, 11], "fake")
+ SimpleDataset([0, 1], [8, 9, 10, 11], ["a", "b"], "fake")
class FakeLazyDataset(LazyDataset):
def __init__(self):
super().__init__("fake")
- __iter__ = MagicMock(side_effect=lambda *args: iter([(1, 1), (2, 4), (3, 9)]))
+ __iter__ = MagicMock(side_effect=lambda *args: iter([(1, 1, "data_1"), (2, 4, "data_2"), (3, 9, "data_3")]))
class TestLazyDataset(TestCase):
@@ -65,7 +73,10 @@ class TestLazyDataset(TestCase):
self.assertEqual([1, 2, 3], list(dataset.examples))
self.assertEqual([1, 4, 9], list(dataset.targets))
- self.assertEqual([(1, 1), (2, 4), (3, 9)], list(zip(dataset.examples, dataset.targets)))
+ self.assertEqual(
+ [(1, 1, "data_1"), (2, 4, "data_2"), (3, 9, "data_3")],
+ list(zip(dataset.examples, dataset.targets, dataset.names)),
+ )
def test_properties_laziness(self):
FakeLazyDataset.__iter__.reset_mock()
@@ -80,6 +91,11 @@ class TestLazyDataset(TestCase):
list(zip(dataset.examples, dataset.targets))
FakeLazyDataset.__iter__.assert_called_once()
+ FakeLazyDataset.__iter__.reset_mock()
+
+ list(dataset.names)
+ FakeLazyDataset.__iter__.assert_not_called()
+
def _make_fake_dataset() -> Dataset[int, int]:
- return SimpleDataset([0, 1, 2, 3], [8, 9, 10, 11], "fake")
+ return SimpleDataset([0, 1, 2, 3], [8, 9, 10, 11], [f"data_{i}" for i in range(1, 5)], "fake")
--
GitLab