def __getitem__(self, idx):
"""
- We do delayed loading of data to reduce the memory size due to pickling of
dataset across dataloader workers.
- Loads the data if not already loaded.
- Sets and initializes the queue if not already initialized
- Depending on the data source (folder or filelist), get the image.
If using the QueueDataset and image is valid, save the image in queue if
not full. Otherwise return a valid seen image from the queue if queue is
not empty.
"""
if not self.is_initialized:
self._load_data(self._path)
self.is_initialized = True
if not self.queue_init and self.enable_queue_dataset:
self._init_queues()
is_success = True
image_path = self.image_dataset[idx]
try:
if self.data_source == "disk_filelist":
image_path = self._replace_img_path_prefix(
image_path,
replace_prefix=self._remove_prefix,
new_prefix=self._new_prefix,
)
with PathManager.open(image_path, "rb") as fopen:
img = Image.open(fopen).convert("RGB")
elif self.data_source == "disk_folder":
img = self.image_dataset[idx][0]
if is_success and self.enable_queue_dataset:
self.on_sucess(img)
except Exception as e:
logging.warning(
f"Couldn't load: {self.image_dataset[idx]}. Exception: \n{e}")
is_success = False
# if we have queue dataset class enabled, we try to use it to get
# the seen valid images
if self.enable_queue_dataset:
img, is_success = self.on_failure()
if img is None:
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
else:
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
return img, is_success
def __getitem__(self, idx):
"""
Simply return the mean dummy image of the specified size and mark
it as a success.
"""
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
is_success = True
return img, is_success
def __getitem__(self, idx: int):
"""
Simply return the mean dummy image of the specified size and mark
it as a success.
"""
crop_size = self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE
if self.cfg["DATA"][self.split].RANDOM_SYNTHETIC_IMAGES:
img = self.generate_image(seed=idx, crop_size=crop_size)
else:
img = get_mean_image(crop_size)
is_success = True
return img, is_success
def __getitem__(self, index) -> Tuple[Image.Image, bool]:
if self._iterator is None:
self._iterator = self._open_iterator()
if not self.queue_init and self.enable_queue_dataset:
self._init_queues()
try:
# TODO (wpc, prigoyal): we should check images are good when we are
# uploading them to airstore.
ImageFile.LOAD_TRUNCATED_IMAGES = True
image_bytes = next(self._iterator)["image"]
img = Image.open(io.BytesIO(image_bytes))
if img.mode != "RGB":
img = img.convert("RGB")
if self.enable_queue_dataset:
self.on_sucess(img)
is_success = True
except Exception as e:
# TODO: airstore should have no failed images
# because they are filtered at prepare time.
# Then, this should be removed.
logging.warning(e)
is_success = False
# if we have queue dataset class enabled, we try to use it to get
# the seen valid images
if self.enable_queue_dataset:
img, is_success = self.on_failure()
if img is None:
raise RuntimeError(
"Encountered invalid image and couldn't load from QueueDataset"
)
else:
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
return img, is_success
def __getitem__(self, idx):
"""
- We do delayed loading of data to reduce the memory size due to pickling of
dataset across dataloader workers.
- Loads the data if not already loaded.
- Sets and initializes the queue if not already initialized
- Depending on the data source (folder or filelist), get the image.
If using the QueueDataset and image is valid, save the image in queue if
not full. Otherwise return a valid seen image from the queue if queue is
not empty.
"""
if not self.is_initialized:
self._load_data(self._path)
self.is_initialized = True
if not self.queue_init and self.enable_queue_dataset:
self._init_queues()
is_success = True
image_path = self.image_dataset[idx]
try:
if self.data_source == "disk_filelist":
image_path = self._replace_img_path_prefix(
image_path,
replace_prefix=self._remove_prefix,
new_prefix=self._new_prefix,
)
with g_pathmgr.open(image_path, "rb") as fopen:
img = Image.open(fopen).convert("RGB")
elif self.data_source == "disk_folder":
img = self.image_dataset[idx][0]
elif self.data_source == "disk_roi_annotations":
with g_pathmgr.open(image_path, "rb") as fopen:
bbox = [float(item) for item in self.image_roi_bbox[idx]]
img = Image.open(fopen).crop(bbox).convert("RGB")
# TODO: move the below to a dedicated transform
# applicable to openimages dataset only
width, height = img.size
bbox_size = min(img.size)
ratio = max(img.size) / min(img.size)
if ratio >= 1.2:
if width < height: # bigger height
bbox = (0, 0, bbox_size, bbox_size)
else: # bigger width.
bbox = (
int((width - bbox_size) / 2),
0,
int((width - bbox_size) / 2) + bbox_size,
height,
)
img = img.crop(bbox)
if is_success and self.enable_queue_dataset:
self.on_sucess(img)
except Exception as e:
logging.warning(
f"Couldn't load: {self.image_dataset[idx]}. Exception: \n{e}")
is_success = False
# if we have queue dataset class enabled, we try to use it to get
# the seen valid images
if self.enable_queue_dataset:
img, is_success = self.on_failure()
if img is None:
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
else:
img = get_mean_image(
self.cfg["DATA"][self.split].DEFAULT_GRAY_IMG_SIZE)
return img, is_success