def evaluate(self):
if self._distributed:
comm.synchronize()
self._predictions = comm.gather(self._predictions, dst=0)
self._predictions = list(itertools.chain(*self._predictions))
if not comm.is_main_process():
return {}
if len(self._predictions) == 0:
self._logger.warning(
"[COCOEvaluator] Did not receive valid predictions.")
return {}
if self._output_dir:
PathManager.mkdirs(self._output_dir)
file_path = os.path.join(self._output_dir,
"instances_predictions.pth")
with PathManager.open(file_path, "wb") as f:
torch.save(self._predictions, f)
self._results = OrderedDict()
if "proposals" in self._predictions[0]:
self._eval_box_proposals()
if "instances" in self._predictions[0]:
self._eval_predictions(set(self._tasks))
# Copy so the caller can do whatever with results
return copy.deepcopy(self._results)
def load(self, path: str):
"""
Load from the given checkpoint. When path points to network file, this
function has to be called on all ranks.
Args:
path (str): path or url to the checkpoint. If empty, will not load
anything.
Returns:
dict:
extra data loaded from the checkpoint that has not been
processed. For example, those saved with
:meth:`.save(**extra_data)`.
"""
if not path:
# no checkpoint provided
self.logger.info(
"No checkpoint found. Initializing model from scratch")
return {}
self.logger.info("Loading checkpoint from {}".format(path))
if not os.path.isfile(path):
path = PathManager.get_local_path(path)
assert PathManager.isfile(path), "Checkpoint {} not found!".format(
path)
checkpoint = self._load_file(path)
self._load_model(checkpoint)
for key, obj in self.checkpointables.items():
if key in checkpoint:
self.logger.info("Loading {} from {}".format(key, path))
obj.load_state_dict(checkpoint.pop(key))
# return any further checkpoint data
return checkpoint
def convert_to_coco_json(dataset_name, output_folder="", allow_cached=True):
"""
Converts dataset into COCO format and saves it to a json file.
dataset_name must be registered in DatastCatalog and in detectron2's standard format.
Args:
dataset_name:
reference from the config file to the catalogs
must be registered in DatastCatalog and in detectron2's standard format
output_folder: where json file will be saved and loaded from
allow_cached: if json file is already present then skip conversion
Returns:
cache_path: path to the COCO-format json file
"""
# TODO: The dataset or the conversion script *may* change,
# a checksum would be useful for validating the cached data
cache_path = os.path.join(output_folder,
f"{dataset_name}_coco_format.json")
PathManager.mkdirs(output_folder)
if os.path.exists(cache_path) and allow_cached:
logger.info(
f"Reading cached annotations in COCO format from:{cache_path} ...")
else:
logger.info(
f"Converting dataset annotations in '{dataset_name}' to COCO format ...)"
)
coco_dict = convert_to_coco_dict(dataset_name)
with PathManager.open(cache_path, "w") as json_file:
logger.info(f"Caching annotations in COCO format: {cache_path}")
json.dump(coco_dict, json_file)
return cache_path
def get_all_checkpoint_files(self):
"""
Returns:
list: All available checkpoint files (.pth files) in target
directory.
"""
all_model_checkpoints = [
os.path.join(self.save_dir, file)
for file in PathManager.ls(self.save_dir)
if PathManager.isfile(os.path.join(self.save_dir, file))
and file.endswith(".pth")
]
return all_model_checkpoints
def default_setup(cfg, args):
"""
Perform some basic common setups at the beginning of a job, including:
1. Set up the dl_lib logger
2. Log basic information about environment, cmdline arguments, and config
3. Backup the config to the output directory
Args:
cfg (CfgNode): the full config to be used
args (argparse.NameSpace): the command line arguments to be logged
"""
output_dir = cfg.OUTPUT_DIR
if comm.is_main_process() and output_dir:
PathManager.mkdirs(output_dir)
rank = comm.get_rank()
# setup_logger(output_dir, distributed_rank=rank, name="dl_lib")
logger = setup_logger(output_dir, distributed_rank=rank)
logger.info("Rank of current process: {}. World size: {}".format(rank, comm.get_world_size()))
logger.info("Environment info:\n" + collect_env_info())
logger.info("Command line arguments: " + str(args))
if hasattr(args, "config_file") and args.config_file != "":
logger.info(
"Contents of args.config_file={}:\n{}".format(
args.config_file, PathManager.open(args.config_file, "r").read()
)
)
logger.info("Running with full config:\n{}".format(cfg))
base_config = cfg.__class__.__base__()
logger.info("different config with base class:\n{}".format(cfg.show_diff(base_config)))
# if comm.is_main_process() and output_dir:
# # Note: some of our scripts may expect the existence of
# # config.yaml in output directory
# path = os.path.join(output_dir, "config.yaml")
# with PathManager.open(path, "w") as f:
# f.write(cfg.dump())
# logger.info("Full config saved to {}".format(os.path.abspath(path)))
# make sure each worker has a different, yet deterministic seed if specified
seed_all_rng(None if cfg.SEED < 0 else cfg.SEED + rank)
# cudnn benchmark has large overhead. It shouldn't be used considering the small size of
# typical validation set.
if not (hasattr(args, "eval_only") and args.eval_only):
torch.backends.cudnn.benchmark = cfg.CUDNN_BENCHMARK
return cfg, logger
def _load_file(self, filename):
if filename.endswith(".pkl"):
with PathManager.open(filename, "rb") as f:
data = pickle.load(f, encoding="latin1")
if "model" in data and "__author__" in data:
# file is in dl_lib model zoo format
self.logger.info("Reading a file from '{}'".format(
data["__author__"]))
return data
else:
# assume file is from Caffe2 / Detectron1 model zoo
if "blobs" in data:
# Detection models have "blobs", but ImageNet models don't
data = data["blobs"]
data = {
k: v
for k, v in data.items() if not k.endswith("_momentum")
}
return {
"model": data,
"__author__": "Caffe2",
"matching_heuristics": True
}
loaded = super()._load_file(filename) # load native pth checkpoint
if "model" not in loaded:
loaded = {"model": loaded}
return loaded
def has_checkpoint(self):
"""
Returns:
bool: whether a checkpoint exists in the target directory.
"""
save_file = os.path.join(self.save_dir, "last_checkpoint")
return PathManager.exists(save_file)
def load_cityscapes_semantic(image_dir, gt_dir):
"""
Args:
image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train".
gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train".
Returns:
list[dict]: a list of dict, each has "file_name" and
"sem_seg_file_name".
"""
ret = []
for image_file in glob.glob(os.path.join(image_dir, "**/*.png")):
suffix = "leftImg8bit.png"
assert image_file.endswith(suffix)
prefix = image_dir
label_file = gt_dir + image_file[
len(prefix):-len(suffix)] + "gtFine_labelTrainIds.png"
assert os.path.isfile(
label_file
), "Please generate labelTrainIds.png with cityscapesscripts/preparation/createTrainIdLabelImgs.py" # noqa
json_file = gt_dir + image_file[
len(prefix):-len(suffix)] + "gtFine_polygons.json"
with PathManager.open(json_file, "r") as f:
jsonobj = json.load(f)
ret.append({
"file_name": image_file,
"sem_seg_file_name": label_file,
"height": jsonobj["imgHeight"],
"width": jsonobj["imgWidth"],
})
return ret
def __init__(self, json_file, window_size=20):
"""
Args:
json_file (str): path to the json file. New data will be appended if the file exists.
window_size (int): the window size of median smoothing for the scalars whose
`smoothing_hint` are True.
"""
self._file_handle = PathManager.open(json_file, "a")
self._window_size = window_size
def process(self, inputs, outputs):
"""
Args:
inputs: the inputs to a model.
It is a list of dicts. Each dict corresponds to an image and
contains keys like "height", "width", "file_name".
outputs: the outputs of a model. It is either list of semantic segmentation predictions
(Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic
segmentation prediction in the same format.
"""
for input, output in zip(inputs, outputs):
if "flops" in output:
flops = output["flops"]
self._real_flops.append(flops["real_flops"])
self._expt_flops.append(flops["expt_flops"])
output = output["sem_seg"].argmax(dim=0).to(self._cpu_device)
pred = np.array(output, dtype=np.int)
# Cityscapes test output
if 'cityscapes' in self._dataset_name and 'test' in self._dataset_name:
pred_converg = pred.copy()
f_name = input["file_name"].split('/')[-1]
pred_unique = list(np.unique(pred_converg))
pred_unique.sort(reverse=True)
for i in range(len(pred_unique)):
pred_converg[pred_converg == int(pred_unique[i])] \
= trainId2label[int(pred_unique[i])].id
save_dir = os.path.join(self._output_dir, 'test_dir')
if not os.path.exists(save_dir):
PathManager.mkdirs(save_dir)
cv2.imwrite(os.path.join(save_dir, f_name), pred_converg)
with PathManager.open(
self.input_file_to_gt_file[input["file_name"]], "rb") as f:
gt = np.array(Image.open(f), dtype=np.int)
gt[gt == self._ignore_label] = self._num_classes
self._conf_matrix += np.bincount(
self._N * pred.reshape(-1) + gt.reshape(-1),
minlength=self._N**2).reshape(self._N, self._N)
self._predictions.extend(
self.encode_json_sem_seg(pred, input["file_name"]))
def load_proposals_into_dataset(dataset_dicts, proposal_file):
r"""
Load precomputed object proposals into the dataset.
The proposal file should be a pickled dict with the following keys:
- "ids": list[int] or list[str], the image ids
- "boxes": list[np.ndarray], each is an Nx4 array of boxes corresponding to the image id
- "objectness_logits": list[np.ndarray], each is an N sized array of objectness scores
corresponding to the boxes.
- "bbox_mode": the BoxMode of the boxes array. Defaults to ``BoxMode.XYXY_ABS``.
Args:
dataset_dicts (list[dict]): annotations in dl_lib Dataset format.
proposal_file (str): file path of pre-computed proposals, in pkl format.
Returns:
list[dict]: the same format as dataset_dicts, but added proposal field.
"""
logger = logging.getLogger(__name__)
logger.info("Loading proposals from: {}".format(proposal_file))
with PathManager.open(proposal_file, "rb") as f:
proposals = pickle.load(f, encoding="latin1")
# Rename the key names in D1 proposal files
rename_keys = {"indexes": "ids", "scores": "objectness_logits"}
for key in rename_keys:
if key in proposals:
proposals[rename_keys[key]] = proposals.pop(key)
# Fetch the indexes of all proposals that are in the dataset
# Convert image_id to str since they could be int.
img_ids = set({str(record["image_id"]) for record in dataset_dicts})
id_to_index = {
str(id): i
for i, id in enumerate(proposals["ids"]) if str(id) in img_ids
}
# Assuming default bbox_mode of precomputed proposals are 'XYXY_ABS'
bbox_mode = BoxMode(proposals["bbox_mode"]
) if "bbox_mode" in proposals else BoxMode.XYXY_ABS
for record in dataset_dicts:
# Get the index of the proposal
i = id_to_index[str(record["image_id"])]
boxes = proposals["boxes"][i]
objectness_logits = proposals["objectness_logits"][i]
# Sort the proposals in descending order of the scores
inds = objectness_logits.argsort()[::-1]
record["proposal_boxes"] = boxes[inds]
record["proposal_objectness_logits"] = objectness_logits[inds]
record["proposal_bbox_mode"] = bbox_mode
return dataset_dicts
def tag_last_checkpoint(self, last_filename_basename: str):
"""
Tag the last checkpoint.
Args:
last_filename_basename (str): the basename of the last filename.
"""
save_file = os.path.join(self.save_dir, "last_checkpoint")
with PathManager.open(save_file, "w") as f:
f.write(last_filename_basename)
def _eval_predictions(self, tasks):
"""
Evaluate self._predictions on the given tasks.
Fill self._results with the metrics of the tasks.
"""
self._logger.info("Preparing results for COCO format ...")
self._coco_results = list(
itertools.chain(*[x["instances"] for x in self._predictions]))
# unmap the category ids for COCO
if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
reverse_id_mapping = {
v: k
for k, v in
self._metadata.thing_dataset_id_to_contiguous_id.items()
}
for result in self._coco_results:
category_id = result["category_id"]
assert (
category_id in reverse_id_mapping
), "A prediction has category_id={}, which is not available in the dataset.".format(
category_id)
result["category_id"] = reverse_id_mapping[category_id]
if self._output_dir:
file_path = os.path.join(self._output_dir,
"coco_instances_results.json")
self._logger.info("Saving results to {}".format(file_path))
with PathManager.open(file_path, "w") as f:
f.write(json.dumps(self._coco_results))
f.flush()
if not self._do_evaluation:
self._logger.info("Annotations are not available for evaluation.")
return
self._logger.info("Evaluating predictions ...")
for task in sorted(tasks):
coco_eval = (
_evaluate_predictions_on_coco(
self._coco_api,
self._coco_results,
task,
kpt_oks_sigmas=self._kpt_oks_sigmas)
if len(self._coco_results) > 0 else
None # cocoapi does not handle empty results very well
)
res = self._derive_coco_results(
coco_eval,
task,
class_names=self._metadata.get("thing_classes"))
self._results[task] = res
def load_voc_instances(dirname: str, split: str):
"""
Load Pascal VOC detection annotations to dl_lib format.
Args:
dirname: Contain "Annotations", "ImageSets", "JPEGImages"
split (str): one of "train", "test", "val", "trainval"
"""
with PathManager.open(
os.path.join(dirname, "ImageSets", "Main", split + ".txt")) as f:
fileids = np.loadtxt(f, dtype=np.str)
dicts = []
for fileid in fileids:
anno_file = os.path.join(dirname, "Annotations", fileid + ".xml")
jpeg_file = os.path.join(dirname, "JPEGImages", fileid + ".jpg")
tree = ET.parse(anno_file)
r = {
"file_name": jpeg_file,
"image_id": fileid,
"height": int(tree.findall("./size/height")[0].text),
"width": int(tree.findall("./size/width")[0].text),
}
instances = []
for obj in tree.findall("object"):
cls = obj.find("name").text
# We include "difficult" samples in training.
# Based on limited experiments, they don't hurt accuracy.
# difficult = int(obj.find("difficult").text)
# if difficult == 1:
# continue
bbox = obj.find("bndbox")
bbox = [
float(bbox.find(x).text)
for x in ["xmin", "ymin", "xmax", "ymax"]
]
# Original annotations are integers in the range [1, W or H]
# Assuming they mean 1-based pixel indices (inclusive),
# a box with annotation (xmin=1, xmax=W) covers the whole image.
# In coordinate space this is represented by (xmin=0, xmax=W)
bbox[0] -= 1.0
bbox[1] -= 1.0
instances.append({
"category_id": CLASS_NAMES.index(cls),
"bbox": bbox,
"bbox_mode": BoxMode.XYXY_ABS
})
r["annotations"] = instances
dicts.append(r)
return dicts
def get_checkpoint_file(self):
"""
Returns:
str: The latest checkpoint file in target directory.
"""
save_file = os.path.join(self.save_dir, "last_checkpoint")
try:
with PathManager.open(save_file, "r") as f:
last_saved = f.read().strip()
except IOError:
# if file doesn't exist, maybe because it has just been
# deleted by a separate process
return ""
return os.path.join(self.save_dir, last_saved)
def _eval_box_proposals(self):
"""
Evaluate the box proposals in self._predictions.
Fill self._results with the metrics for "box_proposals" task.
"""
if self._output_dir:
# Saving generated box proposals to file.
# Predicted box_proposals are in XYXY_ABS mode.
bbox_mode = BoxMode.XYXY_ABS.value
ids, boxes, objectness_logits = [], [], []
for prediction in self._predictions:
ids.append(prediction["image_id"])
boxes.append(
prediction["proposals"].proposal_boxes.tensor.numpy())
objectness_logits.append(
prediction["proposals"].objectness_logits.numpy())
proposal_data = {
"boxes": boxes,
"objectness_logits": objectness_logits,
"ids": ids,
"bbox_mode": bbox_mode,
}
with PathManager.open(
os.path.join(self._output_dir, "box_proposals.pkl"),
"wb") as f:
pickle.dump(proposal_data, f)
if not self._do_evaluation:
self._logger.info("Annotations are not available for evaluation.")
return
self._logger.info("Evaluating bbox proposals ...")
res = {}
areas = {"all": "", "small": "s", "medium": "m", "large": "l"}
for limit in [100, 1000]:
for area, suffix in areas.items():
stats = _evaluate_box_proposals(self._predictions,
self._coco_api,
area=area,
limit=limit)
key = "AR{}@{:d}".format(suffix, limit)
res[key] = float(stats["ar"].item() * 100)
self._logger.info("Proposal metrics: \n" + create_small_table(res))
self._results["box_proposals"] = res
def __init__(self,
dataset_name,
cfg,
distributed,
output_dir=None,
dump=False):
"""
Args:
dataset_name (str): name of the dataset to be evaluated.
It must have either the following corresponding metadata:
"json_file": the path to the COCO format annotation
Or it must be in detectron2's standard dataset format
so it can be converted to COCO format automatically.
cfg (CfgNode): config instance
distributed (True): if True, will collect results from all ranks for evaluation.
Otherwise, will evaluate the results in the current process.
output_dir (str): optional, an output directory to dump results.
"""
self._dump = dump
self._tasks = self._tasks_from_config(cfg)
self._distributed = distributed
self._output_dir = output_dir
self._cpu_device = torch.device("cpu")
self._logger = logging.getLogger(__name__)
self._metadata = MetadataCatalog.get(dataset_name)
if not hasattr(self._metadata, "json_file"):
self._logger.warning(
f"json_file was not found in MetaDataCatalog for '{dataset_name}'"
)
cache_path = convert_to_coco_json(dataset_name, output_dir)
self._metadata.json_file = cache_path
json_file = PathManager.get_local_path(self._metadata.json_file)
with contextlib.redirect_stdout(io.StringIO()):
self._coco_api = COCO(json_file)
self._kpt_oks_sigmas = cfg.TEST.KEYPOINT_OKS_SIGMAS
# Test set json files do not contain annotations (evaluation must be
# performed using the COCO evaluation server).
self._do_evaluation = "annotations" in self._coco_api.dataset
def after_step(self):
if self._profiler is None:
return
self._profiler.__exit__(None, None, None)
out_file = os.path.join(
self._output_dir,
"profiler-trace-iter{}.json".format(self.trainer.iter))
if "://" not in out_file:
self._profiler.export_chrome_trace(out_file)
else:
# Support non-posix filesystems
with tempfile.TemporaryDirectory(prefix="dl_lib_profiler") as d:
tmp_file = os.path.join(d, "tmp.json")
self._profiler.export_chrome_trace(tmp_file)
with open(tmp_file) as f:
content = f.read()
with PathManager.open(out_file, "w") as f:
f.write(content)
def read_image(file_name, format=None):
"""
Read an image into the given format.
Will apply rotation and flipping if the image has such exif information.
Args:
file_name (str): image file path
format (str): one of the supported image modes in PIL, or "BGR"
Returns:
image (np.ndarray): an HWC image
"""
with PathManager.open(file_name, "rb") as f:
image = Image.open(f)
image = ImageOps.exif_transpose(image)
# capture and ignore this bug: https://github.com/python-pillow/Pillow/issues/3973
try:
image = ImageOps.exif_transpose(image)
except Exception:
pass
if format is not None:
# PIL only supports RGB, so convert to RGB and flip channels over below
conversion_format = format
if format == "BGR":
conversion_format = "RGB"
image = image.convert(conversion_format)
image = np.asarray(image)
if format == "BGR":
# flip channels if needed
image = image[:, :, ::-1]
# PIL squeezes out the channel dimension for "L", so make it HWC
if format == "L":
image = np.expand_dims(image, -1)
return image
def save(self, name: str, **kwargs: dict):
"""
Dump model and checkpointables to a file.
Args:
name (str): name of the file.
kwargs (dict): extra arbitrary data to save.
"""
if not self.save_dir or not self.save_to_disk:
return
data = {}
data["model"] = self.model.state_dict()
for key, obj in self.checkpointables.items():
data[key] = obj.state_dict()
data.update(kwargs)
basename = "{}.pth".format(name)
save_file = os.path.join(self.save_dir, basename)
assert os.path.basename(save_file) == basename, basename
self.logger.info("Saving checkpoint to {}".format(save_file))
with PathManager.open(save_file, "wb") as f:
torch.save(data, f)
self.tag_last_checkpoint(basename)
def _get_local_path(self, path):
logger = logging.getLogger(__name__)
catalog_path = ModelCatalog.get(path[len(self.PREFIX):])
logger.info("Catalog entry {} points to {}".format(path, catalog_path))
return PathManager.get_local_path(catalog_path)
def setup_logger(output=None,
distributed_rank=0,
*,
color=True,
name="dl_lib",
abbrev_name=None):
"""
generate a logger with given setting
Args:
output (str): a file name or a directory to save log. If None, will not save log file.
If ends with ".txt" or ".log", assumed to be a file name.
Otherwise, logs will be saved to `output/log.txt`.
name (str): the root module name of this logger
abbrev_name (str): an abbreviation of the module, to avoid long names in logs.
Set to "" to not log the root module in logs.
By default, will abbreviate "dl_lib" to "d2" and leave other
modules unchanged.
"""
logger = logging.getLogger(name)
logger.setLevel(logging.DEBUG)
logger.propagate = False
if abbrev_name is None:
abbrev_name = "c2" if name == "dl_lib" else name
plain_formatter = logging.Formatter(
"[%(asctime)s] %(name)s %(levelname)s: %(message)s",
datefmt="%m/%d %H:%M:%S")
# stdout logging: master only
if distributed_rank == 0:
ch = logging.StreamHandler(stream=sys.stdout)
ch.setLevel(logging.DEBUG)
if color:
formatter = _ColorfulFormatter(
colored("[%(asctime)s %(name)s]: ", "green") + "%(message)s",
datefmt="%m/%d %H:%M:%S",
root_name=name,
abbrev_name=str(abbrev_name),
)
else:
formatter = plain_formatter
ch.setFormatter(formatter)
logger.addHandler(ch)
# file logging: all workers
if output is not None:
if output.endswith(".txt") or output.endswith(".log"):
filename = output
else:
filename = os.path.join(output, "log.txt")
if distributed_rank > 0:
filename = filename + ".rank{}".format(distributed_rank)
PathManager.mkdirs(os.path.dirname(filename))
fh = logging.StreamHandler(_cached_log_stream(filename))
fh.setLevel(logging.DEBUG)
fh.setFormatter(plain_formatter)
logger.addHandler(fh)
return logger
def _cached_log_stream(filename):
return PathManager.open(filename, "a")
def evaluate(self):
"""
Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval):
* Mean intersection-over-union averaged across classes (mIoU)
* Frequency Weighted IoU (fwIoU)
* Mean pixel accuracy averaged across classes (mACC)
* Pixel Accuracy (pACC)
"""
if self._distributed:
synchronize()
conf_matrix_list = all_gather(self._conf_matrix)
self._predictions = all_gather(self._predictions)
self._predictions = list(itertools.chain(*self._predictions))
self._real_flops = all_gather(self._real_flops)
self._real_flops = list(itertools.chain(*self._real_flops))
self._expt_flops = all_gather(self._expt_flops)
self._expt_flops = list(itertools.chain(*self._expt_flops))
if not is_main_process():
return
self._conf_matrix = np.zeros_like(self._conf_matrix)
for conf_matrix in conf_matrix_list:
self._conf_matrix += conf_matrix
if self._output_dir:
PathManager.mkdirs(self._output_dir)
file_path = os.path.join(self._output_dir,
"sem_seg_predictions.json")
with PathManager.open(file_path, "w") as f:
f.write(json.dumps(self._predictions))
acc = np.zeros(self._num_classes, dtype=np.float)
iou = np.zeros(self._num_classes, dtype=np.float)
tp = self._conf_matrix.diagonal()[:-1].astype(np.float)
pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float)
class_weights = pos_gt / np.sum(pos_gt)
pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float)
acc_valid = pos_gt > 0
acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid]
iou_valid = (pos_gt + pos_pred) > 0
union = pos_gt + pos_pred - tp
iou[acc_valid] = tp[acc_valid] / union[acc_valid]
macc = np.sum(acc) / np.sum(acc_valid)
miou = np.sum(iou) / np.sum(iou_valid)
fiou = np.sum(iou * class_weights)
pacc = np.sum(tp) / np.sum(pos_gt)
res = {}
res["mIoU"] = 100 * miou
res["fwIoU"] = 100 * fiou
res["mACC"] = 100 * macc
res["pACC"] = 100 * pacc
# add flops calculation
if len(self._real_flops) > 0 and len(self._expt_flops) > 0:
res["mean_real_flops"] = (sum(self._real_flops) /
len(self._real_flops)) / 1e3
res["max_real_flops"] = max(self._real_flops) / 1e3
res["min_real_flops"] = min(self._real_flops) / 1e3
res["mean_expt_flops"] = (sum(self._expt_flops) /
len(self._expt_flops)) / 1e3
res["max_expt_flops"] = max(self._expt_flops) / 1e3
res["min_expt_flops"] = min(self._expt_flops) / 1e3
if self._output_dir:
file_path = os.path.join(self._output_dir,
"sem_seg_evaluation.pth")
with PathManager.open(file_path, "wb") as f:
torch.save(res, f)
results = OrderedDict({"sem_seg": res})
self._logger.info(results)
return results
def __call__(self, dataset_dict):
"""
Args:
dataset_dict (dict): Metadata of one image, in dl_lib Dataset format.
Returns:
dict: a format that builtin models in dl_lib accept
"""
dataset_dict = copy.deepcopy(
dataset_dict) # it will be modified by code below
# USER: Write your own image loading if it's not from a file
image = utils.read_image(dataset_dict["file_name"],
format=self.img_format)
utils.check_image_size(dataset_dict, image)
if "annotations" not in dataset_dict:
image, transforms = T.apply_transform_gens(
([self.crop_gen] if self.crop_gen else []) + self.tfm_gens,
image)
else:
# Crop around an instance if there are instances in the image.
# USER: Remove if you don't use cropping
if self.crop_gen:
crop_tfm = utils.gen_crop_transform_with_instance(
self.crop_gen.get_crop_size(image.shape[:2]),
image.shape[:2],
np.random.choice(dataset_dict["annotations"]),
)
image = crop_tfm.apply_image(image)
image, transforms = T.apply_transform_gens(self.tfm_gens, image)
if self.crop_gen:
transforms = crop_tfm + transforms
image_shape = image.shape[:2] # h, w
# Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
# but not efficient on large generic data structures due to the use of pickle & mp.Queue.
# Therefore it's important to use torch.Tensor.
dataset_dict["image"] = torch.as_tensor(
image.transpose(2, 0, 1).astype("float32"))
# Can use uint8 if it turns out to be slow some day
# USER: Remove if you don't use pre-computed proposals.
if self.load_proposals:
utils.transform_proposals(dataset_dict, image_shape, transforms,
self.min_box_side_len,
self.proposal_topk)
if not self.is_train and not self.eval_with_gt:
dataset_dict.pop("annotations", None)
dataset_dict.pop("sem_seg_file_name", None)
return dataset_dict
if "annotations" in dataset_dict:
# USER: Modify this if you want to keep them for some reason.
for anno in dataset_dict["annotations"]:
if not self.mask_on:
anno.pop("segmentation", None)
if not self.keypoint_on:
anno.pop("keypoints", None)
# USER: Implement additional transformations if you have other types of data
annos = [
utils.transform_instance_annotations(
obj,
transforms,
image_shape,
keypoint_hflip_indices=self.keypoint_hflip_indices,
dot_number=self.dot_number)
for obj in dataset_dict.pop("annotations")
if obj.get("iscrowd", 0) == 0
]
instances = utils.annotations_to_instances(
annos, image_shape, mask_format=self.mask_format)
# Create a tight bounding box from masks, useful when image is cropped
if self.crop_gen and instances.has("gt_masks"):
instances.gt_boxes = instances.gt_masks.get_bounding_boxes()
dataset_dict["instances"] = utils.filter_empty_instances(instances)
# USER: Remove if you don't do semantic/panoptic segmentation.
if "sem_seg_file_name" in dataset_dict:
with PathManager.open(dataset_dict.pop("sem_seg_file_name"),
"rb") as f:
sem_seg_gt = Image.open(f)
sem_seg_gt = np.asarray(sem_seg_gt, dtype="uint8")
sem_seg_gt = transforms.apply_segmentation(sem_seg_gt)
sem_seg_gt = torch.as_tensor(sem_seg_gt.astype("long"))
dataset_dict["sem_seg"] = sem_seg_gt
return dataset_dict
def cityscapes_files_to_dict(files, from_json, to_polygons):
"""
Parse cityscapes annotation files to a dict.
Args:
files (tuple): consists of (image_file, instance_id_file, label_id_file, json_file)
from_json (bool): whether to read annotations from the raw json file or the png files.
to_polygons (bool): whether to represent the segmentation as polygons
(COCO's format) instead of masks (cityscapes's format).
Returns:
A dict in dl_lib Dataset format.
"""
from cityscapesscripts.helpers.labels import id2label, name2label
image_file, instance_id_file, _, json_file = files
annos = []
if from_json:
from shapely.geometry import MultiPolygon, Polygon
with PathManager.open(json_file, "r") as f:
jsonobj = json.load(f)
ret = {
"file_name": image_file,
"image_id": os.path.basename(image_file),
"height": jsonobj["imgHeight"],
"width": jsonobj["imgWidth"],
}
# `polygons_union` contains the union of all valid polygons.
polygons_union = Polygon()
# CityscapesScripts draw the polygons in sequential order
# and each polygon *overwrites* existing ones. See
# (https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/json2instanceImg.py) # noqa
# We use reverse order, and each polygon *avoids* early ones.
# This will resolve the ploygon overlaps in the same way as CityscapesScripts.
for obj in jsonobj["objects"][::-1]:
if "deleted" in obj: # cityscapes data format specific
continue
label_name = obj["label"]
try:
label = name2label[label_name]
except KeyError:
if label_name.endswith("group"): # crowd area
label = name2label[label_name[:-len("group")]]
else:
raise
if label.id < 0: # cityscapes data format
continue
# Cityscapes's raw annotations uses integer coordinates
# Therefore +0.5 here
poly_coord = np.asarray(obj["polygon"], dtype="f4") + 0.5
# CityscapesScript uses PIL.ImageDraw.polygon to rasterize
# polygons for evaluation. This function operates in integer space
# and draws each pixel whose center falls into the polygon.
# Therefore it draws a polygon which is 0.5 "fatter" in expectation.
# We therefore dilate the input polygon by 0.5 as our input.
poly = Polygon(poly_coord).buffer(0.5, resolution=4)
if not label.hasInstances or label.ignoreInEval:
# even if we won't store the polygon it still contributes to overlaps resolution
polygons_union = polygons_union.union(poly)
continue
# Take non-overlapping part of the polygon
poly_wo_overlaps = poly.difference(polygons_union)
if poly_wo_overlaps.is_empty:
continue
polygons_union = polygons_union.union(poly)
anno = {}
anno["iscrowd"] = label_name.endswith("group")
anno["category_id"] = label.id
if isinstance(poly_wo_overlaps, Polygon):
poly_list = [poly_wo_overlaps]
elif isinstance(poly_wo_overlaps, MultiPolygon):
poly_list = poly_wo_overlaps.geoms
else:
raise NotImplementedError(
"Unknown geometric structure {}".format(poly_wo_overlaps))
poly_coord = []
for poly_el in poly_list:
# COCO API can work only with exterior boundaries now, hence we store only them.
# TODO: store both exterior and interior boundaries once other parts of the
# codebase support holes in polygons.
poly_coord.append(list(chain(*poly_el.exterior.coords)))
anno["segmentation"] = poly_coord
(xmin, ymin, xmax, ymax) = poly_wo_overlaps.bounds
anno["bbox"] = (xmin, ymin, xmax, ymax)
anno["bbox_mode"] = BoxMode.XYXY_ABS
annos.append(anno)
else:
# See also the official annotation parsing scripts at
# https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/instances2dict.py # noqa
with PathManager.open(instance_id_file, "rb") as f:
inst_image = np.asarray(Image.open(f), order="F")
# ids < 24 are stuff labels (filtering them first is about 5% faster)
flattened_ids = np.unique(inst_image[inst_image >= 24])
ret = {
"file_name": image_file,
"image_id": os.path.basename(image_file),
"height": inst_image.shape[0],
"width": inst_image.shape[1],
}
for instance_id in flattened_ids:
# For non-crowd annotations, instance_id // 1000 is the label_id
# Crowd annotations have <1000 instance ids
label_id = instance_id // 1000 if instance_id >= 1000 else instance_id
label = id2label[label_id]
if not label.hasInstances or label.ignoreInEval:
continue
anno = {}
anno["iscrowd"] = instance_id < 1000
anno["category_id"] = label.id
mask = np.asarray(inst_image == instance_id,
dtype=np.uint8,
order="F")
inds = np.nonzero(mask)
ymin, ymax = inds[0].min(), inds[0].max()
xmin, xmax = inds[1].min(), inds[1].max()
anno["bbox"] = (xmin, ymin, xmax, ymax)
if xmax <= xmin or ymax <= ymin:
continue
anno["bbox_mode"] = BoxMode.XYXY_ABS
if to_polygons:
# This conversion comes from D4809743 and D5171122,
# when Mask-RCNN was first developed.
contours = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)[-2]
polygons = [
c.reshape(-1).tolist() for c in contours if len(c) >= 3
]
# opencv's can produce invalid polygons
if len(polygons) == 0:
continue
anno["segmentation"] = polygons
else:
anno["segmentation"] = mask_util.encode(mask[:, :, None])[0]
annos.append(anno)
ret["annotations"] = annos
return ret
def _open(self, path, mode="r", **kwargs):
return PathManager.open(self._get_local_path(path), mode, **kwargs)
def load_coco_json(json_file,
image_root,
dataset_name=None,
extra_annotation_keys=None):
"""
Load a json file with COCO's instances annotation format.
Currently supports instance detection, instance segmentation,
and person keypoints annotations.
Args:
json_file (str): full path to the json file in COCO instances annotation format.
image_root (str): the directory where the images in this json file exists.
dataset_name (str): the name of the dataset (e.g., coco_2017_train).
If provided, this function will also put "thing_classes" into
the metadata associated with this dataset.
extra_annotation_keys (list[str]): list of per-annotation keys that should also be
loaded into the dataset dict (besides "iscrowd", "bbox", "keypoints",
"category_id", "segmentation"). The values for these keys will be returned as-is.
For example, the densepose annotations are loaded in this way.
Returns:
list[dict]: a list of dicts in dl_lib standard format. (See
`Using Custom Datasets </tutorials/datasets.html>`_ )
Notes:
1. This function does not read the image files.
The results do not have the "image" field.
"""
from pycocotools.coco import COCO
timer = Timer()
json_file = PathManager.get_local_path(json_file)
with contextlib.redirect_stdout(io.StringIO()):
coco_api = COCO(json_file)
if timer.seconds() > 1:
logger.info("Loading {} takes {:.2f} seconds.".format(
json_file, timer.seconds()))
id_map = None
if dataset_name is not None:
meta = MetadataCatalog.get(dataset_name)
cat_ids = sorted(coco_api.getCatIds())
cats = coco_api.loadCats(cat_ids)
# The categories in a custom json file may not be sorted.
thing_classes = [
c["name"] for c in sorted(cats, key=lambda x: x["id"])
]
meta.thing_classes = thing_classes
# In COCO, certain category ids are artificially removed,
# and by convention they are always ignored.
# We deal with COCO's id issue and translate
# the category ids to contiguous ids in [0, 80).
# It works by looking at the "categories" field in the json, therefore
# if users' own json also have incontiguous ids, we'll
# apply this mapping as well but print a warning.
if not (min(cat_ids) == 1 and max(cat_ids) == len(cat_ids)):
if "coco" not in dataset_name:
logger.warning("""
Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.
""")
id_map = {v: i for i, v in enumerate(cat_ids)}
meta.thing_dataset_id_to_contiguous_id = id_map
# sort indices for reproducible results
img_ids = sorted(list(coco_api.imgs.keys()))
# imgs is a list of dicts, each looks something like:
# {'license': 4,
# 'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
# 'file_name': 'COCO_val2014_000000001268.jpg',
# 'height': 427,
# 'width': 640,
# 'date_captured': '2013-11-17 05:57:24',
# 'id': 1268}
imgs = coco_api.loadImgs(img_ids)
# anns is a list[list[dict]], where each dict is an annotation
# record for an object. The inner list enumerates the objects in an image
# and the outer list enumerates over images. Example of anns[0]:
# [{'segmentation': [[192.81,
# 247.09,
# ...
# 219.03,
# 249.06]],
# 'area': 1035.749,
# 'iscrowd': 0,
# 'image_id': 1268,
# 'bbox': [192.81, 224.8, 74.73, 33.43],
# 'category_id': 16,
# 'id': 42986},
# ...]
anns = [coco_api.imgToAnns[img_id] for img_id in img_ids]
if "minival" not in json_file:
# The popular valminusminival & minival annotations for COCO2014 contain this bug.
# However the ratio of buggy annotations there is tiny and does not affect accuracy.
# Therefore wce explicitly white-list them.
ann_ids = [
ann["id"] for anns_per_image in anns for ann in anns_per_image
]
assert len(set(ann_ids)) == len(
ann_ids), "Annotation ids in '{}' are not unique!".format(
json_file)
imgs_anns = list(zip(imgs, anns))
logger.info("Loaded {} images in COCO format from {}".format(
len(imgs_anns), json_file))
dataset_dicts = []
ann_keys = ["iscrowd", "bbox", "keypoints", "category_id"
] + (extra_annotation_keys or [])
num_instances_without_valid_segmentation = 0
for (img_dict, anno_dict_list) in imgs_anns:
record = {}
record["file_name"] = os.path.join(image_root, img_dict["file_name"])
record["height"] = img_dict["height"]
record["width"] = img_dict["width"]
image_id = record["image_id"] = img_dict["id"]
objs = []
for anno in anno_dict_list:
# Check that the image_id in this annotation is the same as
# the image_id we're looking at.
# This fails only when the data parsing logic or the annotation file is buggy.
# The original COCO valminusminival2014 & minival2014 annotation files
# actually contains bugs that, together with certain ways of using COCO API,
# can trigger this assertion.
assert anno["image_id"] == image_id
assert anno.get("ignore", 0) == 0
obj = {key: anno[key] for key in ann_keys if key in anno}
segm = anno.get("segmentation", None)
if segm: # either list[list[float]] or dict(RLE)
if not isinstance(segm, dict):
# filter out invalid polygons (< 3 points)
segm = [
poly for poly in segm
if len(poly) % 2 == 0 and len(poly) >= 6
]
if len(segm) == 0:
num_instances_without_valid_segmentation += 1
continue # ignore this instance
obj["segmentation"] = segm
keypts = anno.get("keypoints", None)
if keypts: # list[int]
for idx, v in enumerate(keypts):
if idx % 3 != 2:
# COCO's segmentation coordinates are floating points in [0, H or W],
# but keypoint coordinates are integers in [0, H-1 or W-1]
# Therefore we assume the coordinates are "pixel indices" and
# add 0.5 to convert to floating point coordinates.
keypts[idx] = v + 0.5
obj["keypoints"] = keypts
obj["bbox_mode"] = BoxMode.XYWH_ABS
if id_map:
obj["category_id"] = id_map[obj["category_id"]]
objs.append(obj)
record["annotations"] = objs
dataset_dicts.append(record)
if num_instances_without_valid_segmentation > 0:
logger.warn(
"Filtered out {} instances without valid segmentation. "
"There might be issues in your dataset generation process.".format(
num_instances_without_valid_segmentation))
return dataset_dicts
def load_sem_seg(gt_root, image_root, gt_ext="png", image_ext="jpg"):
"""
Load semantic segmentation datasets. All files under "gt_root" with "gt_ext" extension are
treated as ground truth annotations and all files under "image_root" with "image_ext" extension
as input images. Ground truth and input images are matched using file paths relative to
"gt_root" and "image_root" respectively without taking into account file extensions.
This works for COCO as well as some other datasets.
Args:
gt_root (str): full path to ground truth semantic segmentation files. Semantic segmentation
annotations are stored as images with integer values in pixels that represent
corresponding semantic labels.
image_root (str): the directory where the input images are.
gt_ext (str): file extension for ground truth annotations.
image_ext (str): file extension for input images.
Returns:
list[dict]:
a list of dicts in dl_lib standard format without instance-level
annotation.
Notes:
1. This function does not read the image and ground truth files.
The results do not have the "image" and "sem_seg" fields.
"""
# We match input images with ground truth based on their relative filepaths (without file
# extensions) starting from 'image_root' and 'gt_root' respectively.
def file2id(folder_path, file_path):
# extract relative path starting from `folder_path`
image_id = os.path.normpath(
os.path.relpath(file_path, start=folder_path))
# remove file extension
image_id = os.path.splitext(image_id)[0]
return image_id
input_files = sorted(
(os.path.join(image_root, f)
for f in PathManager.ls(image_root) if f.endswith(image_ext)),
key=lambda file_path: file2id(image_root, file_path),
)
gt_files = sorted(
(os.path.join(gt_root, f)
for f in PathManager.ls(gt_root) if f.endswith(gt_ext)),
key=lambda file_path: file2id(gt_root, file_path),
)
assert len(gt_files) > 0, "No annotations found in {}.".format(gt_root)
# Use the intersection, so that val2017_100 annotations can run smoothly with val2017 images
if len(input_files) != len(gt_files):
logger.warn(
"Directory {} and {} has {} and {} files, respectively.".format(
image_root, gt_root, len(input_files), len(gt_files)))
input_basenames = [
os.path.basename(f)[:-len(image_ext)] for f in input_files
]
gt_basenames = [os.path.basename(f)[:-len(gt_ext)] for f in gt_files]
intersect = list(set(input_basenames) & set(gt_basenames))
# sort, otherwise each worker may obtain a list[dict] in different order
intersect = sorted(intersect)
logger.warn("Will use their intersection of {} files.".format(
len(intersect)))
input_files = [
os.path.join(image_root, f + image_ext) for f in intersect
]
gt_files = [os.path.join(gt_root, f + gt_ext) for f in intersect]
logger.info("Loaded {} images with semantic segmentation from {}".format(
len(input_files), image_root))
dataset_dicts = []
for (img_path, gt_path) in zip(input_files, gt_files):
local_path = PathManager.get_local_path(gt_path)
w, h = imagesize.get(local_path)
record = {}
record["file_name"] = img_path
record["sem_seg_file_name"] = gt_path
record["height"] = h
record["width"] = w
dataset_dicts.append(record)
return dataset_dicts
def _get_local_path(self, path):
name = path[len(self.PREFIX):]
return PathManager.get_local_path(self.S3_DETECTRON2_PREFIX + name)