def evaluate(self,
results,
metric='bbox',
logger=None,
jsonfile_prefix=None,
classwise=False,
proposal_nums=(100, 300, 1000),
iou_thrs=np.arange(0.5, 0.96, 0.05)):
"""Evaluation in LVIS protocol.
Args:
results (list[list | tuple]): Testing results of the dataset.
metric (str | list[str]): Metrics to be evaluated. Options are
'bbox', 'segm', 'proposal', 'proposal_fast'.
logger (logging.Logger | str | None): Logger used for printing
related information during evaluation. Default: None.
jsonfile_prefix (str | None):
classwise (bool): Whether to evaluating the AP for each class.
proposal_nums (Sequence[int]): Proposal number used for evaluating
recalls, such as recall@100, recall@1000.
Default: (100, 300, 1000).
iou_thrs (Sequence[float]): IoU threshold used for evaluating
recalls. If set to a list, the average recall of all IoUs will
also be computed. Default: 0.5.
Returns:
dict[str, float]: LVIS style metrics.
"""
try:
from lvis import LVISResults, LVISEval
except ImportError:
raise ImportError('Please follow config/lvis/README.md to '
'install open-mmlab forked lvis first.')
assert isinstance(results, list), 'results must be a list'
assert len(results) == len(self), (
'The length of results is not equal to the dataset len: {} != {}'.
format(len(results), len(self)))
metrics = metric if isinstance(metric, list) else [metric]
allowed_metrics = ['bbox', 'segm', 'proposal', 'proposal_fast']
for metric in metrics:
if metric not in allowed_metrics:
raise KeyError('metric {} is not supported'.format(metric))
if jsonfile_prefix is None:
tmp_dir = tempfile.TemporaryDirectory()
jsonfile_prefix = osp.join(tmp_dir.name, 'results')
else:
tmp_dir = None
result_files = self.results2json(results, jsonfile_prefix)
eval_results = {}
# get original api
lvis_gt = self.coco
for metric in metrics:
msg = 'Evaluating {}...'.format(metric)
if logger is None:
msg = '\n' + msg
print_log(msg, logger=logger)
if metric == 'proposal_fast':
ar = self.fast_eval_recall(results,
proposal_nums,
iou_thrs,
logger='silent')
log_msg = []
for i, num in enumerate(proposal_nums):
eval_results['AR@{}'.format(num)] = ar[i]
log_msg.append('\nAR@{}\t{:.4f}'.format(num, ar[i]))
log_msg = ''.join(log_msg)
print_log(log_msg, logger=logger)
continue
if metric not in result_files:
raise KeyError('{} is not in results'.format(metric))
try:
lvis_dt = LVISResults(lvis_gt, result_files[metric])
except IndexError:
print_log('The testing results of the whole dataset is empty.',
logger=logger,
level=logging.ERROR)
break
iou_type = 'bbox' if metric == 'proposal' else metric
lvis_eval = LVISEval(lvis_gt, lvis_dt, iou_type)
lvis_eval.params.imgIds = self.img_ids
if metric == 'proposal':
lvis_eval.params.useCats = 0
lvis_eval.params.maxDets = list(proposal_nums)
lvis_eval.evaluate()
lvis_eval.accumulate()
lvis_eval.summarize()
for k, v in lvis_eval.get_results().items():
if k.startswith('AR'):
val = float('{:.3f}'.format(float(v)))
eval_results[k] = val
else:
lvis_eval.evaluate()
lvis_eval.accumulate()
lvis_eval.summarize()
lvis_results = lvis_eval.get_results()
if classwise: # Compute per-category AP
# Compute per-category AP
# from https://github.com/facebookresearch/detectron2/
precisions = lvis_eval.eval['precision']
# precision: (iou, recall, cls, area range, max dets)
assert len(self.cat_ids) == precisions.shape[2]
results_per_category = []
for idx, catId in enumerate(self.cat_ids):
# area range index 0: all area ranges
# max dets index -1: typically 100 per image
nm = self.coco.load_cats(catId)[0]
precision = precisions[:, :, idx, 0, -1]
precision = precision[precision > -1]
if precision.size:
ap = np.mean(precision)
else:
ap = float('nan')
results_per_category.append(
(f'{nm["name"]}', f'{float(ap):0.3f}'))
num_columns = min(6, len(results_per_category) * 2)
results_flatten = list(
itertools.chain(*results_per_category))
headers = ['category', 'AP'] * (num_columns // 2)
results_2d = itertools.zip_longest(*[
results_flatten[i::num_columns]
for i in range(num_columns)
])
table_data = [headers]
table_data += [result for result in results_2d]
table = AsciiTable(table_data)
print_log('\n' + table.table, logger=logger)
for k, v in lvis_results.items():
if k.startswith('AP'):
key = '{}_{}'.format(metric, k)
val = float('{:.3f}'.format(float(v)))
eval_results[key] = val
ap_summary = ' '.join([
'{}:{:.3f}'.format(k, float(v))
for k, v in lvis_results.items() if k.startswith('AP')
])
eval_results['{}_mAP_copypaste'.format(metric)] = ap_summary
lvis_eval.print_results()
if tmp_dir is not None:
tmp_dir.cleanup()
return eval_results
def main():
# Use first line of file docstring as description if it exists.
parser = argparse.ArgumentParser(
description=__doc__.split("\n")[0] if __doc__ else "",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("annotations_json", type=Path)
parser.add_argument("results_json", type=Path)
parser.add_argument("output_dir", type=Path)
parser.add_argument("--type", default="segm", choices=["segm", "bbox"])
parser.add_argument("--dets-per-cat", default=-1, type=int)
parser.add_argument("--max-dets", default=-1, type=int)
parser.add_argument("--ious", nargs="*", type=float)
# NOTE: We default to only using areas=all, since we don't report S/M/L APs here.
parser.add_argument(
"--areas",
nargs="*",
type=str,
choices=["all", "small", "medium", "large"],
default=["all"],
)
parser.add_argument(
"--pools",
nargs="*",
choices=["all", "r", "c", "f"],
default=["all", "r", "c", "f"],
)
args = parser.parse_args()
args.output_dir.mkdir(exist_ok=True, parents=True)
logger = setup_logger(output=str(args.output_dir.resolve()), name=__file__)
log_path = args.output_dir / "log.txt"
results = str(args.results_json)
if args.dets_per_cat > 0:
with open(args.results_json, "r") as f:
results = json.load(f)
by_cat = defaultdict(list)
for ann in results:
by_cat[ann["category_id"]].append(ann)
results = []
topk = args.dets_per_cat
for cat_anns in by_cat.values():
results.extend(
sorted(cat_anns, key=lambda x: x["score"],
reverse=True)[:topk])
gt = LVIS(args.annotations_json)
results = LVISResults(gt, results, max_dets=args.max_dets)
lvis_eval = LVISPooledEval(gt,
results,
iou_type=args.type,
pools=args.pools)
params = lvis_eval.params
params.max_dets = args.max_dets
if args.ious:
params.iou_thrs = args.ious
if args.areas:
indices = [
i for i, area in enumerate(params.area_rng_lbl)
if area in args.areas
]
params.area_rng_lbl = [params.area_rng_lbl[i] for i in indices]
params.area_rng = [params.area_rng[i] for i in indices]
lvis_eval.run()
lvis_eval.print_results()
metrics = {
k: v
for k, v in lvis_eval.results.items() if k.startswith("AP")
}
logger.info("copypaste: %s,%s", ",".join(map(str, metrics.keys())), "path")
logger.info(
"copypaste: %s,%s",
",".join(f"{v*100:.2f}" for v in metrics.values()),
log_path,
)
def reload_annotations(self):
self.lvis = LVIS('/scratch/users/zzweng/datasets/lvis/lvis_v0.5_val.json')
self.dt_path = 'output/inference/lvis_instances_results.json'
self.lvis_dt = LVISResults(self.lvis, self.dt_path)
def evaluate(self,
results,
metric=['track'],
logger=None,
resfile_path=None):
if isinstance(metric, list):
metrics = metric
elif isinstance(metric, str):
metrics = [metric]
else:
raise TypeError('metric must be a list or a str.')
allowed_metrics = ['bbox', 'track']
for metric in metrics:
if metric not in allowed_metrics:
raise KeyError(f'metric {metric} is not supported.')
result_files, tmp_dir = self.format_results(results, resfile_path)
eval_results = dict()
if 'track' in metrics:
from tao.toolkit.tao import TaoEval
print_log('Evaluating TAO results...', logger)
tao_eval = TaoEval(self.ann_file, result_files['track'])
tao_eval.params.img_ids = self.img_ids
tao_eval.params.cat_ids = self.cat_ids
tao_eval.params.iou_thrs = np.array([0.5, 0.75])
tao_eval.run()
tao_eval.print_results()
tao_results = tao_eval.get_results()
for k, v in tao_results.items():
if isinstance(k, str) and k.startswith('AP'):
key = 'track_{}'.format(k)
val = float('{:.3f}'.format(float(v)))
eval_results[key] = val
if 'bbox' in metrics:
try:
import lvis
assert lvis.__version__ >= '10.5.3'
from lvis import LVIS, LVISResults, LVISEval
except AssertionError:
raise AssertionError(
'Incompatible version of lvis is installed. '
'Run pip uninstall lvis first. Then run pip '
'install mmlvis to install open-mmlab forked '
'lvis. ')
except ImportError:
raise ImportError(
'Package lvis is not installed. Please run pip '
'install mmlvis to install open-mmlab forked '
'lvis.')
print_log('Evaluating detection results...', logger)
lvis_gt = LVIS(self.ann_file)
lvis_dt = LVISResults(lvis_gt, result_files['bbox'])
lvis_eval = LVISEval(lvis_gt, lvis_dt, 'bbox')
lvis_eval.params.imgIds = self.img_ids
lvis_eval.params.catIds = self.cat_ids
lvis_eval.evaluate()
lvis_eval.accumulate()
lvis_eval.summarize()
lvis_eval.print_results()
lvis_results = lvis_eval.get_results()
for k, v in lvis_results.items():
if k.startswith('AP'):
key = '{}_{}'.format('bbox', k)
val = float('{:.3f}'.format(float(v)))
eval_results[key] = val
ap_summary = ' '.join([
'{}:{:.3f}'.format(k, float(v))
for k, v in lvis_results.items() if k.startswith('AP')
])
eval_results['bbox_mAP_copypaste'] = ap_summary
if tmp_dir is not None:
tmp_dir.cleanup()
return eval_results
def _evaluate_predictions_on_lvis_per_class(lvis_gt,
lvis_results,
iou_type,
class_names=None):
"""
Args:
iou_type (str):
kpt_oks_sigmas (list[float]):
class_names (None or list[str]): if provided, will use it to predict
per-category AP.
Returns:
a dict of {metric name: score}
"""
metrics = {
"bbox":
["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
"segm":
["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
}[iou_type]
logger = logging.getLogger(__name__)
if len(lvis_results) == 0: # TODO: check if needed
logger.warn("No predictions from the model! Set scores to -1")
return {metric: -1 for metric in metrics}
if iou_type == "segm":
lvis_results = copy.deepcopy(lvis_results)
# When evaluating mask AP, if the results contain bbox, LVIS API will
# use the box area as the area of the instance, instead of the mask area.
# This leads to a different definition of small/medium/large.
# We remove the bbox field to let mask AP use mask area.
for c in lvis_results:
c.pop("bbox", None)
from lvis import LVISEval, LVISResults
lvis_results = LVISResults(lvis_gt, lvis_results)
lvis_eval = LVISEval(lvis_gt, lvis_results, iou_type)
lvis_eval.evaluate()
lvis_eval.accumulate()
precisions = lvis_eval.eval["precision"]
results_per_category = []
for idx, name in enumerate(class_names):
# area range index 0: all area ranges
# max dets index -1: typically 100 per image
precision = precisions[:, :, idx, 0]
precision = precision[precision > -1]
ap = np.mean(precision) if precision.size else float("nan")
results_per_category.append(("{}".format(name), float(ap * 100)))
lvis_eval.summarize()
lvis_eval.print_results()
# Pull the standard metrics from the LVIS results
results = lvis_eval.get_results()
results = {metric: float(results[metric] * 100) for metric in metrics}
logger.info("Evaluation results for {}: \n".format(iou_type) +
create_small_table(results))
results.update({"AP-" + name: ap for name, ap in results_per_category})
return results
def main():
# Use first line of file docstring as description if it exists.
parser = argparse.ArgumentParser(
description=__doc__.split("\n")[0] if __doc__ else "",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("annotations_json", type=Path)
parser.add_argument("results_json", type=Path)
parser.add_argument("output_dir", type=Path)
parser.add_argument("--type", default="segm", choices=["segm", "bbox"])
parser.add_argument("--dets-per-cat", default=10000, type=int)
parser.add_argument("--ious", nargs="*", type=float)
args = parser.parse_args()
args.output_dir.mkdir(exist_ok=True, parents=True)
logger = setup_logger(output=str(args.output_dir.resolve()), name=__file__)
log_path = args.output_dir / "log.txt"
assert args.dets_per_cat > 0
with open(args.results_json, "r") as f:
results = json.load(f)
by_cat = defaultdict(list)
for ann in results:
by_cat[ann["category_id"]].append(ann)
results = []
topk = args.dets_per_cat
missing_dets_cats = set()
for cat, cat_anns in by_cat.items():
if len(cat_anns) < topk:
missing_dets_cats.add(cat)
results.extend(
sorted(cat_anns, key=lambda x: x["score"], reverse=True)[:topk])
if missing_dets_cats:
logger.warning(
f"\n===\n"
f"{len(missing_dets_cats)} classes had less than {topk} detections!\n"
f"Outputting {topk} detections for each class will improve AP further.\n"
f"If using detectron2, please use the lvdevil/infer_topk.py script to "
f"output a results file with {topk} detections for each class.\n"
f"===")
gt = LVIS(args.annotations_json)
results = LVISResults(gt, results, max_dets=-1)
lvis_eval = LVISEval(gt, results, iou_type=args.type)
params = lvis_eval.params
params.max_dets = -1 # No limit on detections per image.
if args.ious:
params.iou_thrs = args.ious
lvis_eval.run()
lvis_eval.print_results()
metrics = {
k: v
for k, v in lvis_eval.results.items() if k.startswith("AP")
}
logger.info("copypaste: %s,%s", ",".join(map(str, metrics.keys())), "path")
logger.info(
"copypaste: %s,%s",
",".join(f"{v*100:.2f}" for v in metrics.values()),
log_path,
)
def eval_cocofied_lvis_result(self, gt_file, result_file, metric='segm'):
def get_lvis_format_result(lvis_params, lvis_results):
template = " {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} catIds={:>3s}] = {:0.3f}"
result_list = []
for key, value in lvis_results.items():
max_dets = lvis_params.max_dets
if "AP" in key:
title = "Average Precision"
_type = "(AP)"
else:
title = "Average Recall"
_type = "(AR)"
if len(key) > 2 and key[2].isdigit():
iou_thr = (float(key[2:]) / 100)
iou = "{:0.2f}".format(iou_thr)
else:
iou = "{:0.2f}:{:0.2f}".format(
lvis_params.iou_thrs[0], lvis_params.iou_thrs[-1]
)
if len(key) > 2 and key[2] in ["r", "c", "f"]:
cat_group_name = key[2]
else:
cat_group_name = "all"
if len(key) > 2 and key[2] in ["s", "m", "l"]:
area_rng = key[2]
else:
area_rng = "all"
result_list.append(template.format(title, _type, iou, area_rng, max_dets, cat_group_name, value))
return result_list
print('load gt json')
lvis_gt = LVIS(gt_file)
cat_ids = lvis_gt.get_cat_ids()
print('load pred json')
lvis_dt = LVISResults(lvis_gt, result_file)
print('evaluating')
lvis_eval = LVISEval(lvis_gt, lvis_dt, metric)
lvis_eval.params.imgIds = lvis_gt.get_img_ids()
lvis_eval.evaluate()
lvis_eval.accumulate()
lvis_eval.summarize()
# Compute per-category AP
precisions = lvis_eval.eval['precision']
assert len(cat_ids) == precisions.shape[2]
results_per_category = []
for idx, catId in enumerate(cat_ids):
nm = lvis_gt.load_cats([catId])[0]
precision = precisions[:, :, idx, 0]
precision = precision[precision > -1]
if precision.size:
ap = np.mean(precision)
else:
ap = float('nan')
results_per_category.append(
(f'{nm["name"]}', f'{float(ap):0.3f}'))
num_columns = min(6, len(results_per_category) * 2)
results_flatten = list(
itertools.chain(*results_per_category))
headers = ['category', 'AP'] * (num_columns // 2)
results_2d = itertools.zip_longest(*[
results_flatten[i::num_columns]
for i in range(num_columns)
])
table_data = [headers]
table_data += [result for result in results_2d]
table = AsciiTable(table_data)
print_log('\n' + table.table)
format_summary_result_list = get_lvis_format_result(lvis_eval.params, lvis_eval.results)
format_summary_result = "\n".join(format_summary_result_list)
with open(f"cocofied_per-category-ap-{metric}.txt", 'w') as f:
f.write(table.table + "\n" + format_summary_result)
lvis_eval.print_results()