COCO mAP metric

docs/source/metrics.rst

@@ -332,11 +332,13 @@ Complete list of metrics
     Frequency
     Loss
     MeanAbsoluteError
+    MeanAveragePrecision
     MeanPairwiseDistance
     MeanSquaredError
     metric.Metric
     metrics_lambda.MetricsLambda
     MultiLabelConfusionMatrix
+    ObjectDetectionMAP
     precision.Precision
     PSNR
     recall.Recall

ignite/distributed/utils.py

@@ -1,9 +1,11 @@
+import itertools
 import socket
 from contextlib import contextmanager
 from functools import wraps
-from typing import Any, Callable, List, Mapping, Optional, Tuple, Union
+from typing import Any, Callable, List, Mapping, Optional, Sequence, Tuple, Union
 
 import torch
+from torch import distributed as dist
 
 from ignite.distributed.comp_models import (
     _SerialModel,
@@ -43,6 +45,7 @@
     "one_rank_only",
     "new_group",
     "one_rank_first",
+    "all_gather_tensors_with_shapes",
 ]
 
 _model = _SerialModel()
@@ -350,6 +353,35 @@ def all_reduce(
     return _model.all_reduce(tensor, op, group=group)
 
 
+def all_gather_tensors_with_shapes(
+    tensor: torch.Tensor, shapes: Sequence[Sequence[int]], group: Optional[Union[Any, List[int]]] = None
+) -> List[torch.Tensor]:
+    if _need_to_sync and isinstance(_model, _SerialModel):
+        sync(temporary=True)
+
+    if isinstance(group, list) and all(isinstance(item, int) for item in group):
+        group = _model.new_group(group)
+
+    if isinstance(_model, _SerialModel) or group == dist.GroupMember.NON_GROUP_MEMBER:
+        return [tensor]
+
+    max_shape = torch.tensor(shapes).amax(dim=0)
+    padding_sizes = (max_shape - torch.tensor(tensor.shape)).tolist()
+    padded_tensor = torch.nn.functional.pad(
+        tensor, tuple(itertools.chain.from_iterable(map(lambda dim_size: (0, dim_size), reversed(padding_sizes))))
+    )
+    all_padded_tensors: torch.Tensor = _model.all_gather(padded_tensor, group=group)  # .split(max_shape[0], dim=0)
+    return [
+        all_padded_tensors[
+            [
+                slice(rank * max_shape[0] if dim == 0 else 0, rank * max_shape[0] + dim_size if dim == 0 else dim_size)
+                for dim, dim_size in enumerate(shape)
+            ]
+        ]
+        for rank, shape in enumerate(shapes)
+    ]
+
+
 def all_gather(
     tensor: Union[torch.Tensor, float, str], group: Optional[Union[Any, List[int]]] = None
 ) -> Union[torch.Tensor, float, List[float], List[str]]:

ignite/metrics/__init__.py

@@ -9,6 +9,7 @@
 from ignite.metrics.gan.inception_score import InceptionScore
 from ignite.metrics.loss import Loss
 from ignite.metrics.mean_absolute_error import MeanAbsoluteError
+from ignite.metrics.mean_average_precision import MeanAveragePrecision
 from ignite.metrics.mean_pairwise_distance import MeanPairwiseDistance
 from ignite.metrics.mean_squared_error import MeanSquaredError
 from ignite.metrics.metric import BatchFiltered, BatchWise, EpochWise, Metric, MetricUsage
@@ -23,6 +24,7 @@
 from ignite.metrics.running_average import RunningAverage
 from ignite.metrics.ssim import SSIM
 from ignite.metrics.top_k_categorical_accuracy import TopKCategoricalAccuracy
+from ignite.metrics.vision.object_detection_map import ObjectDetectionMAP
 
 __all__ = [
     "Metric",
@@ -58,4 +60,6 @@
     "Rouge",
     "RougeN",
     "RougeL",
+    "MeanAveragePrecision",
+    "ObjectDetectionMAP",
 ]

ignite/metrics/metric.py

@@ -761,7 +761,7 @@ def _is_list_of_tensors_or_numbers(x: Sequence[Union[torch.Tensor, float]]) -> b
     return isinstance(x, Sequence) and all([isinstance(t, (torch.Tensor, Number)) for t in x])
 
 
-def _to_batched_tensor(x: Union[torch.Tensor, float], device: Optional[torch.device] = None) -> torch.Tensor:
+def _to_batched_tensor(x: Union[torch.Tensor, Number], device: Optional[torch.device] = None) -> torch.Tensor:
     if isinstance(x, torch.Tensor):
         return x.unsqueeze(dim=0)
     return torch.tensor([x], device=device)

ignite/metrics/vision/__init__.py

@@ -0,0 +1,3 @@
+from ignite.metrics.vision.object_detection_map import ObjectDetectionMAP
+
+__all__ = ["ObjectDetectionMAP"]