Improve ManhattanDistance metric for DDP

ignite/contrib/metrics/regression/manhattan_distance.py

@@ -1,32 +1,43 @@
+from typing import Callable, Union
+
 import torch
 
 from ignite.contrib.metrics.regression._base import _BaseRegression
+from ignite.metrics.metric import reinit__is_reduced, sync_all_reduce
 
 
 class ManhattanDistance(_BaseRegression):
     r"""
     Calculates the Manhattan Distance:
 
-    :math:`\text{MD} = \sum_{j=1}^n (A_j - P_j)`,
+    :math:`\text{MD} = \sum_{j=1}^n |A_j - P_j|`,
 
     where :math:`A_j` is the ground truth and :math:`P_j` is the predicted value.
 
-    More details can be found in `Botchkarev 2018`__.
+    More details can be found in `scikit-learn distance metrics`__.
 
     - ``update`` must receive output of the form ``(y_pred, y)`` or ``{'y_pred': y_pred, 'y': y}``.
     - `y` and `y_pred` must be of same shape `(N, )` or `(N, 1)`.
 
-    __ https://arxiv.org/abs/1809.03006
+    __ https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.DistanceMetric.html
 
     """
 
+    def __init__(
+        self, output_transform: Callable = lambda x: x, device: Union[str, torch.device] = torch.device("cpu")
+    ):
+        self._sum_of_errors = None
+        super(ManhattanDistance, self).__init__(output_transform, device)
+
+    @reinit__is_reduced
     def reset(self):
-        self._sum_of_errors = 0.0
+        self._sum_of_errors = torch.tensor(0.0, device=self._device)
 
     def _update(self, output):
         y_pred, y = output
-        errors = y.view_as(y_pred) - y_pred
-        self._sum_of_errors += torch.sum(errors).item()
+        errors = torch.abs(y - y_pred)
+        self._sum_of_errors += torch.sum(errors).to(self._device)
 
+    @sync_all_reduce("_sum_of_errors")
     def compute(self):
-        return self._sum_of_errors
+        return self._sum_of_errors.item()

tests/ignite/contrib/metrics/regression/test_manhattan_distance.py

@@ -1,7 +1,11 @@
+import os
+
 import numpy as np
 import pytest
 import torch
+from sklearn.neighbors import DistanceMetric
 
+import ignite.distributed as idist
 from ignite.contrib.metrics.regression import ManhattanDistance
 
 
@@ -30,18 +34,124 @@ def test_mahattan_distance():
 
     m = ManhattanDistance()
 
+    manhattan = DistanceMetric.get_metric("manhattan")
+
     m.update((torch.from_numpy(a), torch.from_numpy(ground_truth)))
-    np_ans = (ground_truth - a).sum()
-    assert m.compute() == pytest.approx(np_ans)
+    np_sum = np.abs(ground_truth - a).sum()
+    assert m.compute() == pytest.approx(np_sum)
+    assert manhattan.pairwise([a, ground_truth])[0][1] == pytest.approx(np_sum)
 
     m.update((torch.from_numpy(b), torch.from_numpy(ground_truth)))
-    np_ans += (ground_truth - b).sum()
-    assert m.compute() == pytest.approx(np_ans)
+    np_sum += np.abs(ground_truth - b).sum()
+    assert m.compute() == pytest.approx(np_sum)
+    v1 = np.hstack([a, b])
+    v2 = np.hstack([ground_truth, ground_truth])
+    assert manhattan.pairwise([v1, v2])[0][1] == pytest.approx(np_sum)
 
     m.update((torch.from_numpy(c), torch.from_numpy(ground_truth)))
-    np_ans += (ground_truth - c).sum()
-    assert m.compute() == pytest.approx(np_ans)
+    np_sum += np.abs(ground_truth - c).sum()
+    assert m.compute() == pytest.approx(np_sum)
+    v1 = np.hstack([v1, c])
+    v2 = np.hstack([v2, ground_truth])
+    assert manhattan.pairwise([v1, v2])[0][1] == pytest.approx(np_sum)
 
     m.update((torch.from_numpy(d), torch.from_numpy(ground_truth)))
-    np_ans += (ground_truth - d).sum()
-    assert m.compute() == pytest.approx(np_ans)
+    np_sum += np.abs(ground_truth - d).sum()
+    assert m.compute() == pytest.approx(np_sum)
+    v1 = np.hstack([v1, d])
+    v2 = np.hstack([v2, ground_truth])
+    assert manhattan.pairwise([v1, v2])[0][1] == pytest.approx(np_sum)
+
+
+def _test_distrib_compute(device):
+    rank = idist.get_rank()
+
+    manhattan = DistanceMetric.get_metric("manhattan")
+
+    def _test(metric_device):
+        metric_device = torch.device(metric_device)
+        m = ManhattanDistance(device=metric_device)
+        torch.manual_seed(10 + rank)
+
+        y_pred = torch.randint(0, 10, size=(10,), device=device).float()
+        y = torch.randint(0, 10, size=(10,), device=device).float()
+
+        m.update((y_pred, y))
+
+        # gather y_pred, y
+        y_pred = idist.all_gather(y_pred)
+        y = idist.all_gather(y)
+
+        np_y_pred = y_pred.cpu().numpy()
+        np_y = y.cpu().numpy()
+        res = m.compute()
+        assert manhattan.pairwise([np_y_pred, np_y])[0][1] == pytest.approx(res)
+
+    for _ in range(3):
+        _test("cpu")
+        if device.type != "xla":
+            _test(idist.device())
+
+
+@pytest.mark.distributed
+@pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.skipif(torch.cuda.device_count() < 1, reason="Skip if no GPU")
+def test_distrib_gpu(distributed_context_single_node_nccl):
+    device = torch.device("cuda:{}".format(distributed_context_single_node_nccl["local_rank"]))
+    _test_distrib_compute(device)
+
+
+@pytest.mark.distributed
+@pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")
+def test_distrib_cpu(distributed_context_single_node_gloo):
+
+    device = torch.device("cpu")
+    _test_distrib_compute(device)
+
+
+@pytest.mark.distributed
+@pytest.mark.skipif(not idist.has_hvd_support, reason="Skip if no Horovod dist support")
+@pytest.mark.skipif("WORLD_SIZE" in os.environ, reason="Skip if launched as multiproc")
+def test_distrib_hvd(gloo_hvd_executor):
+
+    device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
+    nproc = 4 if not torch.cuda.is_available() else torch.cuda.device_count()
+
+    gloo_hvd_executor(_test_distrib_compute, (device,), np=nproc, do_init=True)
+
+
+@pytest.mark.multinode_distributed
+@pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.skipif("MULTINODE_DISTRIB" not in os.environ, reason="Skip if not multi-node distributed")
+def test_multinode_distrib_cpu(distributed_context_multi_node_gloo):
+    device = torch.device("cpu")
+    _test_distrib_compute(device)
+
+
+@pytest.mark.multinode_distributed
+@pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.skipif("GPU_MULTINODE_DISTRIB" not in os.environ, reason="Skip if not multi-node distributed")
+def test_multinode_distrib_gpu(distributed_context_multi_node_nccl):
+    device = torch.device("cuda:{}".format(distributed_context_multi_node_nccl["local_rank"]))
+    _test_distrib_compute(device)
+
+
+@pytest.mark.tpu
+@pytest.mark.skipif("NUM_TPU_WORKERS" in os.environ, reason="Skip if NUM_TPU_WORKERS is in env vars")
+@pytest.mark.skipif(not idist.has_xla_support, reason="Skip if no PyTorch XLA package")
+def test_distrib_single_device_xla():
+    device = idist.device()
+    _test_distrib_compute(device)
+
+
+def _test_distrib_xla_nprocs(index):
+    device = idist.device()
+    _test_distrib_compute(device)
+
+
+@pytest.mark.tpu
+@pytest.mark.skipif("NUM_TPU_WORKERS" not in os.environ, reason="Skip if no NUM_TPU_WORKERS in env vars")
+@pytest.mark.skipif(not idist.has_xla_support, reason="Skip if no PyTorch XLA package")
+def test_distrib_xla_nprocs(xmp_executor):
+    n = int(os.environ["NUM_TPU_WORKERS"])
+    xmp_executor(_test_distrib_xla_nprocs, args=(), nprocs=n)