Merge pull request #689 from mkmenta/remove-big-tensor-nonzero

Fixes the "nonzero is not supported for tensors with more than INT_MAX elements" in get_all_triplets_indices

Author: KevinMusgrave

src/pytorch_metric_learning/utils/loss_and_miner_utils.py

@@ -85,9 +85,41 @@ def neg_pairs_from_tuple(indices_tuple):
 
 
 def get_all_triplets_indices(labels, ref_labels=None):
-    matches, diffs = get_matches_and_diffs(labels, ref_labels)
-    triplets = matches.unsqueeze(2) * diffs.unsqueeze(1)
-    return torch.where(triplets)
+    all_matches, all_diffs = get_matches_and_diffs(labels, ref_labels)
+
+    if (all_matches.shape[0] * all_matches.shape[1] * all_matches.shape[1]
+            < torch.iinfo(torch.int32).max):
+        # torch.nonzero is not supported for tensors with more than INT_MAX elements
+        triplets = all_matches.unsqueeze(2) * all_diffs.unsqueeze(1)
+        return torch.where(triplets)
+
+    all_matches, all_diffs = all_matches.bool(), all_diffs.bool()
+
+    # Find anchors with at least a positive and a negative
+    indices = torch.arange(0, len(labels), device=labels.device)
+    indices = indices[all_matches.any(dim=1) & all_diffs.any(dim=1)]
+
+    # No triplets found
+    if len(indices) == 0:
+        return (torch.tensor([], device=labels.device, dtype=labels.dtype),
+                torch.tensor([], device=labels.device, dtype=labels.dtype),
+                torch.tensor([], device=labels.device, dtype=labels.dtype))
+
+    # Compute all triplets
+    anchors = []
+    positives = []
+    negatives = []
+    for i in indices:
+        matches = all_matches[i].nonzero(as_tuple=False).squeeze(1)
+        diffs = all_diffs[i].nonzero(as_tuple=False).squeeze(1)
+        nd = len(diffs)
+        nm = len(matches)
+        matches = matches.repeat_interleave(nd)
+        diffs = diffs.repeat(nm)
+        anchors.append(torch.full((len(matches),), i, dtype=labels.dtype, device=labels.device))
+        positives.append(matches)
+        negatives.append(diffs)
+    return torch.cat(anchors), torch.cat(positives), torch.cat(negatives)
 
 
 # sample triplets, with a weighted distribution if weights is specified.