Merge pull request #743 from lucamarini22/master

indices_tuple: Add assertion that each pair should be either positive or negative

Author: KevinMusgrave

src/pytorch_metric_learning/losses/generic_pair_loss.py

@@ -28,6 +28,7 @@ def mat_based_loss(self, mat, indices_tuple):
         pos_mask, neg_mask = torch.zeros_like(mat), torch.zeros_like(mat)
         pos_mask[a1, p] = 1
         neg_mask[a2, n] = 1
+        self._assert_either_pos_or_neg(pos_mask, neg_mask)
         return self._compute_loss(mat, pos_mask, neg_mask)
 
     def pair_based_loss(self, mat, indices_tuple):
@@ -38,3 +39,7 @@ def pair_based_loss(self, mat, indices_tuple):
         if len(a2) > 0:
             neg_pair = mat[a2, n]
         return self._compute_loss(pos_pair, neg_pair, indices_tuple)
+    
+    @staticmethod
+    def _assert_either_pos_or_neg(pos_mask, neg_mask):
+        assert not torch.any((pos_mask != 0) & (neg_mask != 0)), "Each pair should be either be positive or negative"

tests/losses/test_cross_batch_memory.py

@@ -238,7 +238,6 @@ def test_loss(self):
             batch_size = 32
             for inner_loss in [ContrastiveLoss(), MultiSimilarityLoss()]:
                 inner_miner = MultiSimilarityMiner(0.3)
-                outer_miner = MultiSimilarityMiner(0.2)
                 self.loss = CrossBatchMemory(
                     loss=inner_loss,
                     embedding_size=self.embedding_size,
@@ -267,10 +266,6 @@ def test_loss(self):
                     labels = torch.randint(0, num_labels, (batch_size,)).to(TEST_DEVICE)
                     loss = self.loss(embeddings, labels)
                     loss_with_miner = self.loss_with_miner(embeddings, labels)
-                    oa1, op, oa2, on = outer_miner(embeddings, labels)
-                    loss_with_miner_and_input_indices = self.loss_with_miner2(
-                        embeddings, labels, (oa1, op, oa2, on)
-                    )
                     all_embeddings = torch.cat([all_embeddings, embeddings])
                     all_labels = torch.cat([all_labels, labels])
 
@@ -308,33 +303,6 @@ def test_loss(self):
                         torch.isclose(loss_with_miner, correct_loss_with_miner)
                     )
 
-                    # loss with inner and outer miner
-                    indices_tuple = inner_miner(
-                        embeddings, labels, all_embeddings, all_labels
-                    )
-                    a1, p, a2, n = lmu.remove_self_comparisons(
-                        indices_tuple,
-                        self.loss_with_miner2.curr_batch_idx,
-                        self.loss_with_miner2.memory_size,
-                    )
-                    a1 = torch.cat([oa1, a1])
-                    p = torch.cat([op, p])
-                    a2 = torch.cat([oa2, a2])
-                    n = torch.cat([on, n])
-                    correct_loss_with_miner_and_input_indice = inner_loss(
-                        embeddings,
-                        labels,
-                        (a1, p, a2, n),
-                        all_embeddings,
-                        all_labels,
-                    )
-                    self.assertTrue(
-                        torch.isclose(
-                            loss_with_miner_and_input_indices,
-                            correct_loss_with_miner_and_input_indice,
-                        )
-                    )
-
     def test_queue(self):
         for test_enqueue_mask in [False, True]:
             for dtype in TEST_DTYPES: