Add a test that running contrastive loss doesn't kill the training

stanza/tests/constituency/test_trainer.py

@@ -277,6 +277,20 @@ def test_early_dropout(self, wordvec_pretrain_file):
             if all(module.p == 0.0 for _, module in dropouts):
                 raise AssertionError("All dropouts were 0 after training even though early_dropout was set to -1")
 
+    def test_contrastive(self, wordvec_pretrain_file):
+        """
+        Test that things don't blow up when a contrastive loss is used for a few iterations
+        """
+        with tempfile.TemporaryDirectory(dir=TEST_WORKING_DIR) as tmpdirname:
+            args = ['--contrastive_learning_rate', '0.1']
+            self.run_train_test(wordvec_pretrain_file, tmpdirname, num_epochs=6, extra_args=args)
+
+        with tempfile.TemporaryDirectory(dir=TEST_WORKING_DIR) as tmpdirname:
+            # TODO: get some kind of loss record back from the training process
+            # so that we can check it is being properly applied?
+            args = ['--contrastive_learning_rate', '0.1', '--contrastive_initial_epoch', '3']
+            self.run_train_test(wordvec_pretrain_file, tmpdirname, num_epochs=6, extra_args=args)
+
     def test_train_silver(self, wordvec_pretrain_file):
         """
         Test the whole thing for a few iterations on the fake data