diff --git a/tests/test_tensor_generators.py b/tests/test_tensor_generators.py
index 8128b3d51..7687cd5c4 100644
--- a/tests/test_tensor_generators.py
+++ b/tests/test_tensor_generators.py
@@ -5,6 +5,7 @@
 from collections import defaultdict
 
 from ml4cvd.defines import TENSOR_EXT
+from ml4cvd.tensor_generators import TensorGenerator
 from ml4cvd.tensor_generators import _sample_csv_to_set, get_train_valid_test_paths, get_train_valid_test_paths_split_by_csvs
 
 
@@ -92,6 +93,36 @@ def test_set(request, train_valid_test_csv):
     return train_valid_test_csv[2]
 
 
+@pytest.fixture(scope='function')
+def train_valid_test_paths(default_arguments, train_valid_test_csv):
+    args = default_arguments
+    (train_csv, train_ids), (valid_csv, valid_ids), (test_csv, test_ids) = train_valid_test_csv
+    return get_train_valid_test_paths(
+        tensors=args.tensors,
+        valid_ratio=args.valid_ratio,
+        test_ratio=args.test_ratio,
+        sample_csv=None,
+        train_csv=train_csv,
+        valid_csv=valid_csv,
+        test_csv=test_csv,
+    )
+
+
+@pytest.fixture(scope='function')
+def train_paths(train_valid_test_paths):
+    return train_valid_test_paths[0]
+
+
+@pytest.fixture(scope='function')
+def valid_paths(train_valid_test_paths):
+    return train_valid_test_paths[1]
+
+
+@pytest.fixture(scope='function')
+def test_paths(train_valid_test_paths):
+    return train_valid_test_paths[2]
+
+
 @pytest.fixture(scope='function')
 def valid_test_ratio():
     valid_ratio = np.random.randint(1, 5) / 10
@@ -113,6 +144,43 @@ def test_ratio(request, valid_test_ratio):
     return valid_test_ratio[1]
 
 
+class TestTensorGenerator:
+
+    # this test should currently fail - the failure is flaky
+    def test_get_true_epoch(self, default_arguments, train_paths):
+        num_workers = 2
+        num_tensors = len(train_paths) # 8 paths by default
+        batch_size = 2
+        num_steps = 20 # each path should be visited exactly 5 times
+
+        generator = TensorGenerator(
+            paths=train_paths,
+            keep_paths=True,
+            batch_size=batch_size,
+            input_maps=default_arguments.tensor_maps_in,
+            output_maps=default_arguments.tensor_maps_out,
+            num_workers=num_workers,
+            cache_size=default_arguments.cache_size,
+        )
+
+        rets = []
+        for i in range(num_steps):
+            rets.append(next(generator))
+
+        paths = []
+        for ret in rets:
+            paths.extend(ret[3])
+        unique_paths, counts = np.unique(paths, return_counts=True)
+        unique_counts = np.unique(counts)
+
+        try:
+            assert len(unique_counts) == 1 # make sure the tensors visited were seen the same number of times
+            assert set(unique_paths) == set(train_paths) # make sure all the tensors are visited
+            assert unique_counts[0] == batch_size * num_steps / num_tensors # make sure the tensors are visited the expected number of times
+        finally:
+            del generator
+
+
 class TestSampleCsvToSet:
     def test_sample_csv(self, sample_csv):
         csv_path, sample_ids = sample_csv
@@ -230,9 +298,3 @@ def test_get_paths_overlap(self, default_arguments, train_valid_test_csv):
             )
 
     # TODO test method with balance csvs
-
-
-class TestTensorGenerator:
-
-    def test_get_true_epoch(self):
-        pass