test_tensorflow_functional(batch_size=32, shuffle_config=2, ...) fails on pytest 8.0.0

[elisno]

Stack trace

From Github Actions logs:

=================================== FAILURES ===================================
_______________________ test_tensorflow_functional[32-2] _______________________

batch_size = 32, shuffle_config = 2
data = {'X': array([[-1.00724718, -0.92444024, -1.1659146 ],
       [-1.0525093 , -1.15150253, -1.13007914],
       [-1.09[831](https://github.com/cleanlab/cleanlab/actions/runs/7730565156/job/21076241756#step:8:832)...4693,  1.15885651,  0.84745434]]), 'error_indices': [47, 48, 49, 50, 51, 52], 'num_classes': 2, 'num_features': 3, ...}
hidden_units = 64

    @pytest.mark.slow
    @pytest.mark.skipif("not python_version_ok()", reason="need at least python 3.7")
    @pytest.mark.parametrize("batch_size,shuffle_config", [(1, 0), (32, 0), (32, 1), (32, 2)])
    def test_tensorflow_functional(batch_size, shuffle_config, data=DATA, hidden_units=64):
        dataset_tf = tf.data.Dataset.from_tensor_slices((data["X"], data["y"]))
        if shuffle_config == 0:  # proper shuffling for SGD
            dataset_shuffled = dataset_tf.shuffle(buffer_size=len(data["X"]))
        elif shuffle_config == 1:  # shuffling for datasets that don't fit in memory
            dataset_shuffled = dataset_tf.shuffle(buffer_size=60)
        else:
            dataset_shuffled = dataset_tf  # no shuffling
    
        dataset_og_order = dataset_tf.batch(batch_size)
        dataset_tf = dataset_shuffled.batch(batch_size)
    
        def make_model(num_features, num_classes):
            inputs = tf.keras.Input(shape=(num_features,))
            x = tf.keras.layers.Dense(hidden_units, activation="relu")(inputs)
            outputs = tf.keras.layers.Dense(num_classes)(x)
            model = tf.keras.Model(inputs=inputs, outputs=outputs, name="test_model")
    
            return model
    
        model = KerasWrapperModel(
            make_model,
            model_kwargs={"num_features": data["num_features"], "num_classes": data["num_classes"]},
        )
    
        model.summary()
    
        # Test base model works:
        model.fit(
            X=dataset_tf,
            y=data["y"],
            epochs=2,
        )
        preds_base = model.predict_proba(dataset_tf)
    
        # Test CleanLearning performs well:
        cl = CleanLearning(model)
        cl.fit(dataset_tf, data["y"], clf_kwargs={"epochs": 10}, clf_final_kwargs={"epochs": 15})
    
        preds = cl.predict(dataset_og_order)
        err = np.sum(preds != data["y_og"]) / len(data["y_og"])
        issue_indices = list(cl.label_issues_df[cl.label_issues_df["is_label_issue"]].index.values)
>       assert len(set(issue_indices) & set(data["error_indices"])) != 0
E       assert 0 != 0
E        +  where 0 = len((set() & {47, 48, 49, 50, 51, 52}))
E        +    where set() = set([])
E        +    and   {47, 48, 49, 50, 51, 52} = set([47, 48, 49, 50, 51, 52])

Steps to reproduce

• Run one of the Python 3.8+ test jobs in the CI.
• Only one test case fails where CleanLearning finds no issues in an un-shuffled dataset using the KerasModelWrapper

Additional information

• Cleanlab version: 5852a0a
• Operating system: ubuntu-latest
• Python version: 3.10

Capping pytest~=7.4 seems to resolve the issue. Done in #975.
This version cap should be reverted asap.