Fixture not found errors when fastai star imports and pytest

fastai
1

:wave: I’ve been writing some unit tests that use pytest fixtures (a fastai learner and dataloader with random data) and I’ve run into an odd set of errors when I use the star imports from fastai==2.2.7.

My workaround here is to not use star imports - even though it looks using star imports is the recommended approach.

Here’s a toy example:

import numpy as np
import pandas as pd
import pytest

# Using the star imports
from fastai.tabular.all import *

@pytest.fixture
def fastai_dl(tmp_path):
    # Returns a TabularDataLoader with some random data 
    m = np.random.random_sample((10, 10))
    df = pd.DataFrame(m, columns=[f"x{i}" for i in range(10)])
    df["y"] = (np.random.rand(10, 1) > 0.5).astype(int)
    return TabularDataLoaders.from_df(df, path=tmp_path, y_names=["y"])

@pytest.fixture
def fastai_learner(fastai_dl):
    learner = tabular_learner(fastai_dl)
    learner.fit_one_cycle(n_epoch=1)
    return learner

def test_example(fastai_learner):
    # Run a simple assertion just as a demo
    assert type(fastai_learner).__name__ == "TabularLearner"

When running this, I get a ton of errors. They are all fixture '<name>' not found.

❯ pytest test_fastai.py
================================================ test session starts ================================================
platform darwin -- Python 3.8.7, pytest-6.2.2, py-1.10.0, pluggy-0.13.1
rootdir: /Users/neal/src/github.com/operatorai/modelstore, configfile: pytest.ini
collected 15 items                                                                                                  

test_fastai.py EEEEEEEEEEEEEE.                                                                                [100%]

====================================================== ERRORS =======================================================
__________________________________________ ERROR at setup of test_overflow __________________________________________
file /Users/neal/.pyenv/versions/3.8.7/envs/operatorai.example.fastai/lib/python3.8/site-packages/fastai/callback/fp16.py, line 82
  def test_overflow(x):
E       fixture 'x' not found
>       available fixtures: cache, capfd, capfdbinary, caplog, capsys, capsysbinary, doctest_namespace, fastai_dl, fastai_learner, monkeypatch, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory
>       use 'pytest --fixtures [testpath]' for help on them.

[...]
ERROR test_fastai.py::test_overflow
ERROR test_fastai.py::test_set
ERROR test_fastai.py::test_sig
ERROR test_fastai.py::test_fail
ERROR test_fastai.py::test
ERROR test_fastai.py::test_eq
ERROR test_fastai.py::test_eq_type
ERROR test_fastai.py::test_ne
ERROR test_fastai.py::test_close
ERROR test_fastai.py::test_is
ERROR test_fastai.py::test_shuffled
ERROR test_fastai.py::test_stdout
ERROR test_fastai.py::test_warns
ERROR test_fastai.py::test_fig_exists

If I replace the * import with explicit imports, then my tests are all green:

from fastai.callback.schedule import fit_one_cycle
from fastai.learner import load_learner
from fastai.tabular.data import TabularDataLoaders
from fastai.tabular.learner import tabular_learner

(For completeness, I had to figure out to also explicitly import fit_one_cycle to resolve the 'TabularModel' object has no attribute 'fit_one_cycle' problem that would otherwise pop up)

I haven’t been able to find any similar problems on this forum or stackoverflow relating to this – perhaps I have haven’t configured pytest in the right way? Either way, I thought I’d share it here in case anyone has a better solution or runs into a similar problem: feedback welcome :pray: