def test_hdf5_load(self):
hdf5_summary = "celltype_summary.hdf5"
orig_ast = Astir(self.expr, self.marker_dict)
orig_ast.fit_type(max_epochs=5, n_init=1, n_init_epochs=1)
orig_ast.fit_state(max_epochs=5, n_init=1, n_init_epochs=1)
orig_ast.save_models(hdf5_summary)
new_ast = Astir()
new_ast.load_model(hdf5_summary)
orig_type_run_info = orig_ast.get_type_run_info()
orig_state_run_info = orig_ast.get_state_run_info()
new_type_run_info = new_ast.get_type_run_info()
new_state_run_info = new_ast.get_state_run_info()
for key, val in orig_type_run_info.items():
if val != new_type_run_info[key]:
raise AssertionError(
"variable " + key +
" is different in original model and loaded model")
for key, val in orig_state_run_info.items():
if val != new_state_run_info[key]:
raise AssertionError(
"variable " + key +
" is different in original model and loaded model")
orig_type_losses = orig_ast.get_type_losses()
orig_state_losses = orig_ast.get_state_losses()
new_type_losses = new_ast.get_type_losses()
new_state_losses = new_ast.get_state_losses()
if not (all(orig_type_losses == new_type_losses)
and all(orig_state_losses == new_state_losses)):
raise AssertionError(
"loss is different in original model and loaded model")
class TestAstir(TestCase):
def __init__(self, *args, **kwargs):
super(TestAstir, self).__init__(*args, **kwargs)
warnings.filterwarnings("ignore", category=UserWarning)
self.expr_csv_file = os.path.join(os.path.dirname(__file__),
"test-data/test_data.csv")
self.marker_yaml_file = os.path.join(
os.path.dirname(__file__), "test-data/jackson-2020-markers.yml")
self.design_file = os.path.join(os.path.dirname(__file__),
"test-data/design.csv")
self.test_dir = os.path.join(os.path.dirname(__file__),
"test-data/test-dir-read")
self.adata_file = os.path.join(os.path.dirname(__file__),
"test-data/adata_small.h5ad")
self.expr = pd.read_csv(self.expr_csv_file, index_col=0)
with open(self.marker_yaml_file, "r") as stream:
self.marker_dict = yaml.safe_load(stream)
self.a = Astir(self.expr, self.marker_dict)
self._device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
def test_basic_instance_creation(self):
"""Tests basic instance creation"""
self.assertIsInstance(self.a, Astir)
# self.assertTrue(isinstance(a, str))
def test_csv_reading(self):
"""Test from_csv_yaml function"""
a = from_csv_yaml(self.expr_csv_file, self.marker_yaml_file)
self.assertIsInstance(a, Astir)
def test_dir_reading(self):
a = from_csv_dir_yaml(self.test_dir, self.marker_yaml_file)
self.assertIsInstance(a, Astir)
## Make sure the design matrix has been constructed correctly
self.assertTrue(a._type_dset._design.shape[0] == len(a._type_dset))
files = os.listdir(self.test_dir)
files = [f for f in files if f.endswith(".csv")]
self.assertTrue(a._type_dset._design.shape[1] == len(files))
def test_csv_reading_with_design(self):
a = from_csv_yaml(self.expr_csv_file,
self.marker_yaml_file,
design_csv=self.design_file)
self.assertIsInstance(a, Astir)
def test_fitting_type(self):
epochs = 2
with open(os.devnull, "w") as devnull:
with contextlib.redirect_stdout(devnull):
self.a.fit_type(max_epochs=epochs)
# Check probability matrix looks ok
probabilities = self.a.get_celltype_probabilities()
self.assertTrue(probabilities.shape[0] == self.expr.shape[0])
self.assertIsInstance(probabilities, pd.DataFrame)
# Check assignments look ok
assignments = self.a.get_celltypes()
self.assertIsInstance(assignments, pd.DataFrame)
self.assertTrue(assignments.shape[0] == self.expr.shape[0])
self.assertTrue(assignments.columns[0] == "cell_type")
# Check diagnostics look ok
type_diagnostics = self.a.diagnostics_celltype(threshold=0.2, alpha=0)
self.assertIsInstance(type_diagnostics, pd.DataFrame)
self.assertTrue(type_diagnostics.shape[1] ==
7) # make sure we have the standard 6 columns
def test_no_overlap(self):
bad_file = os.path.join(os.path.dirname(__file__),
"test-data/bad_data.csv")
bad_data = pd.read_csv(bad_file)
raised = False
try:
test = Astir(bad_data, self.marker_dict)
except (RuntimeError):
raised = True
self.assertTrue(raised == True)
def test_missing_marker(self):
bad_marker = os.path.join(os.path.dirname(__file__),
"test-data/bad_marker.yml")
with open(bad_marker, "r") as stream:
bad_dict = yaml.safe_load(stream)
raised = False
try:
test = Astir(self.expr, bad_dict)
test.fit_type()
except (RuntimeError):
raised = True
self.assertTrue(raised)
# # Uncomment below test functions to test private variables
# # Commented it out because these tests can be highly overlapping with
# # future unittests
def test_sanitize_dict_state(self):
"""Testing the method _sanitize_dict"""
expected_state_dict = self.marker_dict["cell_states"]
(_, actual_state_dict, _) = self.a._sanitize_dict(self.marker_dict)
expected_state_dict = {
i: sorted(j) if isinstance(j, list) else j
for i, j in expected_state_dict.items()
}
actual_state_dict = {
i: sorted(j) if isinstance(j, list) else j
for i, j in actual_state_dict.items()
}
self.assertDictEqual(
expected_state_dict,
actual_state_dict,
"state_dict is different from its expected value",
)
def test_state_names(self):
"""Test _state_names field"""
expected_state_names = sorted(self.marker_dict["cell_states"].keys())
(_, actual_state_dict, _) = self.a._sanitize_dict(self.marker_dict)
actual_state_names = sorted(actual_state_dict.keys())
self.assertListEqual(expected_state_names, actual_state_names,
"unexpected state_names value")
def test_celltype_same_seed_same_result(self):
"""Test whether the loss after one epoch one two different models
with the same random seed have the same losses after one epochs
"""
warnings.filterwarnings("ignore", category=UserWarning)
model1 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model2 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model1.fit_type(max_epochs=10)
model1_loss = model1.get_type_losses()
model2.fit_type(max_epochs=10)
model2_loss = model2.get_type_losses()
self.assertTrue(np.abs(model1_loss - model2_loss)[-1] < 1e-6)
def test_celltype_diff_seed_diff_result(self):
"""Test whether the loss after one epoch one two different models
with the different random seed have different losses after one epoch
"""
warnings.filterwarnings("ignore", category=UserWarning)
model1 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model2 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=1234,
)
model1.fit_type(max_epochs=10)
model1_loss = model1.get_type_losses()
model2.fit_type(max_epochs=10)
model2_loss = model2.get_type_losses()
self.assertFalse(np.abs(model1_loss - model2_loss)[-1] < 1e-6)
def test_cellstate_same_seed_same_result(self):
"""Test whether the loss after one epoch one two different models
with the same random seed have the same losses after one epochs
"""
warnings.filterwarnings("ignore", category=UserWarning)
model1 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model2 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model1.fit_state(max_epochs=5)
model1_loss = model1.get_state_losses()
model2.fit_state(max_epochs=5)
model2_loss = model2.get_state_losses()
self.assertTrue(np.abs(model1_loss - model2_loss)[-1] < 1e-6)
# @pytest.mark.filterwarnings("ignore")
def test_cellstate_diff_seed_diff_result(self):
"""Test whether the loss after one epoch one two different models
with the different random seed have different losses after one epoch
"""
warnings.filterwarnings("ignore", category=UserWarning)
model1 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=42,
)
model2 = Astir(
input_expr=self.expr,
marker_dict=self.marker_dict,
design=None,
random_seed=1234,
)
model1.fit_state(max_epochs=5)
model1_loss = model1.get_state_losses()
model2.fit_state(max_epochs=5)
model2_loss = model2.get_state_losses()
self.assertFalse(np.abs(model1_loss - model2_loss)[-1] < 1e-6)
def test_cellstate_assignment(self):
warnings.filterwarnings("ignore", category=UserWarning)
self.a.fit_state(max_epochs=50, n_init=1)
state_assignments = self.a.get_cellstates()
n_classes = len(list(self.marker_dict.keys()))
self.assertTrue(state_assignments.shape, (len(self.expr), n_classes))
def test_cellstate_predicted_assignment(self):
warnings.filterwarnings("ignore", category=UserWarning)
dset = SCDataset(
expr_input=self.expr,
marker_dict=self.marker_dict["cell_states"],
design=None,
include_other_column=False,
device=self._device,
)
self.a.fit_state(max_epochs=50, n_init=1)
state_assignments = self.a.predict_cellstates(dset)
self.assertTrue(state_assignments.shape,
(len(dset), dset.get_n_classes()))
def test_celltype_assignment(self):
warnings.filterwarnings("ignore", category=UserWarning)
self.a.fit_type(max_epochs=50, n_init=1)
type_assignments = self.a.get_celltypes()
n_classes = len(list(self.marker_dict.keys()))
self.assertTrue(type_assignments.shape,
(len(self.expr), n_classes + 1))
def test_celltype_predicted_assignment(self):
warnings.filterwarnings("ignore", category=UserWarning)
self.a.fit_type(max_epochs=50, n_init=1)
type_predict = self.a.predict_celltypes()
type_assignment = self.a.get_celltype_probabilities()
comp = type_predict == type_assignment
self.assertTrue(comp.all().all())
# def test_adata_reading(self):
# ast = from_anndata_yaml(
# self.adata_file,
# self.marker_yaml_file,
# protein_name="protein",
# cell_name="cell_name",
# batch_name="batch",
# )
# self.assertTrue(ast.get_type_dataset().get_n_features() == 14)
# self.assertTrue(ast.get_type_dataset().get_n_classes() == 6)
# self.assertIsInstance(ast.get_type_dataset().get_exprs(), torch.Tensor)
# self.assertEqual(ast.get_type_dataset().get_exprs().shape[0], 10)
def test_cellstate_diagnostics(self):
warnings.filterwarnings("ignore", category=UserWarning)
self.a.fit_state(max_epochs=50, n_init=1)
state_diagnostics = self.a.diagnostics_cellstate()
self.assertIsInstance(state_diagnostics, pd.DataFrame)
def test_type_hdf5_summary(self):
hdf5_summary = "celltype_summary.hdf5"
info = {
"max_epochs": 5,
"learning_rate": 0.001,
"batch_size": 24,
"delta_loss": 0.001,
"n_init": 1,
"n_init_epochs": 1,
}
self.a.fit_type(
max_epochs=info["max_epochs"],
learning_rate=info["learning_rate"],
batch_size=info["batch_size"],
delta_loss=info["delta_loss"],
n_init=info["n_init"],
n_init_epochs=info["n_init_epochs"],
)
self.a.save_models(hdf5_summary)
params = list(self.a.get_type_model().get_data().items()) + list(
self.a.get_type_model().get_variables().items())
recog_params = []
for key, val in self.a.get_type_model().get_recognet(
).named_parameters():
recog_params.append((key, val.detach().cpu().numpy()))
same = True
with h5py.File(hdf5_summary, "r") as f:
f_params = f["/celltype_model/parameters"]
for key, val in params:
if not (val.detach().cpu().numpy()
== f_params[key][()]).all().all():
same = False
f_recog = f["/celltype_model/recog_net"]
for key, val in recog_params:
if not (f_recog[key][()] == val).all():
same = False
f_info = f["/celltype_model/run_info"]
for key, val in info.items():
if val != f_info[key][()]:
same = False
if not (self.a.get_type_model().get_losses().cpu().numpy()
== f["/celltype_model/losses"]["losses"][()]).all():
same = False
self.assertTrue(same)
def test_state_summary(self):
hdf5_summary = "cellstate_summary.hdf5"
info = {
"max_epochs": 5,
"learning_rate": 0.001,
"batch_size": 24,
"delta_loss": 0.001,
"n_init": 1,
"n_init_epochs": 1,
"delta_loss_batch": 2,
}
self.a.fit_state(
max_epochs=info["max_epochs"],
learning_rate=info["learning_rate"],
batch_size=info["batch_size"],
delta_loss=info["delta_loss"],
n_init=info["n_init"],
n_init_epochs=info["n_init_epochs"],
delta_loss_batch=info["delta_loss_batch"],
)
self.a.save_models(hdf5_summary)
params = list(self.a.get_state_model().get_data().items()) + list(
self.a.get_state_model().get_variables().items())
recog_params = []
for key, val in self.a.get_state_model().get_recognet(
).named_parameters():
recog_params.append((key, val.detach().cpu().numpy()))
same = True
with h5py.File(hdf5_summary, "r") as f:
f_params = f["/cellstate_model/parameters"]
for key, val in params:
if not (val.detach().cpu().numpy()
== f_params[key][()]).all().all():
same = False
f_recog = f["/cellstate_model/recog_net"]
for key, val in recog_params:
if not (f_recog[key][()] == val).all():
same = False
f_info = f["/cellstate_model/run_info"]
for key, val in info.items():
if val != f_info[key][()]:
same = False
if not (self.a.get_state_model().get_losses().cpu().numpy()
== f["/cellstate_model/losses"]["losses"][()]).all():
same = False
self.assertTrue(same)
def test_hierarchy_assignment(self):
self.a.fit_type(max_epochs=5, n_init=1, n_init_epochs=1)
original_assignment = self.a.get_celltype_probabilities()
hier_dict = self.a.get_hierarchy_dict()
# expected_assignment = pd.DataFrame()
# for key, cells in hier_dict.items():
# expected_assignment[key] = original_assignment[cells].sum(axis=1)
actual_assignment = self.a.assign_celltype_hierarchy(depth=3)
# self.assertTrue((original_assignment == actual_assignment).all().all())
for cell in actual_assignment.columns:
self.assertTrue(
(actual_assignment[cell] == original_assignment[cell]).all())
def test_hdf5_load(self):
hdf5_summary = "celltype_summary.hdf5"
orig_ast = Astir(self.expr, self.marker_dict)
orig_ast.fit_type(max_epochs=5, n_init=1, n_init_epochs=1)
orig_ast.fit_state(max_epochs=5, n_init=1, n_init_epochs=1)
orig_ast.save_models(hdf5_summary)
new_ast = Astir()
new_ast.load_model(hdf5_summary)
orig_type_run_info = orig_ast.get_type_run_info()
orig_state_run_info = orig_ast.get_state_run_info()
new_type_run_info = new_ast.get_type_run_info()
new_state_run_info = new_ast.get_state_run_info()
for key, val in orig_type_run_info.items():
if val != new_type_run_info[key]:
raise AssertionError(
"variable " + key +
" is different in original model and loaded model")
for key, val in orig_state_run_info.items():
if val != new_state_run_info[key]:
raise AssertionError(
"variable " + key +
" is different in original model and loaded model")
orig_type_losses = orig_ast.get_type_losses()
orig_state_losses = orig_ast.get_state_losses()
new_type_losses = new_ast.get_type_losses()
new_state_losses = new_ast.get_state_losses()
if not (all(orig_type_losses == new_type_losses)
and all(orig_state_losses == new_state_losses)):
raise AssertionError(
"loss is different in original model and loaded model")