def test_mutate():
example = np.array([[0, 0, 1, 0], [0, 1, 0, 0]])
m = Mutation(DummyModel(), "dat1", ['diff'])
for ret, idxs in m._mutate_sample(example):
assert example[idxs[0], idxs[1]] == 0
kept_sel = np.arange(example.shape[0]) != idxs[0]
sel_j = np.arange(example.shape[1]) == idxs[1]
assert np.all(example[kept_sel, :] == ret[kept_sel, :])
assert np.all(ret[~kept_sel, sel_j] == 1)
assert np.all(ret[~kept_sel, ~sel_j] != 1)
def test_get_correct_model_input():
for model_input_type in input_examples:
# define batch input data:
batch_input = batch_inputs[model_input_type]
m = Mutation(get_dummy_model(model_input_type), "dat1", ['diff'])
idx = m.get_correct_model_input_id('dat1')
if model_input_type == "dict":
assert idx == 'dat1'
else:
assert idx == 0
def test_score():
# define batch input data:
batch_input = {
"dat1": np.array([[[0, 0, 1, 0], [0, 1, 0, 0]]])
} # 1 sample, seqlen 2, onehot encoded
m = Mutation(DummyModel(), "dat1", ['diff'])
scores_ret = m.score(batch_input)
# expected output:
for smpl_i, smpl in enumerate(scores_ret):
for i in range(len(smpl)):
for j in range(len(smpl[i])):
exp = (np.arange(0, 4)
== j).astype(int) - batch_input['dat1'][smpl_i, i, :]
if np.all(exp == 0):
assert smpl[i][j] is None
else:
smpl_diff = smpl[i][j][
0] # select the score i,j and the score 0 which is 'diff' here
model_out_diff = smpl_diff['dat1']
assert np.all(model_out_diff[i, :] == exp)
# test with selector_fn
sel_fn = lambda x: x['dat1']
m = Mutation(DummyModel(), "dat1", ['diff'], output_sel_fn=sel_fn)
scores_ret = m.score(batch_input)
for smpl_i, smpl in enumerate(scores_ret):
for i in range(len(smpl)):
for j in range(len(smpl[i])):
exp = (np.arange(0, 4)
== j).astype(int) - batch_input['dat1'][smpl_i, i, :]
if np.all(exp == 0):
assert smpl[i][j] is None
else:
smpl_diff = smpl[i][j][
0] # select the score i,j and the score 0 which is 'diff' here
model_out_diff = smpl_diff
assert np.all(model_out_diff[i, :] == exp)
def test_is_compatible():
is_compat_model = kipoi.get_model("DeepSEA/variantEffects")
m = Mutation(is_compat_model, "dat1", ['diff'])
assert m.is_compatible(is_compat_model)
def test_score():
for model_input_type in input_examples:
# define batch input data:
batch_input = batch_inputs[model_input_type]
m = Mutation(get_dummy_model(model_input_type), "dat1", ['diff'])
scores_ret = m.score(batch_input)
# expected output:
for smpl_i, smpl in enumerate(scores_ret):
for i in range(len(smpl)):
for j in range(len(smpl[i])):
bi = batch_input
if model_input_type == "dict":
bi = batch_input['dat1']
elif model_input_type == "list":
bi = batch_input[0]
exp = (np.arange(0, 4) == j).astype(int) - bi[smpl_i, i, :]
if np.all(exp == 0):
assert smpl[i][j] is None
else:
smpl_diff = smpl[i][j][
0] # select the score i,j and the score 0 which is 'diff' here
model_out_diff = smpl_diff
if model_input_type == "dict":
model_out_diff = smpl_diff['dat1']
elif model_input_type == "list":
model_out_diff = smpl_diff[0]
assert np.all(model_out_diff[i, :] == exp)
# test with selector_fn
if model_input_type == "dict":
sel_fn = lambda x: x['dat1']
elif model_input_type == "list":
sel_fn = lambda x: x[0]
else:
sel_fn = lambda x: x
m = Mutation(get_dummy_model(model_input_type),
"dat1", ['diff'],
output_sel_fn=sel_fn)
scores_ret = m.score(batch_input)
for smpl_i, smpl in enumerate(scores_ret):
for i in range(len(smpl)):
for j in range(len(smpl[i])):
bi = batch_input
if model_input_type == "dict":
bi = batch_input['dat1']
elif model_input_type == "list":
bi = batch_input[0]
exp = (np.arange(0, 4) == j).astype(int) - bi[smpl_i, i, :]
if np.all(exp == 0):
assert smpl[i][j] is None
else:
smpl_diff = smpl[i][j][
0] # select the score i,j and the score 0 which is 'diff' here
model_out_diff = smpl_diff
assert np.all(model_out_diff[i, :] == exp)
# test the test_ref_ref functionality
m = Mutation(get_dummy_model(model_input_type),
"dat1", ['diff'],
test_ref_ref=True)
scores_ret = m.score(batch_input)
# expected output:
for smpl_i, smpl in enumerate(scores_ret):
for i in range(len(smpl)):
for j in range(len(smpl[i])):
bi = batch_input
if model_input_type == "dict":
bi = batch_input['dat1']
elif model_input_type == "list":
bi = batch_input[0]
exp = (np.arange(0, 4) == j).astype(int) - bi[smpl_i, i, :]
# if test_ref_ref is set the score has to be returned for all!
smpl_diff = smpl[i][j][
0] # select the score i,j and the score 0 which is 'diff' here
model_out_diff = smpl_diff
if model_input_type == "dict":
model_out_diff = smpl_diff['dat1']
elif model_input_type == "list":
model_out_diff = smpl_diff[0]
assert np.all(model_out_diff[i, :] == exp)
def test_is_compatible():
return None # Circle-ci is rediciulous about this
is_compat_model = kipoi.get_model("tests/models/pyt", "dir")
m = Mutation(is_compat_model, "dat1", ['diff'])
assert m.is_compatible(is_compat_model)