def test_decapitate_model_lazy_input():
"""Test an error is raised when the model has a lazy input layer initialization"""
# Raise warning when model has lazy input layer initialization
error_model = Sequential(
[Dense(40, input_shape=(100, )),
Dense(20),
Activation('softmax')])
with warnings.catch_warnings(record=True) as warning_check:
_decapitate_model(error_model, 1)
assert len(warning_check) == 1
assert "depth issues" in str(warning_check[-1].message)
def test_decapitate_model():
"""
This test creates a toy network, and checks that it calls the right errors
and checks that it decapitates the network correctly:
"""
# Create test model
test_model = _decapitate_model(CHECK_MODEL, 5)
# Make checks for all of the necessary features: the model outputs, the
# last layer, the last layer's connections, and the last layer's shape
assert test_model.layers[-1] == test_model.layers[3]
assert test_model.layers[3].outbound_nodes == []
assert test_model.outputs == [test_model.layers[3].output]
assert test_model.layers[-1].output_shape == (None, 20)
def test_decapitate_model_too_deep():
"""Test error raised when model is decapitated too deep"""
# Check for Value Error when passed a depth >= (# of layers in network) - 1
with pytest.raises(ValueError):
_decapitate_model(CHECK_MODEL, 8)