def fit(
self, X: Optional[np.ndarray] = None, batch_size: int = 32, **kwargs
) -> None:
"""
Takes the same parameters as Keras's :func:`model.fit` function.
"""
generator = iutils.BatchSequence(X, batch_size)
return self._fit_generator(generator, **kwargs) # type: ignore
def compute(self, X, batch_size=32, verbose=0):
"""
Compute and return the patterns for the model and the data `X`.
:param X: Data to compute patterns.
:param batch_size: Batch size to use.
:param verbose: As for keras model.fit.
"""
generator = iutils.BatchSequence(X, batch_size)
return self.compute_generator(generator, verbose=verbose)
def test_fast__PerturbationAnalysis():
# Some test data
if keras.backend.image_data_format() == "channels_first":
input_shape = (2, 1, 4, 4)
else:
input_shape = (2, 4, 4, 1)
x = np.arange(2 * 4 * 4).reshape(input_shape)
generator = iutils.BatchSequence([x, np.zeros(x.shape[0])],
batch_size=x.shape[0])
# Simple model
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=x.shape[1:]),
keras.layers.Dense(1, use_bias=False),
])
weights = np.arange(4 * 4 * 1).reshape((4 * 4, 1))
model.layers[-1].set_weights([weights])
model.compile(loss="mean_squared_error", optimizer="sgd")
expected_output = np.array([[1240.0], [3160.0]])
assert np.all(np.isclose(model.predict(x), expected_output))
# Analyzer
analyzer = innvestigate.create_analyzer("gradient",
model,
postprocess="abs")
# Run perturbation analysis
perturbation = innvestigate.tools.perturbate.Perturbation("zeros",
region_shape=(2,
2),
in_place=False)
perturbation_analysis = innvestigate.tools.perturbate.PerturbationAnalysis(
analyzer,
model,
generator,
perturbation,
recompute_analysis=False,
steps=3,
regions_per_step=1,
verbose=False,
)
scores = perturbation_analysis.compute_perturbation_analysis()
expected_scores = np.array([5761600.0, 1654564.0, 182672.0, 21284.0])
assert np.all(np.isclose(scores, expected_scores))
###############################################################################
###############################################################################
###############################################################################
if __name__ == "__main__":
print("This script is deprecated and has been replaced by a Jupyter Notebook version (mnist_perturbation.ipynb).")
num_classes = 10
batch_size = 256
data = fetch_data()
data_preprocessed = (preprocess(data[0]), data[1],
preprocess(data[2]), data[3])
x_test, y_test = data_preprocessed[2:]
y_test = keras.utils.to_categorical(y_test, num_classes)
generator = iutils.BatchSequence([x_test, y_test], batch_size=batch_size)
# Build and train model
model_without_softmax, model_with_softmax = create_model()
train_model(model_with_softmax, data_preprocessed)
model_without_softmax.set_weights(model_with_softmax.get_weights())
# Setup analyzer
method = ("lrp.z_baseline", {}, ivis.heatmap, "LRP-Z")
analyzer = innvestigate.create_analyzer(method[0],
model_without_softmax,
**method[1])
# Perturbation analysis
perturbation_function = "mean"
perturbation = Perturbation(perturbation_function, region_shape=(5, 5))