def test_boston_housing_no_fit_invalid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
explained_model = RandomForestRegressor(n_estimators=64,
max_depth=5,
random_state=1)
explained_model.fit(x_train, y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=32,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=32,
learning_rate=0.001,
num_epochs=2,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = mean_squared_error
explainer = CXPlain(explained_model, model_builder, masking_operation,
loss)
with self.assertRaises(AssertionError):
explainer.predict(x_test, y_test)
with self.assertRaises(AssertionError):
explainer.score(x_test, y_test)
def test_time_series_valid(self):
num_samples = 1024
fixed_length = 99
(x_train,
y_train), (x_test, y_test) = TestUtil.get_random_fixed_length_dataset(
num_samples=num_samples, fixed_length=fixed_length)
model_builder = RNNModelBuilder(with_embedding=False,
num_layers=2,
num_units=32,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=32,
learning_rate=0.001,
num_epochs=2,
early_stopping_patience=128)
explained_model = MLPClassifier()
explained_model.fit(x_train.reshape((-1, np.prod(x_train.shape[1:]))),
y_train)
masking_operation = ZeroMasking()
loss = binary_crossentropy
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
flatten_for_explained_model=True)
explainer.fit(x_train, y_train)
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median = explainer.predict(x_test)
self.assertTrue(median.shape == x_test.shape)
def test_boston_housing_load_save_valid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
explained_model = RandomForestRegressor(n_estimators=64, max_depth=5, random_state=1)
explained_model.fit(x_train, y_train)
model_builder = MLPModelBuilder(num_layers=2, num_units=32, activation="relu", p_dropout=0.2, verbose=0,
batch_size=32, learning_rate=0.001, num_epochs=2, early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = mean_squared_error
num_models_settings = [1, 2]
for num_models in num_models_settings:
explainer = CXPlain(explained_model, model_builder, masking_operation, loss,
num_models=num_models)
explainer.fit(x_train, y_train)
median_1 = explainer.predict(x_test)
tmp_dir_name = tempfile.mkdtemp()
explainer.save(tmp_dir_name)
with self.assertRaises(ValueError):
explainer.save(tmp_dir_name, overwrite=False)
explainer.save(tmp_dir_name, overwrite=True)
explainer.load(tmp_dir_name)
median_2 = explainer.predict(x_test)
self.assertTrue(np.array_equal(median_1, median_2))
shutil.rmtree(tmp_dir_name) # Cleanup.
def test_boston_housing_valid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
explained_model = RandomForestRegressor(n_estimators=64,
max_depth=5,
random_state=1)
explained_model.fit(x_train, y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=32,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=32,
learning_rate=0.001,
num_epochs=2,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = mean_squared_error
explainer = CXPlain(explained_model, model_builder, masking_operation,
loss)
explainer.fit(x_train, y_train)
self.assertEqual(explainer.prediction_model.output_shape,
(None, np.prod(x_test.shape[1:])))
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median = explainer.predict(x_test)
self.assertTrue(median.shape == x_test.shape)
def test_boston_housing_confidence_level_invalid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
explained_model = RandomForestRegressor(n_estimators=64,
max_depth=5,
random_state=1)
explained_model.fit(x_train, y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=32,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=32,
learning_rate=0.001,
num_epochs=3,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = mean_squared_error
num_models = 2
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
num_models=num_models)
explainer.fit(x_train, y_train)
invalid_confidence_levels = [1.01, -0.5, -0.01]
for confidence_level in invalid_confidence_levels:
with self.assertRaises(ValueError):
explainer.predict(x_test, confidence_level=confidence_level)
def test_mnist_unet_valid(self):
num_subsamples = 100
(x_train, y_train), (x_test, y_test) = TestUtil.get_mnist(flattened=False, num_subsamples=num_subsamples)
explained_model = MLPClassifier(solver='lbfgs', alpha=1e-5,
hidden_layer_sizes=(64, 32), random_state=1)
explained_model.fit(x_train.reshape((len(x_train), -1)), y_train)
masking_operation = ZeroMasking()
loss = categorical_crossentropy
downsample_factors = [(2, 2), (4, 4), (4, 7), (7, 4), (7, 7)]
with_bns = [True if i % 2 == 0 else False for i in range(len(downsample_factors))]
for downsample_factor, with_bn in zip(downsample_factors, with_bns):
model_builder = UNetModelBuilder(downsample_factor, num_layers=2, num_units=64, activation="relu",
p_dropout=0.2, verbose=0, batch_size=256, learning_rate=0.001,
num_epochs=2, early_stopping_patience=128, with_bn=with_bn)
explainer = CXPlain(explained_model, model_builder, masking_operation, loss,
downsample_factors=downsample_factor, flatten_for_explained_model=True)
explainer.fit(x_train, y_train)
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median = explainer.predict(x_test)
self.assertTrue(median.shape == x_test.shape)
def test_mnist_unet_with_shape_valid(self):
num_subsamples = 100
(x_train,
y_train), (x_test,
y_test) = TestUtil.get_mnist(flattened=False,
num_subsamples=num_subsamples)
explained_model_builder = MLPModelBuilder(num_layers=2,
num_units=64,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=256,
learning_rate=0.001,
num_epochs=2,
early_stopping_patience=128)
input_shape = x_train.shape[1:]
input_layer = Input(shape=input_shape)
last_layer = Flatten()(input_layer)
last_layer = explained_model_builder.build(last_layer)
last_layer = Dense(y_train.shape[-1], activation="softmax")(last_layer)
explained_model = Model(input_layer, last_layer)
explained_model.compile(loss="categorical_crossentropy",
optimizer="adam")
explained_model.fit(x_train, y_train)
masking_operation = ZeroMasking()
loss = categorical_crossentropy
downsample_factors = [(2, 2), (4, 4), (4, 7), (7, 4), (7, 7)]
with_bns = [
True if i % 2 == 0 else False
for i in range(len(downsample_factors))
]
for downsample_factor, with_bn in zip(downsample_factors, with_bns):
model_builder = UNetModelBuilder(downsample_factor,
num_layers=2,
num_units=64,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=256,
learning_rate=0.001,
num_epochs=2,
early_stopping_patience=128,
with_bn=with_bn)
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
downsample_factors=downsample_factor)
explainer.fit(x_train, y_train)
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median = explainer.predict(x_test)
self.assertTrue(median.shape == x_test.shape)
def test_mnist_valid(self):
num_subsamples = 100
(x_train,
y_train), (x_test,
y_test) = TestUtil.get_mnist(flattened=False,
num_subsamples=num_subsamples)
explained_model = MLPClassifier(solver='lbfgs',
alpha=1e-5,
hidden_layer_sizes=(64, 32),
random_state=1)
explained_model.fit(x_train.reshape((len(x_train), -1)), y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=64,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=256,
learning_rate=0.001,
num_epochs=3,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = categorical_crossentropy
downsample_factors = [(2, 2), (4, 4), (4, 7), (7, 4), (7, 7)]
for downsample_factor in downsample_factors:
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
num_models=2,
downsample_factors=downsample_factor,
flatten_for_explained_model=True)
explainer.fit(x_train, y_train)
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median, confidence = explainer.predict(x_test,
confidence_level=0.95)
self.assertTrue(median.shape == x_test.shape)
self.assertTrue(confidence.shape == x_test.shape[:-1] + (2, ))
# Flatten predictions for iteration below.
median = median.reshape((len(x_test), -1))
confidence = confidence.reshape((len(x_test), -1, 2))
for sample_idx in range(len(x_test)):
for feature_idx in range(len(x_test[sample_idx])):
self.assertTrue(confidence[sample_idx][feature_idx][0] <=
median[sample_idx][feature_idx] <=
confidence[sample_idx][feature_idx][1])
self.assertTrue(
confidence[sample_idx][feature_idx][0] >= 0)
self.assertTrue(
confidence[sample_idx][feature_idx][1] >= 0)
def test_overwrite_ensemble_model_invalid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
model_builder = MLPModelBuilder()
explained_model = RandomForestRegressor(n_estimators=64,
max_depth=5,
random_state=1)
explained_model.fit(x_train, y_train)
masking_operation = ZeroMasking()
loss = binary_crossentropy
num_models = 5
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
num_models=num_models)
file_names = [
CXPlain.get_config_file_name(),
CXPlain.get_explained_model_file_name(".pkl"),
CXPlain.get_loss_pkl_file_name(),
CXPlain.get_model_builder_pkl_file_name(),
CXPlain.get_masking_operation_pkl_file_name()
]
# Test with untrained explanation model.
for file_name in file_names:
tmp_dir = TestExplanationModel.make_at_tmp(file_name)
with self.assertRaises(ValueError):
explainer.save(tmp_dir, overwrite=False)
# Test with trained explanation model.
explainer.fit(x_train, y_train)
file_names = [
CXPlain.get_config_file_name(),
CXPlain.get_explained_model_file_name(".pkl"),
CXPlain.get_loss_pkl_file_name(),
CXPlain.get_model_builder_pkl_file_name(),
CXPlain.get_masking_operation_pkl_file_name()
] + [
CXPlain.get_prediction_model_h5_file_name(i)
for i in range(num_models)
]
for file_name in file_names:
tmp_dir = TestExplanationModel.make_at_tmp(file_name)
with self.assertRaises(ValueError):
explainer.save(tmp_dir, overwrite=False)
def test_boston_housing_valid(self):
(x_train, y_train), (x_test, y_test) = TestUtil.get_boston_housing()
explained_model = RandomForestRegressor(n_estimators=64,
max_depth=5,
random_state=1)
explained_model.fit(x_train, y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=32,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=32,
learning_rate=0.001,
num_epochs=3,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = mean_squared_error
for num_models in [2, 5, 10]:
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
num_models=num_models)
explainer.fit(x_train, y_train)
eval_score = explainer.score(x_test, y_test)
train_score = explainer.get_last_fit_score()
median, confidence = explainer.predict(x_test,
confidence_level=0.95)
self.assertTrue(median.shape == x_test.shape)
self.assertTrue(confidence.shape == x_test.shape + (2, ))
# Flatten predictions for iteration below.
median = median.reshape((len(x_test), -1))
confidence = confidence.reshape((len(x_test), -1, 2))
for sample_idx in range(len(x_test)):
for feature_idx in range(len(x_test[sample_idx])):
self.assertTrue(confidence[sample_idx][feature_idx][0] <=
median[sample_idx][feature_idx] <=
confidence[sample_idx][feature_idx][1])
self.assertTrue(
confidence[sample_idx][feature_idx][0] >= 0)
self.assertTrue(
confidence[sample_idx][feature_idx][1] >= 0)
def test_mnist_confidence_levels_valid(self):
num_subsamples = 100
(x_train,
y_train), (x_test,
y_test) = TestUtil.get_mnist(flattened=False,
num_subsamples=num_subsamples)
explained_model = MLPClassifier(solver='lbfgs',
alpha=1e-5,
hidden_layer_sizes=(64, 32),
random_state=1)
explained_model.fit(x_train.reshape((len(x_train), -1)), y_train)
model_builder = MLPModelBuilder(num_layers=2,
num_units=64,
activation="relu",
p_dropout=0.2,
verbose=0,
batch_size=256,
learning_rate=0.001,
num_epochs=3,
early_stopping_patience=128)
masking_operation = ZeroMasking()
loss = categorical_crossentropy
confidence_levels = [0.0, 1.0, 1.01, -0.01]
for confidence_level in confidence_levels:
downsample_factor = (2, 2)
explainer = CXPlain(explained_model,
model_builder,
masking_operation,
loss,
num_models=2,
downsample_factors=downsample_factor,
flatten_for_explained_model=True)
explainer.fit(x_train, y_train)
with self.assertRaises(ValueError):
_ = explainer.predict(x_test,
confidence_level=confidence_level)
def get_feature_importances(self, model):
from cxplain import ZeroMasking
from cxplain.util.test_util import TestUtil
from cxplain.backend.masking.masking_util import MaskingUtil
from cxplain.backend.causal_loss import calculate_delta_errors
from cxplain.backend.numpy_math_interface import NumpyInterface
x, y = self.test_set[0], self.test_set[1]
masking = ZeroMasking()
if isinstance(model, Pipeline):
transform = model.steps[0][1]
x = transform.transform(np.array(x))
model = model.steps[1][1]
num_features = np.array(x).shape[-1]
max_num_feature_groups = int(
np.rint(self.args["max_num_feature_groups"]))
if max_num_feature_groups >= num_features:
_, y_pred, all_y_pred_imputed = masking.get_predictions_after_masking(
model,
x,
y,
batch_size=len(x),
downsample_factors=(1, ),
flatten=True)
auxiliary_outputs = y_pred
all_but_one_auxiliary_outputs = all_y_pred_imputed
all_but_one_auxiliary_outputs = TestUtil.split_auxiliary_outputs_on_feature_dim(
all_but_one_auxiliary_outputs)
delta_errors = calculate_delta_errors(
np.expand_dims(y, axis=-1),
auxiliary_outputs,
all_but_one_auxiliary_outputs,
NumpyInterface.binary_crossentropy,
math_ops=NumpyInterface)
group_importances = np.mean(delta_errors, axis=0)
feature_groups = np.expand_dims(np.arange(x[0].shape[-1]),
axis=-1).tolist()
else:
class ModelWrapper(object):
def __init__(self, wrapped_model, real_data,
dummy_to_real_mapping):
self.wrapped_model = wrapped_model
self.real_data = real_data
self.dummy_to_real_mapping = dummy_to_real_mapping
def map_from_dummy(self, dummy):
mask = np.ones(np.array(self.real_data).shape)
for i, row in enumerate(dummy):
for j, group in enumerate(self.dummy_to_real_mapping):
if row[j] == 0.:
mask[i, group] = 0
return self.real_data * mask
def predict(self, x):
x = self.map_from_dummy(x)
y = MaskingUtil.predict_proxy(model, x)
if len(y.shape) == 1:
y = np.expand_dims(y, axis=-1)
return y
num_groups = 1
feature_groups = [np.random.permutation(np.arange(x[0].shape[-1]))]
group_importances = [1.0]
while num_groups < max_num_feature_groups:
num_groups += 1
# Recurse into largest relative importance group.
did_find_splittable = False
highest_importances = np.argsort(group_importances)[::-1]
for highest_importance in highest_importances:
if len(feature_groups[highest_importance]) != 1:
did_find_splittable = True
break
else:
continue
if not did_find_splittable:
# max_num_groups > len(features) - abort.
break
rest = len(feature_groups[highest_importance]) % 2
if rest != 0:
carry = feature_groups[highest_importance][-rest:].tolist()
feature_groups[highest_importance] = feature_groups[
highest_importance][:-rest]
else:
carry = []
feature_groups[highest_importance] = np.split(
feature_groups[highest_importance], 2)
feature_groups[highest_importance][0] = np.array(
feature_groups[highest_importance][0].tolist() + carry)
recombined = feature_groups[:highest_importance] + \
feature_groups[highest_importance] + \
feature_groups[highest_importance + 1:]
assert 0 not in map(len, recombined)
feature_groups = recombined
wrapped_model = ModelWrapper(model, x, feature_groups)
dummy_data = np.ones((len(x), num_groups))
_, y_pred, all_y_pred_imputed = masking.get_predictions_after_masking(
wrapped_model,
dummy_data,
y,
batch_size=len(x),
downsample_factors=(1, ),
flatten=True)
auxiliary_outputs = y_pred
all_but_one_auxiliary_outputs = all_y_pred_imputed
all_but_one_auxiliary_outputs = TestUtil.split_auxiliary_outputs_on_feature_dim(
all_but_one_auxiliary_outputs)
delta_errors = calculate_delta_errors(
np.expand_dims(y, axis=-1),
auxiliary_outputs,
all_but_one_auxiliary_outputs,
NumpyInterface.binary_crossentropy,
math_ops=NumpyInterface)
group_importances = np.mean(delta_errors, axis=0)
return feature_groups, group_importances
def cxpl(model_dir, data_dir, results_subdir, random_seed, resolution):
np.random.seed(random_seed)
tf.set_random_seed(np.random.randint(1 << 31))
session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)
set_session(sess)
# parser config
config_file = model_dir+ "/config.ini"
print("Config File Path:", config_file,flush=True)
assert os.path.isfile(config_file)
cp = ConfigParser()
cp.read(config_file)
output_dir = os.path.join(results_subdir, "classification_results/test")
print("Output Directory:", output_dir,flush=True)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
# default config
image_dimension = cp["TRAIN"].getint("image_dimension")
gan_resolution = resolution
batch_size = cp["TEST"].getint("batch_size")
use_best_weights = cp["TEST"].getboolean("use_best_weights")
if use_best_weights:
print("** Using BEST weights",flush=True)
model_weights_path = os.path.join(results_subdir, "classification_results/train/best_weights.h5")
else:
print("** Using LAST weights",flush=True)
model_weights_path = os.path.join(results_subdir, "classification_results/train/weights.h5")
print("** DenseNet Input Resolution:", image_dimension, flush=True)
print("** GAN Image Resolution:", gan_resolution, flush=True)
# get test sample count
test_dir = os.path.join(results_subdir, "inference/test")
shutil.copy(test_dir+"/test.csv", output_dir)
# Get class names
class_names = get_class_names(output_dir,"test")
tfrecord_dir_te = os.path.join(data_dir, "test")
test_counts, _ = get_sample_counts(output_dir, "test", class_names)
# get indicies (all of csv file for validation)
print("** test counts:", test_counts, flush=True)
# compute steps
test_steps = int(np.floor(test_counts / batch_size))
print("** test_steps:", test_steps, flush=True)
log2_record = int(np.log2(gan_resolution))
record_file_ending = "*"+ np.str(log2_record)+ ".tfrecords"
print("** resolution ", gan_resolution, " corresponds to ", record_file_ending, " TFRecord file.", flush=True)
# Get Model
# ------------------------------------
input_shape=(image_dimension, image_dimension, 3)
img_input = Input(shape=input_shape)
base_model = DenseNet121(
include_top = False,
weights = None,
input_tensor = img_input,
input_shape = input_shape,
pooling = "avg")
x = base_model.output
predictions = Dense(len(class_names), activation="sigmoid", name="predictions")(x)
model = Model(inputs=img_input, outputs = predictions)
print(" ** load model from:", model_weights_path, flush=True)
model.load_weights(model_weights_path)
# ------------------------------------
print("** load test generator **", flush=True)
test_seq = TFWrapper(
tfrecord_dir=tfrecord_dir_te,
record_file_endings = record_file_ending,
batch_size = batch_size,
model_target_size = (image_dimension, image_dimension),
steps = None,
augment=False,
shuffle=False,
prefetch=True,
repeat=False)
print("** make prediction **", flush=True)
test_seq.initialise()
x_all, y_all = test_seq.get_all_test_data()
print("X-Test Shape:", x_all.shape,flush=True)
print("Y-Test Shape:", y_all.shape,flush=True)
print("----------------------------------------", flush=True)
print("Test Model AUROC", flush=True)
y_pred = model.predict(x_all)
current_auroc = []
for i in range(len(class_names)):
try:
score = roc_auc_score(y_all[:, i], y_pred[:, i])
except ValueError:
score = 0
current_auroc.append(score)
print(i+1,class_names[i],": ", score, flush=True)
mean_auroc = np.mean(current_auroc)
print("Mean auroc: ", mean_auroc,flush=True)
print("----------------------------------------", flush=True)
downscale_factor = 8
num_models_to_use = 3
num_test_images = 100
print("Number of Models to use:", num_models_to_use, flush=True)
print("Number of Test images:", num_test_images, flush=True)
x_tr, y_tr = x_all[num_test_images:], y_all[num_test_images:]
x_te, y_te = x_all[0:num_test_images], y_all[0:num_test_images]
downsample_factors = (downscale_factor,downscale_factor)
print("Downsample Factors:", downsample_factors,flush=True)
model_builder = UNetModelBuilder(downsample_factors, num_layers=2, num_units=8, activation="relu",
p_dropout=0.0, verbose=0, batch_size=32, learning_rate=0.001)
print("Model build done.",flush=True)
masking_operation = ZeroMasking()
loss = categorical_crossentropy
explainer = CXPlain(model, model_builder, masking_operation, loss,
num_models=num_models_to_use, downsample_factors=downsample_factors, flatten_for_explained_model=False)
print("Explainer build done.",flush=True)
explainer.fit(x_tr, y_tr);
print("Explainer fit done.",flush=True)
try:
attr, conf = explainer.explain(x_te, confidence_level=0.80)
np.save(output_dir+"/x_cxpl.npy", x_te)
np.save(output_dir+"/y_cxpl.npy", y_te)
np.save(output_dir+"/attr.npy", attr)
np.save(output_dir+"/conf.npy", conf)
print("Explainer explain done and saved.",flush=True)
except Exception as ef: print(ef,flush=True)