def _test_train_model_from_features(self, sparse_gp, multiple_kernels):
gp_repurposer = GpRepurposer(self.source_model,
self.source_model_layers)
num_inducing = self.num_data_points_to_train
gp_repurposer.NUM_INDUCING_SPARSE_GP = num_inducing
if not sparse_gp: # Select a small data set to apply normal GP classification
self.train_features = self.train_features[:num_inducing]
self.train_labels = self.train_labels[:num_inducing]
self.feature_mean = self.train_features.mean(axis=0)
if multiple_kernels:
trained_model = gp_repurposer._train_model_from_features(
self.train_features, self.train_labels, {
'l1': self.train_feature_indices[:4],
'l2': self.train_feature_indices[4:]
})
else:
trained_model = gp_repurposer._train_model_from_features(
self.train_features, self.train_labels,
{'l1': self.train_feature_indices})
assert np.array_equal(gp_repurposer.feature_mean, self.feature_mean)
self._validate_trained_gp_model(trained_model, sparse_gp, num_inducing,
multiple_kernels)
def test_repurpose(self, mock_model_handler):
# Patch model_handler and then create gp_repurposer
mock_model_handler.return_value = RepurposerTestUtils.get_mock_model_handler_object()
mock_model_handler.return_value.get_layer_output.return_value = {'l1': self.train_features}, self.train_labels
gp_repurposer = GpRepurposer(self.source_model, self.source_model_layers)
gp_repurposer.NUM_INDUCING_SPARSE_GP = 5 # To speed-up unit test running time
self._run_common_repurposer_tests(gp_repurposer)
def _test_gp_serialisation(self, sparse_gp, multiple_kernels):
gp_repurposer = GpRepurposer(self.source_model,
self.source_model_layers,
apply_l2_norm=True)
num_inducing = 2
gp_repurposer.NUM_INDUCING_SPARSE_GP = num_inducing
if not sparse_gp: # Select a small data set to apply normal GP classification
self.train_features = self.train_features[:num_inducing]
self.train_labels = self.train_labels[:num_inducing]
self.feature_mean = self.train_features.mean(axis=0)
if multiple_kernels:
gp_repurposer.target_model = gp_repurposer._train_model_from_features(
self.train_features, self.train_labels, {
'l1': self.train_feature_indices[:4],
'l2': self.train_feature_indices[4:]
})
else:
gp_repurposer.target_model = gp_repurposer._train_model_from_features(
self.train_features, self.train_labels,
{'l1': self.train_feature_indices})
# Save and load repurposer to test serialization
loaded_repurposer = self._save_and_load_repurposer(gp_repurposer)
# Validate repurposer properties
self._compare_gp_repurposers(gp_repurposer, loaded_repurposer)
# Get prediction results using both repurposers
predictions_before = gp_repurposer._predict_probability_from_features(
self.test_features[:self.num_data_points_to_predict])
predictions_after = loaded_repurposer._predict_probability_from_features(
self.test_features[:self.num_data_points_to_predict])
# Compare probabilities predicted per test instance
self.assertTrue(
predictions_before.shape == predictions_after.shape,
"Prediction shape is incorrect. Expected: {} Actual: {}".format(
predictions_before.shape, predictions_after.shape))
for sample_id, prediction in enumerate(predictions_before):
self.assertTrue(
np.allclose(prediction, predictions_after[sample_id]),
"Incorrect prediction for sample id: {}. Expected: {} Actual: {}"
.format(sample_id, predictions_before[sample_id],
predictions_after[sample_id]))
# Validate if accuracy is above expected threshold
predicted_labels = np.argmax(predictions_after, axis=1)
accuracy = np.mean(predicted_labels ==
self.test_labels[:self.num_data_points_to_predict])
expected_accuracy = 0.3
self.assertTrue(
accuracy >= expected_accuracy,
"Accuracy {} less than {}".format(accuracy, expected_accuracy))