def _fit_single(self, model, X, y, masked_data=None):
Validation.check_dataset(X, y)
if len(X) != 0:
# Pre-compute target outputs if none are passed.
if masked_data is None:
output_dim = Validation.get_output_dimension(y)
masked_data = self.masking_operation.get_predictions_after_masking(self.explained_model, X, y,
batch_size=
self.model_builder.batch_size,
downsample_factors=
self.downsample_factors,
flatten=
self.flatten_for_explained_model)
masked_data = TensorflowCXPlain._clean_output_dims(output_dim, masked_data)
self.last_masked_data = masked_data
if self.model_filepath is None:
from tempfile import NamedTemporaryFile
model_filepath = NamedTemporaryFile(delete=False).name
else:
model_filepath = self.model_filepath
self.last_history = self.model_builder.fit(model, masked_data, y, model_filepath)
return self
def get_predictions_after_masking(self, explained_model, X, y, downsample_factors=(1,), batch_size=64,
flatten=False):
Validation.check_dataset(X, y)
num_batches = int(np.ceil(len(X) / float(batch_size)))
all_outputs = []
for batch_idx in range(num_batches):
x_batch = X[batch_idx*batch_size:(batch_idx+1)*batch_size]
y_pred = MaskingUtil.get_prediction(explained_model, x_batch, flatten=flatten)
x_imputed = []
for x_i in x_batch:
x_curr = []
for j in range(len(x_i)):
x_i_imputed_j = np.concatenate([x_i[:j], x_i[j+1:]], axis=0) # Drop entry at each index j.
x_curr.append(x_i_imputed_j)
x_imputed.append(x_curr)
all_y_pred_imputed = []
for j, x_imputed_curr in enumerate(x_imputed):
if len(x_imputed_curr) == 0:
y_pred_imputed = y_pred[j].reshape((1, -1))
else:
y_pred_imputed = MaskingUtil.get_prediction(explained_model, x_imputed_curr, flatten=flatten)
all_y_pred_imputed.append(y_pred_imputed)
all_outputs.append((x_batch, y_pred, all_y_pred_imputed))
all_outputs = [np.concatenate(list(map(partial(lambda x, dim: x[dim], dim=dim), all_outputs)))
for dim in range(len(all_outputs[0]))]
return all_outputs
def _build_model(self, X, y):
Validation.check_dataset(X, y)
if Validation.is_variable_length(X):
raise ValueError("Variable length inputs to CXPlain are currently not supported.")
n, p = Validation.get_input_dimension(X)
output_dim = Validation.get_output_dimension(y)
if self.model is None:
if self.num_models == 1:
build_fun = self._build_single
else:
build_fun = self._build_ensemble
self.model, self.prediction_model = build_fun(input_dim=p, output_dim=output_dim)
def get_predictions_after_masking(self,
explained_model,
X,
y,
downsample_factors=(1, ),
batch_size=64,
flatten=False):
Validation.check_dataset(X, y)
num_batches = int(np.ceil(len(X) / float(batch_size)))
all_outputs = []
for batch_idx in range(num_batches):
x = X[batch_idx * batch_size:(batch_idx + 1) * batch_size]
y_pred = MaskingUtil.get_prediction(explained_model,
x,
flatten=flatten)
x_imputed = MaskingUtil.get_x_imputed(x,
downsample_factors,
math_ops=NumpyInterface)
all_y_pred_imputed = []
for x_imputed_curr in x_imputed:
y_pred_imputed = MaskingUtil.get_prediction(explained_model,
x_imputed_curr,
flatten=flatten)
all_y_pred_imputed.append(y_pred_imputed)
all_y_pred_imputed = np.stack(all_y_pred_imputed).swapaxes(0, 1)
all_outputs.append((x, y_pred, all_y_pred_imputed))
all_outputs = [
np.concatenate(
list(map(partial(lambda x, dim: x[dim], dim=dim),
all_outputs))) for dim in range(len(all_outputs[0]))
]
return all_outputs