Callable
📞 Callable or Models handle by BlackBox Attribution methods¶
The model can be something else than a tf.keras.Model if it respects one of the following condition:
- model(inputs: np.ndarray) return either a np.ndarray or a tf.Tensor of shape \((N, L)\) where \(N\) is the number of samples and \(L\) the number of targets
- The model has a scikit-learn API and has a predict_proba function
- The model is a xgboost.XGBModel from the XGBoost python library
- The model is a TF Lite model. Note this feature is experimental.
On the other hand, a PyTorch model can be used with method having Callable as type of model. In order to makes it work you should write a wrapper as follow:
class TemplateTorchWrapper(nn.Module):
def __init__(self, torch_model):
super(TemplateTorchWrapper, self).__init__()
self.model = torch_model
def __call__(self, inputs):
# transform your numpy inputs to torch
torch_inputs = self._transform_np_inputs(inputs)
# mak predictions
with torch.no_grad():
outputs = self.model(torch_inputs)
# convert to numpy
outputs = outputs.detach().numpy()
# convert to tf.Tensor
outputs = tf.cast(outputs, tf.float32)
return outputs
def _transform_np_inputs(self, np_inputs):
# include in this function all transformation
# needed for your torch model to work, here
# for example we swap from channels last to
# channels first
np_inputs = np.swapaxes(np_inputs, -1, 1)
torch_inputs = torch.Tensor(np_inputs)
return torch_inputs
wrapped_model = TemplateTorchWrapper(torch_model)
explainer = Lime(wrapped_model)
explanations = explainer.explain(images, labels)
As a matter of fact, if the instance of your model doesn't belong to [tf.keras.Model, tf.lite.Interpreter, sklearn.base.BaseEstimator, xgboost.XGBModel] when the explainer will need
to make inference the following will happen:
# inputs are automatically transform to tf.Tensor when using an explainer
pred = model(inputs.numpy())
pred = tf.cast(pred, dtype=tf.float32)
scores = tf.reduce_sum(pred * targets, axis=-1)