d0b03af329a2b2bf679ae33981f34636ce19690a,deepexplain/tensorflow/methods.py,,deepexplain_grad,#Any#Any#,162

Before Change

@ops.RegisterGradient("DeepExplainGrad")
def deepexplain_grad(op, grad):
    mode = tf.get_default_graph().get_tensor_by_name("deepexplain_mode:0")
    //mode = tf.Print(mode, [mode], "mode flag: ")

    def default():
        input = op.inputs[0]
        return grad * grad_activation(op.name)(input)

    cases = OrderedDict({
        tf.equal(mode, 0): (lambda: DummyZero.nonlinearity_grad_override(op, grad)),
        tf.equal(mode, 1): (lambda: Saliency.nonlinearity_grad_override(op, grad)),
        tf.equal(mode, 2): (lambda: GradientXInput.nonlinearity_grad_override(op, grad)),
        tf.equal(mode, 3): (lambda: IntegratedGradients.nonlinearity_grad_override(op, grad)),
        tf.equal(mode, 4): (lambda: EpsilonLRP.nonlinearity_grad_override(op, grad)),
    })
    return tf.case(cases, default=default, exclusive=True)

After Change

@ops.RegisterGradient("DeepExplainGrad")
def deepexplain_grad(op, grad):
    global _ENABLED_METHOD_CLASS
    if _ENABLED_METHOD_CLASS is not None:
        return _ENABLED_METHOD_CLASS.nonlinearity_grad_override(op, grad)
    else:
        return grad * grad_activation(op.name)(op.inputs[0])


class DeepExplain(object):

    def __init__(self, graph=tf.get_default_graph(), sess=tf.get_default_session()):
        self.method = None

Instances