Questions on masking process for UCI dataset

[raghuramdr]

In the file notmiwae.py, consider the following function

´´´
def bernoulli_decoder_miss(self, z):

    if self.missing_process == 'selfmasking':

        self.W = tf.get_variable('W', shape=[1, 1, self.d])
        self.b = tf.get_variable('b', shape=[1, 1, self.d])

        logits = - self.W * (z - self.b)

    elif self.missing_process == 'selfmasking_known':

        self.W = tf.get_variable('W', shape=[1, 1, self.d])
        self.W = tf.nn.softplus(self.W)
        self.b = tf.get_variable('b', shape=[1, 1, self.d])

        logits = - self.W * (z - self.b)

    elif self.missing_process == 'linear':

        logits = keras.layers.Dense(units=self.d, activation=None, name='y')(z)

    elif self.missing_process == 'nonlinear':

        z = keras.layers.Dense(units=self.n_hidden, activation=tf.nn.tanh, name='y')(z)
        logits = keras.layers.Dense(units=self.d, activation=None, name='y')(z)

    else:
        print("use 'selfmasking', 'selfmasking_known', 'linear' or 'nonlinear' as 'missing_process'")
        logits = None

    # ---- return logits since it goes better with tfp bernoulli
    return logits

´´´

1. Do the nonlinear and linear options correspond to the agnostic process described in the paper?
2. selfmasking_known should correspond to the third approach described in the paper, which reads where the sign of the weights in the logistic regression is known. I don't follow how doing a tf.nn.softplus is linked to the knowledge of the sign of the weights. Can you please elaborate?

Cheers!