Possible NaN bug in logistic regression

[llylly]

Thanks for offering this great repository! I found that https://github.com/gabrieleangeletti/Deep-Learning-TensorFlow/blob/ddeb1f2848da7b7bee166ad2152b4afc46bb2086/yadlt/models/linear/logistic_regression.py#L52-L55 may contain a numerical bug when the loss_func is "cross_entropy" since self.mod_y can be close to zero.

To fix this, maybe we can replace

self.mod_y = tf.nn.softmax( 
     tf.add(tf.matmul(self.input_data, self.W_), self.b_)) 

by

self.mod_y = tf.nn.softmax( 
     tf.clip_by_value(tf.add(tf.matmul(self.input_data, self.W_), self.b_), 1e-10,1.0)

Thanks!