ndrplz / transforming-autoencoders Goto Github PK

What's parameter to change to train the model on different data set than Minst etc..

I guess you've not implemented the part which was related to making $30 \times 30$ girds, because the connections are fully connected. In the paper, it was pointed out that for the affine transformations, we need smaller patch connections. Anyway, I wanna report a bug. Based on linear algebra and the paper itself the following line of code should be reversed:

learnt_transformation_extended = tf.matmul(learnt_transformation, self.extra_input)

extra_input should be multiplied from left. The correct one:

learnt_transformation_extended = tf.matmul(self.extra_input, learnt_transformation)