Open ghost opened 6 years ago
I haven't checked this thoroughly, but as far as I can understand, meshgrid() initializes the flow to be neutral, instead of everybody picking out the same pixel. That's a much more reasonable starting point.
Note that the meshgrid() here is not the same as tf.meshgrid()! The code appears to be written for pre-1.0 TF, which probably didn't have tf.meshgrid().
Hi, I'm trying to play with your code because I'm fascinated by your result. However, the 'meshgrid' function in the model really confuse me, it seems like a fixed bias applied to the flow, but I cannot see the meaning of doing this. Can't this be trained with the previous layer instead of applying a fixed one?