danielrapp / cnn-gol Goto Github PK

Hi, I saw your post because I got interested by chance on the same topic. Spent last night coding a solution in TensorFlow. After thinking about it for a while I came to realize that 2 channels would be enough for the first 3x3 convolution:

• The first one would count the number of cells
• The second one would say if the cell in the middle was alive or not.

After that you just need a 1x1 convolution with many channels to encode the logic, and an output 1x1 convolution of 1 channel. Here is a snipped of the code

.convolution2d(2, [3, 3], scope='relu_conv_3x3', activation_fn=tf.nn.relu, padding='VALID')
.convolution2d(16, [1, 1], scope='relu_conv_1x1', activation_fn=tf.nn.sigmoid)
.convolution2d(1, [1, 1], scope='linear_conv_1x1_2', activation_fn=None),

The 1x1 convolutions worked as expected, since they are equivalent to a Fully Connected operation between channels per pixel with all parameters shared, those can be seen as the "regression" mechanism.

The solution worked great, it simulates perfectly infact :)