Capsule-Net-mnist" Architecture by Offir Shmulevich
Train acc: 0.991734693878 Test(Unseen data) acc: 0.991666666667
input_1 (InputLayer) | (None, 28, 28, 1) | 0
__________________________________________________________________________________________________
conv1 (Conv2D) | (None, 20, 20, 128)| 10496 | input_1[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) | (None, 6, 6, 128) | 1327232 | conv1[0][0]
__________________________________________________________________________________________________
reshape_1 (Reshape) | (None, 576, 8) | 0 | conv2d_1[0][0]
__________________________________________________________________________________________________
lambda_1 (Lambda) | (None, 576, 8) | 0 | reshape_1[0][0]
__________________________________________________________________________________________________
digitcaps (CapsuleLayer) | (None, 10, 16) | 743040 | lambda_1[0][0]
__________________________________________________________________________________________________
input_2 (InputLayer) | (None, 10) | 0 |
__________________________________________________________________________________________________
mask_1 (Mask) | (None, 16) | 0 | digitcaps[0][0]
input_2[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) | (None, 512) | 8704 | mask_1[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) | (None, 1024) | 525312 | dense_1[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) | (None, 784) | 803600 | dense_2[0][0]
__________________________________________________________________________________________________
out_caps (Length) | (None, 10) | 0 | digitcaps[0][0]
__________________________________________________________________________________________________
out_recon (Reshape) | (None, 28, 28, 1) | 0 | dense_3[0][0]
==================================================================================================
Total params: 3,418,384
Trainable params: 3,412,624
Non-trainable params: 5,760