├── README.md
└── GloRe.py


/README.md:
--------------------------------------------------------------------------------
1 | # GloRe
2 | Tensorflow implementation of Global Reasoning unit (GloRe) from Graph-Based Global Reasoning Networks.
3 | 


--------------------------------------------------------------------------------
/GloRe.py:
--------------------------------------------------------------------------------
 1 | import tensorflow as tf
 2 | slim = tf.contrib.slim
 3 | 
 4 | def GloRe(X, C,  N, activation_fn=None):
 5 |     imput_chancel = X.get_shape().as_list()[-1]
 6 |     inputs_shape = tf.shape(X)
 7 | 
 8 |     B = slim.conv2d(X, N, [1, 1])
 9 |     B = tf.reshape(B, [inputs_shape[0], -1, N]) # [B, H*W, N]
10 | 
11 |     x_reduced = slim.conv2d(X, C, [1, 1])
12 |     x_reduced = tf.reshape(x_reduced, [inputs_shape[0], -1, C]) # [B,  H*W, C]
13 |     x_reduced = tf.transpose(x_reduced, perm=[0, 2, 1])  # [B, C, H*W]
14 | 
15 |     # [B, C, H * W] * [B, H*W, N] —>#[B, C, N]
16 |     v = tf.matmul(x_reduced, B) # [B, C, N]
17 |     v = tf.expand_dims(v, axis=1)  # [B, 1, C, N]
18 | 
19 |     def GCN(Vnode, nodeN, mid_chancel):
20 | 
21 |         net = slim.conv2d(Vnode, nodeN, [1, 1], ) # [B, 1, C, N]
22 | 
23 |         net = Vnode -net  #(I-Ag)V
24 | 
25 |         net = tf.transpose(net, perm=[0, 3, 1, 2]) # [B, N, 1, C]
26 | 
27 |         net = slim.conv2d(net, mid_chancel, [1, 1]) # [B, N, 1, C]
28 | 
29 |         return net
30 | 
31 |     z = GCN(v, N, C) # [B, N, 1, C]
32 |     z = tf.reshape(z, [inputs_shape[0], N, C])  # [B, N, C]
33 | 
34 |     # [B, H*W, N] * [B, N, C] => [B, H*W, C]
35 |     y = tf.matmul(B, z) # [B, H*W, C]
36 |     y = tf.expand_dims(y, axis=1)  #[B, 1, H*W, C]
37 |     y = tf.reshape(y, [inputs_shape[0], inputs_shape[1], inputs_shape[2], C])  # [B, H, W, C]
38 |     x_res = slim.conv2d(y, imput_chancel, [1, 1])
39 | 
40 |     return X + x_res
41 | 


--------------------------------------------------------------------------------