└── res2net_block.py


/res2net_block.py:
--------------------------------------------------------------------------------
 1 | from keras.layers import Conv2D, GlobalAveragePooling2D, Dense, BatchNormalization, Activation, Lambda, Concatenate, Add, Multiply
 2 | 
 3 | def res2net_bottleneck_block(x, f, s=4, expansion=4, use_se_block=False):
 4 |     """
 5 |     Arguments:
 6 |         x: input tensor
 7 |         f: number of output  channels
 8 |         s: scale dimension
 9 |     """
10 |     
11 |     num_channels = int(x._keras_shape[-1])
12 |     assert num_channels % s == 0, f"Number of input channel should be a multiple of s. Received nc={num_channels} and s={s}."
13 |     
14 |     input_tensor = x
15 |     
16 |     # Conv 1x1
17 |     x = BatchNormalization()(x)
18 |     x = Activation('relu')(x)
19 |     x = Conv2D(f, 1, kernel_initializer='he_normal', use_bias=False)(x)
20 |     
21 |     # Conv 3x3
22 |     subset_x = []
23 |     n = f
24 |     w = n // s
25 |     for i in range(s):
26 |         slice_x = Lambda(lambda x: x[..., i*w:(i+1)*w])(x)
27 |         if i > 1:
28 |             slice_x = Add()([slice_x, subset_x[-1]])
29 |         if i > 0:
30 |             slice_x = BatchNormalization()(slice_x)
31 |             slice_x = Activation('relu')(slice_x)
32 |             slice_x = Conv2D(w, 3, kernel_initializer='he_normal', padding='same', use_bias=False)(slice_x)        
33 |         subset_x.append(slice_x)
34 |     x = Concatenate()(subset_x)
35 |         
36 |     # Conv 1x1
37 |     x = BatchNormalization()(x)
38 |     x = Activation('relu')(x)
39 |     x = Conv2D(f*expansion, 1, kernel_initializer='he_normal', use_bias=False)(x)
40 |     
41 |     if use_se_block:
42 |         x = se_block(x)
43 |     
44 |     # Add
45 |     if num_channels == f*expansion:
46 |         skip = input_tensor
47 |     else:
48 |         skip = input_tensor
49 |         skip = Conv2D(f*expansion, 1, kernel_initializer='he_normal')(skip)
50 |     out = Add()([x, skip])
51 |     return out
52 | 
53 | def se_block(input_tensor, c=16):
54 |     num_channels = int(input_tensor._keras_shape[-1]) # Tensorflow backend
55 |     bottleneck = int(num_channels // c)
56 |  
57 |     se_branch = GlobalAveragePooling2D()(input_tensor)
58 |     se_branch = Dense(bottleneck, use_bias=False, activation='relu')(se_branch)
59 |     se_branch = Dense(num_channels, use_bias=False, activation='sigmoid')(se_branch)
60 |     
61 |     out = Multiply()([input_tensor, se_branch]) 
62 |     return out


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