Question About Attentive Feature Network

[ccJia]

Hi Liu,
I want make sure the construct of AF-Net. Could you help me check it?
I modified the branch behind "ch_concat_3a_chconcat" layer and I just use L = 4.

And this is my prototxt for caffe .

layer {
name: "ch_concat_3a_chconcat"
type: "Concat"
bottom: "conv_3a_1x1"
bottom: "conv_3a_3x3"
bottom: "conv_3a_double_3x3_1"
bottom: "conv_3a_proj"
top: "ch_concat_3a_chconcat"
}

layer {
name: "attention_conv_3b_1x1"
type: "Convolution"
bottom: "ch_concat_3a_chconcat"
top: "attention_conv_3b_1x1"
convolution_param {
num_output: 4
kernel_size: 1
stride: 1
pad: 0
}
}

layer {
name: "slice_attention_conv_3b_1x1"
type: "Slice"
bottom: "attention_conv_3b_1x1"
top: "slice_attention_conv_3b_1x1_0"
top: "slice_attention_conv_3b_1x1_1"
top: "slice_attention_conv_3b_1x1_2"
top: "slice_attention_conv_3b_1x1_3"

slice_param {
axis: 1
slice_point: 1
slice_point: 2
slice_point: 3
slice_point: 4
}
}

layer
{
name: "attention_mul_feature_0"
type: "Eltwise"
bottom: "ch_concat_3a_chconcat"
bottom: "slice_attention_conv_3b_1x1_0"
top: "attention_mul_feature_0"
eltwise_param {
operation: PROD
}
}
layer
{
name: "attention_mul_feature_1"
type: "Eltwise"
bottom: "ch_concat_3a_chconcat"
bottom: "slice_attention_conv_3b_1x1_1"
top: "attention_mul_feature_1"
eltwise_param {
operation: PROD
}
}
layer
{
name: "attention_mul_feature_2"
type: "Eltwise"
bottom: "ch_concat_3a_chconcat"
bottom: "slice_attention_conv_3b_1x1_2"
top: "attention_mul_feature_2"
eltwise_param {
operation: PROD
}
}
layer
{
name: "attention_mul_feature_3"
type: "Eltwise"
bottom: "ch_concat_3a_chconcat"
bottom: "slice_attention_conv_3b_1x1_3"
top: "attention_mul_feature_3"
eltwise_param {
operation: PROD
}
}
layer {
name: "attention_3a_chconcat"
type: "Concat"
bottom: "attention_mul_feature_0"
bottom: "attention_mul_feature_1"
bottom: "attention_mul_feature_2"
bottom: "attention_mul_feature_3"
top: "attention_3a_chconcat"
}
Thank you.

[xh-liu]

Hi,
Your basic structure is correct, however there are some parts which is different from my implementation:

1. The layers from attention_mul_feature_0 to attention_mul_feature_3 should be element wise multiplication of ch_concat_3a_chconcat and slice_attention_conv_3b_1x1_tile, where slice_attention_conv_3b_1x1_tile is tiled to be have the same number of channels as ch_concat_3a_chconcat. Otherwise the dimension between ch_concat_3a_chconcat and slice_attention_conv_3b_1x1 mismatches, and there will be error in element wise production.
2. I did not concat attention_mul_feature_0 to attention_mul_feature_3 to get attention_3a_chconcat and pass through the following blocks, but just let concat attention_mul_feature_0 to concat attention_mul_feature_3 pass through the following blocks respectively.
   Hope this will help you!
   Best,
   Xihui

[ccJia]

Hi,

According to the snapshot above, the input "F" is [C,H,W] and the output attention map "a" is [L,H,W]. In this case , L, as your suggestions is 8.
My question is how to do the element-wise multiplication between "F" and "a"?
If I understand , I will get one slice of "a" and do the element-wise multiplication for each channel of "F". Or “a” is indeed [L*C,H,W], we could get L copies of attention maps and each with [C,H,W], then we can perform element-wise multiplication.

[ccJia]

@xh-liu Could you help us ? T-T

[Li1991]

Hi, have you re-implemented this paper? Can you give a prototxt example? Thank you very much! @ccJia

[ccJia]

@Li1991 I haven't finished it . The AF Net is confusing me.....And I don't know how to implement it.

[bilipa]

@ccJia
It may said: for each channel for attention map, use it to multiplicate with F.
It can see from Fig 4

[xh-liu]

@ccJia Yes, your understanding is right. We get one slice of "a" and do the element-wise multiplication for each channel of "F".

[ccJia]

@xh-liu Thank you ^-^ !

[hezhenjun123]

@xh-liu Hi,it seems that the @ccJia prototxt has some error else,i think the number of each output about the "a" element-wise multiplicate with the "F" is 24(38),and the total number output to the GAP is 72(243).Which means each of the "a" need element-wise multiplicate with the "F1,F2 and F3",is that right?

[hezhenjun123]

@xh-liu 24(3x8) and 72(24x3),sorry for the typewriting...

[bilipa]

@hezhenjun123 I think the GAP input is (24x3 + 1) which mean (hydra, plus)

[hezhenjun123]

@bilipa yeah, i think you are right!