how to split the feature in two

[EtheneXiang]

how to split the feature in two in your paper i see, the frist feture map was splitted in two, one for concat, and another for conv and copy. But I didn't find any relevant information in your train cfg file: csresnext50-panet-spp-original-optimal.cfg @WongKinYiu @AlexeyAB

[WongKinYiu]

There are two equivalent implementation.

1. separate it to two convolution

   
   # 1-1
   [convolutional]
   batch_normalize=1
   filters=64
   size=1
   stride=1
   pad=1
   activation=leaky

[route] layers = -2

1-2

[convolutional] batch_normalize=1 filters=64 size=1 stride=1 pad=1 activation=leaky

2. split one convolution into two parts

[convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky

1-1

[route] layers = -1 group_id=0 groups=2

1-2

[route] layers = -2 group_id=1 groups=2

Our old cfg use the first implementation, and new cfg use the second implementation.
Very thanks @AlexeyAB for supporting the second implementation.

[EtheneXiang]

@WongKinYiu thanks but I think you said is this condition:

but what i mean and what i want is this :

after 1x1x128 conv

the out feature map is nxnx128, and i want split the 128-feature in two, one half for OpreationA, and another half for OpreationB

[WongKinYiu]

Hello, the second case is what you want.

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

# 1-1
[route]
layers = -1
group_id=0
groups=2

# 1-2
[route]
layers = -2
group_id=1
groups=2

The parameter groups=2 means separate 128 to 64/64.

[EtheneXiang]

oh really?, it's great. but i do not find this uasge in your or AlexeyAB's github, i cheacked the cfg file just now. can you show me some example cfg aobut the the second case? thanks very much

[WongKinYiu]

@EtheneXiang Hello,

This is an example https://github.com/WongKinYiu/CrossStagePartialNetworks/blob/master/cfg/csresnext50-elastic.cfg#L65

[EtheneXiang]

thank you and i have another porblem, do you have tools to visualize the cfg. It's troublesome to red with text editor such as vim in linux or notepad++ in windows.

[WongKinYiu]

no, i usually draw the architecture by myself.

[EtheneXiang]

OK Anyway, thank you very much

[AlexeyAB]

@EtheneXiang

and i have another porblem, do you have tools to visualize the cfg.

You can try: https://github.com/lutzroeder/netron

[syjeon121]

@WongKinYiu hi, thanks for your explanation but i am still confused about separate feature map

There are two equivalent implementation.

1. separate it to two convolution

   
   # 1-1
   [convolutional]
   batch_normalize=1
   filters=64
   size=1
   stride=1
   pad=1
   activation=leaky

[route] layers = -2

1-2

[convolutional] batch_normalize=1 filters=64 size=1 stride=1 pad=1 activation=leaky

2. split one convolution into two parts

[convolutional] batch_normalize=1 filters=128 size=1 stride=1 pad=1 activation=leaky

1-1

[route] layers = -1 group_id=0 groups=2

1-2

[route] layers = -2 group_id=1 groups=2

as stated in your paper, does the first method divide the feature map into two flow like the second method?

[WongKinYiu]

they are two equivalent implementation of same method.

[syjeon121]

oh, sorry i made mistake. thank you

[WongKinYiu]

[abhiagwl4262]

@WongKinYiu It seems.. The concept of splitting the input into two parts is exactly same as group convolution ? Please explain If I am wrong.

[dengxinlong]

no, i usually draw the architecture by myself.

Hello, can you tell me what tool you used to draw the architecture. thank you!

[brandhsu]

Hi, I was wondering if the main benefit of splitting the channels in the CSPBlock is computation.