need help

[tuteming]

by
https://pjreddie.github.io/uwnet/slides/13#3
this is a classifier config
however, by
https://pjreddie.github.io/uwnet/slides/13#11
that is a Discriminator config

I write as follows:
[net]
batch=4
subdivisions=1
height=256
width=256
channels=3

momentum=0.9
decay=0.0001
learning_rate=0.0001
max_batches = 8000
policy=steps
steps=6400,7200
scales=0.1,0.1

#1
[convolutional]
batch_normalize=1
filters=16
size=3
stride=1
pad=1
activation=leaky

#2
[maxpool]
size=2
stride=2

#3
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky

#4
[maxpool]
size=2
stride=2

#5
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky

#6
[maxpool]
size=2
stride=2

#7
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky

#8
[maxpool]
size=2
stride=2

#9
[avgpool]

[connected]
output=2
activation=linear

[logistic]

//end
///////////////////
In the last layers, what I miss?
please correct it for me. Very thanks!

note: the last [logistic] layer is not in AlexeyAB but has in pjreddie.
I have add [logistic] layer to AlexeyAB version.

[WongKinYiu]

maybe add:

[cost]
sse

information is not sufficient to solve the problem. what is your command? what is the error? how you add logistic layer? ...

[tuteming]

首先謝謝你的快速回覆 上面的鏈結是pjreddie說明GAN的投影片(可用-> or <- 移動來看)
在pix2pix 中 他用的鑑別器在第2個鏈結(第1個鏈結則是一個簡單的分類器)給你對應參考.
logistic層在AlexeyAB version沒有,必須由pjreddie version拷貝,加在相關的地方.
pix2pix源自於pjreddie version darknet/examples/lsd.c 中,不過這版本不全

https://github.com/vaiv/OpenANPR/blob/master/distribute/src/lsd.c 比較完整
對應的東西也是由這裡的src加在AlexeyAB version相關的地方,而不是由pjreddie version.

第2個鏈結鑑別器其實是一個binary classifier,最後2層[connected]與logistic
logistic is binary cross-entropy, softmax is then cross-entropy .
生成器我用的是modify Unet (單獨用確認無誤),鑑別器最後2層我寫得有問題,所以loss=0. 導致生成器loss收斂的有問題
所以想請你幫忙看看.最後2層該怎麼設定, 真是麻煩你了. 謝謝

[tuteming]

鑑別器用binary classifier and binary cross-entropy 的目的
根據pjreddie的投影片說明:鑑別生成影像的好壞, 來修正生成器. (見第2個鏈結左邊說明)

[WongKinYiu]

您好

請問您的訓練指令是什麼, 遇到的錯誤是什麼, 在pireddie與AlexeyAB中加入新layer的方式並不一樣, 您是怎麼加入logistic layer的.
此外, 程式在pireddie運行順暢嗎.

[tuteming]

程式在pireddie運行順暢嗎?
run on AlexeyAB win 10.運行順暢 But Alos=0.

怎麼加入logistic layer的

1. in darknet.h
   add:
   typedef enum {
   ...
   LOGXENT,
   ...
   } LAYER_TYPE;

2. in parser.c
   add:
   #include "logistic_layer.h"

add:
LAYER_TYPE string_to_layer_type(char * type)
{
...
if (strcmp(type, "[logistic]")==0) return LOGXENT;//why LOGXENT see logistic_layer.c
...
}

add:
layer parse_logistic(list *options, size_params params)
{
layer l = make_logistic_layer(params.batch, params.inputs);
l.h = l.out_h = params.h;
l.w = l.out_w = params.w;
l.c = l.out_c = params.c;
return l;
}

in line 144x
add:
}else if(lt == LOGXENT){
l = parse_local(options, params);

from P-version
copy logistic_layer.c and logistic_layer.h
to A-version
but
line 57 copy_gpu->copy_ongpu
line 58 activate_array_gpu->activate_array_ongpu
line 68 axpy_gpu->axpy_ongpu

訓練指令:ref lsd.c
darknet.exe lsd_1 train2 data/data.txt cfg/gnet.cfg cfg/anet.cfg -clear

[WongKinYiu]

AlexeyAB darknet 沒有 lsd.c. 相關的資訊量實在太少了, 無法釐清問題.