확률적인 경사 하강 훈련 오류

Question

전 컨볼 루션 네트워크를 사용하고 있습니다. 내 이미지의 크기는 512x512입니다. 다음은 데이터 세트 형성 코드입니다.

다음

unet = nn.Sequential() 
-- ndim=input:size(2) 
ndim=512 
unet:add(nn.SpatialReflectionPadding(1+ndim/2,1+ndim/2,1+ndim/2,1+ndim/2)) --reflection padding 
unet:add(nn.SpatialConvolution(1, 64, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(64, 64, 3, 3)) 
unet:add(nn.SpatialMaxPooling(2,2,2,2))  

unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(64, 128, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(128,128, 3, 3)) 
unet:add(nn.SpatialMaxPooling(2,2,2,2)) 

unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(128, 256, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(256, 256, 3, 3)) 
unet:add(nn.SpatialMaxPooling(2,2,2,2)) 

unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(256, 512, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(512, 512, 3, 3)) 
unet:add(nn.SpatialMaxPooling(2,2,2,2)) 

unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(512, 1024, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(1024, 1024, 3, 3)) 
unet:add(nn.SpatialMaxPooling(2,2,2,2)) 

unet:add(nn.SpatialFullConvolution(1024, 1024, 2, 2, 2, 2)) -- up convolve 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(1024, 512, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(512, 512, 3, 3)) 

unet:add(nn.SpatialFullConvolution(512, 512, 2, 2, 2, 2)) -- up convolve 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(512, 256, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(256, 256, 3, 3)) 

unet:add(nn.SpatialFullConvolution(256, 256, 2, 2, 2, 2)) -- up convolve 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(256, 128, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(128, 128, 3, 3)) 

unet:add(nn.SpatialFullConvolution(128, 128, 2, 2, 2, 2)) -- up convolve 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(128, 64, 3, 3)) 
unet:add(nn.SpatialReflectionPadding(1,1,1,1)) --reflection padding 
unet:add(nn.SpatialConvolution(64, 64, 3, 3)) 
unet:add(nn.SpatialConvolution(64, 8, 1, 1)) --output size 

내가 그것을 훈련하고 방법은 다음과 같습니다 :

require 'image' 
require 'nn 
npy4th = require 'npy4th' 

function scandir(directory) 
    local i, t, popen = 0, {}, io.popen 
    local pfile = popen('ls -a "'..directory..'"') 
    for filename in pfile:lines() do 
     i = i + 1 
     t[i] = filename 
    end 
    pfile:close() 
    return t 
end 

dataset={} 

path1='/home/dataset/train/data/' 

path2='/home/dataset/train/label/' 

x=scandir(path1) 
table.remove(x,1) 
table.remove(x,1) 

y=scandir(path2) 
table.remove(y,1) 
table.remove(y,1) 

num_files=0 
for _ in pairs(x) do 
    num_files= num_files + 1 
end 

num_files=2 
--dataset of form input, output creation 
function dataset:size() return num_files end 

for i=1, dataset:size() do 
    data=npy4th.loadnpy(path1..x[i]) 
    label=npy4th.loadnpy(path2..y[i]) 
    output=torch.Tensor(8,data:size(1),data:size(2)):zero() 
    output[1][label:eq(1)]=1 --healthy 
    output[2][label:eq(2)]=2 --emphysema 
    output[3][label:eq(3)]=3 --GGO              
    output[4][label:eq(4)]=4 --fibrosis 
    output[4][label:eq(14)]=4 --early_fibrosis 
    output[5][label:eq(5)]=5 --micronodules 
    output[6][label:eq(6)]=6 --consolidation 
    output[7][label:eq(8)]=7 --reticulation 
    output[8][label:eq(11)]=8 --peripheral_micronodules 
    input=torch.reshape(data,1,data:size(1),data:size(2)) 
    dataset[i]={input:double(),output}; 
end 

I는 다음과 성화 (7)의 완전 컨볼 루션 네트워크를 구현하고 교육에

criterion=nn.CrossEntropyCriterion() 
trainer=nn.StochasticGradient(unet,criterion) 
trainer.learningRate = 0.01 
trainer:train(dataset) 

나는 다음과 같은 오류가 발생합니다 :

# StochasticGradient: training 
/home/torch/install/share/lua/5.1/nn/THNN.lua:109: bad argument #2 to 'v' (vector or matrix expected at /home/SparseConvolution/nn/lib/THNN/generic/LogSoftMax.c:26) 
stack traceback: 
    [C]: in function 'v' 
    /home/torch/install/share/lua/5.1/nn/THNN.lua:109: in function 'LogSoftMax_updateOutput' 
    ...e/torch/install/share/lua/5.1/nn/LogSoftMax.lua:4: in function 'updateOutput' 
    ...torch/install/share/lua/5.1/nn/CrossEntropyCriterion.lua:12: in function 'forward' 
    ...le/torch/install/share/lua/5.1/nn/StochasticGradient.lua:35: in function 'train' 
    [string "_RESULT={trainer:train(dataset)}"]:1: in main chunk 
    [C]: in function 'xpcall' 
    /home/torch/install/share/lua/5.1/trepl/init.lua:651: in function 'repl' 
    ...me/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:199: in main chunk 
    [C]: at 0x00406670 

내가 토치를 사용하는 첫 번째 사람입니다, 내가 누락 된 것을 누군가 도울 수 있습니까? 감사.

Answer 1

사용중인 기준 CrossEntropyCriterion()은 모듈 LogSoftMax과 ClassNLLCriterion을 결합합니다. yourModel:forward()은 크기가 64 * 8 인 매트릭스를 반환하며이 기준 (ClassNLLCriterion)은 특히 1D Tensor 만 허용합니다. CrossEntropyCriterion