lua – Torch,为什么我的人工神经网络总是预测零？

我在 Linux CentOS 7机器上使用Torch7.
我正在尝试将人工神经网络(ANN)应用于我的数据集,以解决二进制分类问题.我正在使用一个简单的多层感知器.

我正在使用以下火炬包：optim,torch.

问题是我的感知器总是预测零值(被归类为零的元素),我无法理解为什么……

这是我的数据集(“dataset_file.csv”).有34个功能和1个标签目标(最后一列,可能是0或1)：

0.55,1,0.29,0.46,0.67,0.37,0.41,0.08,0.47,0.23,0.13,0.82,0.25,0.04,0.52,0.33,0
0.65,0.64,0.02,0.32,0.18,0.2,0.38,0.24,0
0.34,0.5,0.55,0.06,0.15,0.51,0.22,0.6,0.42,1
0.46,0.14,0.17,0.1,0.94,0.65,0.75,0.3,0
0.55,0.03,0.16,0.12,0.73,0.54,0.44,0.35,0.11,0
0.67,0.71,0.74,0.69,0.27,0.61,0.48,1
0.52,0.21,0.01,0.34,0.85,0.05,0.36,0
0.58,0.57,0.19,0
0.66,0.07,0.45,0.92,0
0.39,0.31,0.81,0
0.26,0.26,0.43,0
0.96,0.63,0.86,0.72,0.53,0.4,0.09,0.8,0.28,0
0.6,0
0.72,0.78,0.68,0
0.56,0.56,0.49,0.62,0.76,0.88,1
0.61,0.58,0
0.59,0.87,0
0.74,0.93,0
0.64,1
0.36,0.79,0.59,0.7,1

这是我的Torch Lua代码：

-- add comma to separate thousands
function comma_value(amount)
  local formatted = amount
  while true do  
    formatted,k = string.gsub(formatted,"^(-?%d+)(%d%d%d)",'%1,%2')
    if (k==0) then
      break
    end
  end
  return formatted
end

-- function that computes the confusion matrix
function confusion_matrix(predictionTestVect,truthVect,threshold,printValues)

  local tp = 0
  local tn = 0
  local fp = 0
  local fn = 0
  local MatthewsCC = -2
  local accuracy = -2
  local arrayFPindices = {}
  local arrayFPvalues = {}
  local arrayTPvalues = {}
  local areaRoc = 0

  local fpRateVett = {}
  local tpRateVett = {}
  local precisionVett = {}
  local recallVett = {}

  for i=1,#predictionTestVect do

    if printValues == true then
      io.write("predictionTestVect["..i.."] = ".. round(predictionTestVect[i],4).."\ttruthVect["..i.."] = "..truthVect[i].." ");
      io.flush();
    end

    if predictionTestVect[i] >= threshold and truthVect[i] >= threshold then
      tp = tp + 1
      arrayTPvalues[#arrayTPvalues+1] = predictionTestVect[i]
      if printValues == true then print(" TP ") end
    elseif  predictionTestVect[i] < threshold and truthVect[i] >= threshold then
      fn = fn + 1
      if printValues == true then print(" FN ") end
    elseif  predictionTestVect[i] >= threshold and truthVect[i] < threshold then
      fp = fp + 1
      if printValues == true then print(" FP ") end
      arrayFPindices[#arrayFPindices+1] = i;
      arrayFPvalues[#arrayFPvalues+1] = predictionTestVect[i]  
    elseif  predictionTestVect[i] < threshold and truthVect[i] < threshold then
      tn = tn + 1
      if printValues == true then print(" TN ") end
    end
  end

    print("TOTAL:")
    print(" FN = "..comma_value(fn).." / "..comma_value(tonumber(fn+tp)).."\t (truth == 1) & (prediction < threshold)");
    print(" TP = "..comma_value(tp).." / "..comma_value(tonumber(fn+tp)).."\t (truth == 1) & (prediction >= threshold)\n");

    print(" FP = "..comma_value(fp).." / "..comma_value(tonumber(fp+tn)).."\t (truth == 0) & (prediction >= threshold)");
    print(" TN = "..comma_value(tn).." / "..comma_value(tonumber(fp+tn)).."\t (truth == 0) & (prediction < threshold)\n");

  local continueLabel = true

    if continueLabel then
      upperMCC = (tP*tn) - (fP*fn)
      innerSquare = (tp+fp)*(tp+fn)*(tn+fp)*(tn+fn)
      lowerMCC = math.sqrt(innerSquare)

      MatthewsCC = -2
      if lowerMCC>0 then MatthewsCC = upperMCC/lowerMCC end
      local signedMCC = MatthewsCC
      print("signedMCC = "..signedMCC)

      if MatthewsCC > -2 then print("\n::::\tMatthews correlation coefficient = "..signedMCC.."\t::::\n");
      else print("Matthews correlation coefficient = NOT computable");  end

      accuracy = (tp + tn)/(tp + tn +fn + fp)
      print("accuracy = "..round(accuracy,2).. " = (tp + tn) / (tp + tn +fn + fp) \t  \t [worst = -1,best =  +1]");

      local f1_score = -2
      if (tp+fp+fn)>0 then   
    f1_score = (2*tp) / (2*tp+fp+fn)
    print("f1_score = "..round(f1_score,2).." = (2*tp) / (2*tp+fp+fn) \t [worst = 0,best = 1]");
      else
    print("f1_score CANNOT be computed because (tp+fp+fn)==0")    
      end

      local totalRate = 0
      if MatthewsCC > -2 and f1_score > -2 then 
    totalRate = MatthewsCC + accuracy + f1_score 
    print("total rate = "..round(totalRate,2).." in [-1,+3] that is "..round((totalRate+1)*100/4,2).."% of possible correctness");
      end

      local numberOfPredictedOnes = tp + fp;
      print("numberOfPredictedOnes = (TP + FP) = "..comma_value(numberOfPredictedOnes).." = "..round(numberOfPredictedOnes*100/(tp + tn + fn + fp),2).."%");

      io.write("\nDiagnosis: ");
      if (fn >= tp and (fn+tp)>0) then print("too many FN false negatives"); end
      if (fp >= tn and (fp+tn)>0) then print("too many FP false positives"); end


      if (tn > (10*fp) and tp > (10*fn)) then print("Excellent ! ! !");
      elseif (tn > (5*fp) and tp > (5*fn)) then print("Very good ! !"); 
      elseif (tn > (2*fp) and tp > (2*fn)) then print("Good !"); 
      elseif (tn >= fp and tp >= fn) then print("Alright"); 
      else print("Baaaad"); end
    end

    return {accuracy,arrayFPindices,arrayFPvalues,MatthewsCC};
end


-- Permutations
-- tab = {1,2,3,4,5,6,7,8,9,10}
-- permute(tab,10,10)
function permute(tab,n,count)
      n = n or #tab
      for i = 1,count or n do
        local j = math.random(i,n)
        tab[i],tab[j] = tab[j],tab[i]
      end
      return tab
end

-- round a real value
function round(num,idp)
  local mult = 10^(idp or 0)
  return math.floor(num * mult + 0.5) / mult
end



-- ##############################3

local profile_vett = {}
local csv = require("csv")
local fileName = "dataset_file.csv" 

print("Readin' "..tostring(fileName))
local f = csv.open(fileName)
local column_names = {}

local j = 0
for fields in f:lines() do

  if j>0 then
    profile_vett[j] = {}
      for i,v in ipairs(fields) do 
    profile_vett[j][i] = tonumber(v);
      end
    j = j + 1
  else
    for i,v in ipairs(fields) do 
    column_names[i] = v
     end
    j = j + 1
  end
end

OPTIM_PACKAGE = true
local output_number = 1
THRESHOLD = 0.5 -- ORIGINAL
DROPOUT_FLAG = false
MOMENTUM = false
MOMENTUM_ALPHA = 0.5
MAX_MSE = 4
LEARN_RATE = 0.001
IteraTIONS = 100

local hidden_units = 2000
local hidden_layers = 1

local hiddenUnitVect = {2000,4000,6000,8000,10000}
-- local hiddenLayerVect = {1,5}
local hiddenLayerVect = {1}

local profile_vett_data = {}
local label_vett = {}

for i=1,#profile_vett do
  profile_vett_data[i] = {}

  for j=1,#(profile_vett[1]) do  
    if j<#(profile_vett[1]) then
      profile_vett_data[i][j] = profile_vett[i][j]
    else
      label_vett[i] = profile_vett[i][j]
    end    
  end
end

print("Number of value profiles (rows) = "..#profile_vett_data);
print("Number features (columns) = "..#(profile_vett_data[1]));
print("Number of targets (rows) = "..#label_vett);

local table_row_outcome = label_vett
local table_rows_vett = profile_vett

-- ########################################################

-- START

local indexVect = {}; 
for i=1,#table_rows_vett do indexVect[i] = i;  end
permutedindexVect = permute(indexVect,#indexVect,#indexVect);

TEST_SET_PERC = 20
local test_set_size = round((TEST_SET_PERC*#table_rows_vett)/100)

print("training_set_size = "..(#table_rows_vett-test_set_size).." elements");
print("test_set_size = "..test_set_size.." elements\n");

local train_table_row_profile = {}
local test_table_row_profile = {}
local original_test_indexes = {}

for i=1,#table_rows_vett do
  if i<=(tonumber(#table_rows_vett)-test_set_size) then
    train_table_row_profile[#train_table_row_profile+1] = {torch.Tensor(table_rows_vett[permutedindexVect[i]]),torch.Tensor{table_row_outcome[permutedindexVect[i]]}}
  else

    original_test_indexes[#original_test_indexes+1] = permutedindexVect[i];

    test_table_row_profile[#test_table_row_profile+1] = {torch.Tensor(table_rows_vett[permutedindexVect[i]]),torch.Tensor{table_row_outcome[permutedindexVect[i]]}}
  end
end

require 'nn'
perceptron = nn.Sequential()
input_number = #table_rows_vett[1]

perceptron:add(nn.Linear(input_number,hidden_units))
perceptron:add(nn.Sigmoid())
if DROPOUT_FLAG==true then perceptron:add(nn.Dropout()) end

for w=1,hidden_layers do
  perceptron:add(nn.Linear(hidden_units,hidden_units))
  perceptron:add(nn.Sigmoid())
  if DROPOUT_FLAG==true then perceptron:add(nn.Dropout()) end
end
perceptron:add(nn.Linear(hidden_units,output_number))


function train_table_row_profile:size() return #train_table_row_profile end 
function test_table_row_profile:size() return #test_table_row_profile end 


-- OPTIMIZATION LOOPS  
local MCC_vect = {}  

for a=1,#hiddenUnitVect do
  for b=1,#hiddenLayerVect do

    local hidden_units = hiddenUnitVect[a]
    local hidden_layers = hiddenLayerVect[b]
    print("hidden_units = "..hidden_units.."\t output_number = "..output_number.." hidden_layers = "..hidden_layers)


    local criterion = nn.MSECriterion()  
    local lossSum = 0
    local error_progress = 0

      require 'optim'
      local params,gradParams = perceptron:getParameters()     
      local optimstate = nil

      if MOMENTUM==true then 
    optimstate = {learningRate = LEARN_RATE}
      else 
    optimstate = {learningRate = LEARN_RATE,momentum = MOMENTUM_ALPHA }
      end

      local total_runs = IteraTIONS*#train_table_row_profile

      local loopIterations = 1
      for epoch=1,IteraTIONS do
    for k=1,#train_table_row_profile do

        -- Function feval 
        local function feval(params)
        gradParams:zero()

        local thisProfile = train_table_row_profile[k][1]
        local thisLabel = train_table_row_profile[k][2]

        local thisPrediction = perceptron:forward(thisProfile)
        local loss = criterion:forward(thisPrediction,thisLabel)

        -- print("thisPrediction = "..round(thisPrediction[1],2).." thisLabel = "..thisLabel[1])

        lossSum = lossSum + loss
        error_progress = lossSum*100 / (loopIterations*MAX_MSE)

        if ((loopIterations*100/total_runs)*10)%10==0 then
          io.write("completion: ",round((loopIterations*100/total_runs),2).."%" )
          io.write(" (epoch="..epoch..")(element="..k..") loss = "..round(loss,2).." ")      
          io.write("\terror progress = "..round(error_progress,5).."%\n")
        end

        local dloss_doutput = criterion:backward(thisPrediction,thisLabel)

        perceptron:backward(thisProfile,dloss_doutput)

        return loss,gradParams
        end
      optim.sgd(feval,params,optimstate)
      loopIterations = loopIterations+1
    end     
      end


    local correctPredictions = 0
    local atleastOneTrue = false
    local atleastOneFalse = false
    local predictionTestVect = {}
    local truthVect = {}

    for i=1,#test_table_row_profile do
      local current_label = test_table_row_profile[i][2][1]
      local prediction = perceptron:forward(test_table_row_profile[i][1])[1]

      predictionTestVect[i] = prediction
      truthVect[i] = current_label

      local labelResult = false

      if current_label >= THRESHOLD and prediction >= THRESHOLD  then
    labelResult = true
      elseif current_label < THRESHOLD and prediction < THRESHOLD  then
    labelResult = true
      end

      if labelResult==true then correctPredictions = correctPredictions + 1; end

    print("\nCorrect predictions = "..round(correctPredictions*100/#test_table_row_profile,2).."%")

     local printValues = false
     local output_confusion_matrix = confusion_matrix(predictionTestVect,THRESHOLD,printValues)
  end
end

有没有人知道为什么我的脚本只预测零元素？

编辑：我用原始数据集替换了我在脚本中使用的规范化版本

解决方法

当我运行您的原始代码时,我有时会预测所有零,我有时会获得完美的性能.这表明您的原始模型对参数值的初始化非常敏感.

如果我使用种子值torch.manualSeed(0)(所以我们总是有相同的初始化),我每次都会得到完美的表现.但这不是一般的解决方案.

为了获得更全面的改进,我做了以下更改：

>减少隐藏单位的数量.在原始代码中你有一个
单个隐藏层的2000个单位.但是你只有34个输入和
1输出通常你只需要隐藏单位的数量
输入和输出的数量之间.我减少了它
50.
>标签是不对称的,只有5/27(19％)的标签是1,所以你应该真正划分列车|测试集,以保持1与0的比率.目前我只是将测试集大小增加到’50’％.
>我也将学习率提高到’0.01′,开启MOMENTUM,并将IteraTIONS增加到200.

当我运行这个模型20次(未播种)时,我获得了19次优异的表现.为了进一步改进,您可以进一步调整超参数.并且还应该使用单独的验证集来查看多个初始化,以选择“最佳”模型(尽管这将进一步细分已经非常小的数据集).

-- add comma to separate thousands
function comma_value(amount)
  local formatted = amount
  while true do  
    formatted,v in ipairs(fields) do 
    column_names[i] = v
     end
    j = j + 1
  end
end

OPTIM_PACKAGE = true
local output_number = 1
THRESHOLD = 0.5 -- ORIGINAL
DROPOUT_FLAG = false
MOMENTUM_ALPHA = 0.5
MAX_MSE = 4

-- CHANGE: increased learn_rate to 0.01,reduced hidden units to 50,turned momentum on,increased iterations to 200
LEARN_RATE = 0.01
local hidden_units = 50
MOMENTUM = true
IteraTIONS = 200
-------------------------------------

local hidden_layers = 1

local hiddenUnitVect = {2000,#(profile_vett[1]) do  
    if j<#(profile_vett[1]) then
      profile_vett_data[i][j] = profile_vett[i][j]
    else
      label_vett[i] = profile_vett[i][j]
    end    
  end
end

print("Number of value profiles (rows) = "..#profile_vett_data);
print("Number features (columns) = "..#(profile_vett_data[1]));
print("Number of targets (rows) = "..#label_vett);

local table_row_outcome = label_vett
local table_rows_vett = profile_vett

-- ########################################################

-- START

-- Seed random number generator
-- torch.manualSeed(0)

local indexVect = {}; 
for i=1,#indexVect);

-- CHANGE: increase test_set to 50%
TEST_SET_PERC = 50
---------------------------

local test_set_size = round((TEST_SET_PERC*#table_rows_vett)/100)

print("training_set_size = "..(#table_rows_vett-test_set_size).." elements");
print("test_set_size = "..test_set_size.." elements\n");

local train_table_row_profile = {}
local test_table_row_profile = {}
local original_test_indexes = {}

for i=1,printValues)
  end
end
end

下面粘贴的是20次运行中的1次输出：

Correct predictions = 100%  
TOTAL:  
 FN = 0 / 4  (truth == 1) & (prediction < threshold)    
 TP = 4 / 4  (truth == 1) & (prediction >= threshold)

 FP = 0 / 9  (truth == 0) & (prediction >= threshold)   
 TN = 9 / 9  (truth == 0) & (prediction < threshold)

signedMCC = 1   

::::    Matthews correlation coefficient = 1    ::::

accuracy = 1 = (tp + tn) / (tp + tn +fn + fp)        [worst = -1,best =  +1]   
f1_score = 1 = (2*tp) / (2*tp+fp+fn)     [worst = 0,best = 1]  
total rate = 3 in [-1,+3] that is 100% of possible correctness 
numberOfPredictedOnes = (TP + FP) = 4 = 30.77%  

Diagnosis: Excellent ! ! !