Torche, erreur dans l'entraînement StochasticGradient

Question

J'essaie de créer un réseau neuronal autoencoder dans Torch, mais j'ai une erreur quand j'essaie de lancer la phase d'entraînement.

Voilà mes données d'entrée dans data.csv:

chrName-chrStart-chrEnd,gm12878,h1-hesc,hela-s3,hepg2,huvec,k562 
    chr21-10861860-10862010,0.6904761905,0,0,1.2555555556,0.25,1.05 
    chr21-10896160-10896310,0.35,2.4934782609,0.8181818182,1.05,0.75,1.9611111111 
    chr21-10936080-10936230,1.525,0.2,0.525,1.0823529412,0.7409090909,0.85 
    chr21-10972180-10972330,0.7333333333,1,0.67,1.7227272727,0.9571428571,0.8823529412 
    chr21-11059120-11059270,3.7268292683,2.9347826087,1.3833333333,3.328358209,4.35,2.31 
    chr21-11101660-11101810,1.78,5.328125,1.05,0,2.9916666667,1.25 
    chr21-11181780-11181930,2.9542857143,4.3173913043,0.9666666667,2.65,4.772972973,1.01 
    chr21-14911600-14911750,0.7882352941,0,0,0,0.3,1.25 
    chr21-14973280-14973430,0.3,0.65,0,0.4,0.1,0 

Voici mon code

-- see if the file exists 
function file_exists(file) 
    local f = io.open(file, "rb") 
    if f then f:close() end 
    return f ~= nil 
end 

-- get all lines from a file, returns an empty 
-- list/table if the file does not exist 
function lines_from(file) 
    if not file_exists(file) then return {} end 
    lines = {} 
    for line in io.lines(file) do 
    lines[#lines + 1] = line 
    end 
    return lines 
end 


-- Torch program that creates a simple AutoEncoder-style neural network, where output=input 
print('\n\n @ @ @ @ @ @ START @ @ @ @ @ @ @ '); 
print('Torch program that creates a simple AutoEncoder-style neural network, where output=input'); 

require 'os'; 
require 'lfs'; 

file = 'data.csv'; 

print('selected file: ' .. file); 

chromRegions = {} 
cellTypes = {} 
dnaseIsignals = {} 
cr = 0 


chromRegion = {}; 
gm12878 = {}; 
h1hesc = {}; 
helas3 = {}; 
hepg2 = {}; 
huvec = {}; 
k562 = {}; 
dataset = {}; 

i = 0; 
CellTypesNumber = 6 

print('\n-- dataset building --'); 
-- data building 

local file1 = io.open(file) 
for line in file1:lines() do 
    if i >= 1 then 
     chromRegion[i-1], gm12878[i-1], h1hesc[i-1], helas3[i-1], hepg2[i-1], huvec[i-1], k562[i-1] = unpack(line:split(",")) 
    profile = torch.Tensor(CellTypesNumber); 
    profile = {tonumber(gm12878[i-1]), tonumber(h1hesc[i-1]), tonumber(helas3[i-1]), tonumber(hepg2[i-1]), tonumber(huvec[i-1]), tonumber(k562[i-1])} 
    dnaseIsignals[i-1] = profile; 
    profile = null; 
    end 
i = i + 1 
end 

chromRegionsNumber = #chromRegion+1 
print('chromRegionsNumber '..chromRegionsNumber); 

inputs=chromRegionsNumber; 
outputs=inputs; -- 

print(dnaseIsignals[0]) 
print(dnaseIsignals[chromRegionsNumber-1]) 

for i=1,chromRegionsNumber do -- 15 times 

    dataset[i] = {dnaseIsignals[i-1], dnaseIsignals[i-1]} 

    print(dataset[i]); 
end 
function dataset:size() return chromRegionsNumber end -- input dataset hgas 
function dataset:dim() return chromRegionsNumber end -- input dataset hgas 


HUs = 5; 

-- Creation of the neural network 
print('-- creation of the neural network --'); 
require "nn" 
mlp=nn.Sequential(); -- make a multi-layer perceptron 
mlp:add(nn.Linear(inputs,HUs)) -- adds a (input x HUs) layer to the network 
mlp:add(nn.Sigmoid()); 
mlp:add(nn.Linear(HUs,outputs)) -- adds a (HUs x outputs) layer to the network 
mlp:add(nn.Sigmoid()); -- Sigmoid also in output, Peter says 

-- Training 
print('-- training --'); 
criterion = nn.MSECriterion() 
trainer = nn.StochasticGradient(mlp, criterion) 
trainer.learningRate = 0.01 
trainer:train(dataset) --> THE PROBLEMS START HERE! :-(

-- Testing 
print('-- testing --'); 
geneProfileChosen = 0; 
x=torch.Tensor(inputs); -- create a test example Tensor. Test element: first gene 

Et voici mon erreur:

-- training -- 
# StochasticGradient: training 
/home/david/torch/install/bin/luajit: /home/david/torch/install/share/lua/5.1/nn/Linear.lua:34: attempt to call method 'dim' (a nil value) 
stack traceback: 
    /home/david/torch/install/share/lua/5.1/nn/Linear.lua:34: in function 'updateOutput' 
    /home/david/torch/install/share/lua/5.1/nn/Sequential.lua:25: in function 'forward' 
    ...id/torch/install/share/lua/5.1/nn/StochasticGradient.lua:35: in function 'train' 
    autoencoder_chromatin_interactions.lua:103: in main chunk 
    [C]: in function 'dofile' 
    ...avid/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:131: in main chunk 
    [C]: at 0x00405330 

Avez-vous une idée de pourquoi je 'm' gettin 'cette erreur?

Answer 1

dataset[i] = {dnaseIsignals[i-1], dnaseIsignals[i-1]} 

Ceci est faux. Ce que vous voulez donner en entrée au réseau de neurones est un torch.Tensor, mais dans ce cas, vous initialisez les éléments de l'ensemble de données pour qu'ils soient des tables. Convertissez votre entrée en un tenseur, et vous devriez être bon.