MNIST classification



In [1]:

    
require 'nn';
require 'rnn';
matio = require 'matio'



In [2]:

    
data = matio.load('ex4data1.mat')
trainset = {}
trainset.data = data.X
trainset.label = data.y[{ {}, 1 }]



In [4]:

    
setmetatable(trainset,
    {__index = function(t,i)
                return {t.data[i], t.label[i]}
        end}
);
 
function trainset:size()
    return self.data:size(1)
end

Data Normalization



In [6]:

    
mean = {}
stdv = {}
for i=1,400 do
    mean[i] = trainset.data[{ {},{i} }]:mean()
    stdv[i] = trainset.data[{ {}, {i} }]:std()
    --print(i .. 'th mean: ' .. mean[i])
    --print(i .. 'th std dev: ' .. stdv[i])
    trainset.data[{ {},{i} }]:add(-mean[i])
    if stdv[i] ~= 0 then
        trainset.data[{ {},{i} }]:div(stdv[i])
    end
end

define model



In [7]:

    
batchSize = 5000
rho = 20
hiddenSize = 10
nIndex = 20
nClass = 10



In [8]:

    
rnn = nn.Sequential()
r = nn.Recurrent(
    hiddenSize, nn.Linear(nIndex, hiddenSize), --nn.Linear(nIndex, hiddenSize) = W_hx
    nn.Linear(hiddenSize, hiddenSize), nn.Tanh(), --nn.Linear(hiddenSize, hiddenSize) = W_hh
    rho
)
rnn:add(r)
rnn:add(nn.Linear(hiddenSize, nClass))
rnn:add(nn.LogSoftMax())
rnn = nn.Sequencer(rnn)



In [9]:

    
criterion = nn.SequencerCriterion(nn.ClassNLLCriterion())

train



In [10]:

    
lr = 0.1
i = 1



In [11]:

    
prev = 100
for epoch = 1,1e3 do
    local inputs, targets = {}, {}
    for step=1,rho do -- 1 ~ 20
        table.insert(inputs, trainset.data[{{},{(step-1)*rho+1, step*rho}}])
        table.insert(targets, trainset.label[{{}}])
    end

    local outputs = rnn:forward(inputs)
    local err = criterion:forward(outputs, targets)
    if epoch%10 == 1 then print(epoch, err/rho) end
    local gradOutputs = criterion:backward(outputs, targets)
    rnn:backward(inputs, gradOutputs)
    rnn:updateParameters(lr)
    rnn:zeroGradParameters()
end









    Out[11]:





1	2.3391271816177	







    Out[11]:





11	2.1955281583623	







    Out[11]:





21	2.037658252251	







    Out[11]:





31	1.9690607747917	







    Out[11]:





41	1.9593234961249	







    Out[11]:





51	1.9831929945857	







    Out[11]:





61	1.8620890182392	







    Out[11]:





71	1.9148388041667	






    Out[11]:












    Out[11]:





81	1.9557262654271	







    Out[11]:





91	1.8021422117858	







    Out[11]:





101	1.8188248065218	







    Out[11]:





111	1.717234843504	







    Out[11]:





121	1.7623135490069	







    Out[11]:





131	1.79423168286	







    Out[11]:





141	1.7726373670219	







    Out[11]:





151	1.9834949072924	







    Out[11]:





161	1.6997501949595	







    Out[11]:





171	1.6492612479463	







    Out[11]:





181	2.1208070556762	







    Out[11]:





191	1.6821686974183	







    Out[11]:





201	1.6318084080856	







    Out[11]:





211	1.6194401872758	







    Out[11]:





221	1.9092329085307	







    Out[11]:





231	1.856933600125	







    Out[11]:





241	1.8243470661334	







    Out[11]:





251	2.031291997915	







    Out[11]:





261	2.0166399091812	







    Out[11]:





271	1.8939151789164	







    Out[11]:





281	1.9844990549649	







    Out[11]:





291	1.8686006093978	







    Out[11]:





301	1.8129094266696	







    Out[11]:





311	1.8059165071629	







    Out[11]:





321	1.9269985363757	







    Out[11]:





331	1.6999124079312	







    Out[11]:





341	1.8645111286057	







    Out[11]:





351	2.5128108530264	







    Out[11]:





361	1.7570424130973	







    Out[11]:





371	1.8349937668285	







    Out[11]:





381	1.7687974118329	







    Out[11]:





391	1.8403859580368	







    Out[11]:





401	1.8246828940703	







    Out[11]:





411	1.7423012112847	







    Out[11]:





421	1.7294762827616	







    Out[11]:





431	1.8694934743739	







    Out[11]:





441	1.7682192822609	







    Out[11]:





451	1.7605444764079	







    Out[11]:





461	1.7484330770191	







    Out[11]:





471	1.7650141805605	







    Out[11]:





481	1.789059489947	







    Out[11]:





491	1.6934223480175	







    Out[11]:





501	1.6660960211912	







    Out[11]:





511	1.7443063285868	







    Out[11]:





521	1.7009674322367	







    Out[11]:





531	1.6986856051417	







    Out[11]:





541	1.6865454695006	







    Out[11]:





551	1.7411362033103	







    Out[11]:





561	1.6912126938625	







    Out[11]:





571	1.6347865508689	







    Out[11]:





581	1.7103972993512	







    Out[11]:





591	1.8859219735329	







    Out[11]:





601	1.7403357852309	







    Out[11]:





611	1.7748609068793	







    Out[11]:





621	1.7583254468828	







    Out[11]:





631	1.8857419940146	







    Out[11]:





641	1.7992770553966	







    Out[11]:





651	2.2019843001456	







    Out[11]:





661	2.1009637692277	







    Out[11]:





671	2.0553492381859	







    Out[11]:





681	2.1873517365684	







    Out[11]:





691	2.0910587881434	







    Out[11]:





701	2.0645586680785	







    Out[11]:





711	2.0411537749646	







    Out[11]:





721	2.0337288829184	







    Out[11]:





731	2.0386315002356	







    Out[11]:





741	2.0114692246773	







    Out[11]:





751	2.0042521928585	







    Out[11]:





761	2.076581891392	







    Out[11]:





771	2.0004553883316	







    Out[11]:





781	2.0174584730144	







    Out[11]:





791	1.9738821894372	







    Out[11]:





801	2.1543788260467	







    Out[11]:





811	






    Out[11]:





2.1238615624285	







    Out[11]:





821	2.1007468713526	







    Out[11]:





831	2.083947453716	







    Out[11]:





841	2.0535403511206	







    Out[11]:





851	1.9729857218067	







    Out[11]:





861	1.957627588988	







    Out[11]:





871	1.9522227621764	







    Out[11]:





881	1.948584517834	







    Out[11]:





891	1.9398604464103	







    Out[11]:





901	1.927979918414	







    Out[11]:





911	1.9267730202584	







    Out[11]:





921	1.9137384341865	







    Out[11]:





931	1.9048116611769	







    Out[11]:





941	1.9156521368639	







    Out[11]:





951	1.9037856805473	







    Out[11]:





961	1.8979352848381	







    Out[11]:





971	1.8926875283293	







    Out[11]:





981	1.8950950975504	







    Out[11]:





991	1.8761098743622

test



In [12]:

    
correction = {}
for i=10,20 do
    correction[i] = 0
end
for i=1,trainset:size() do
    local answer = trainset.label[i]    
    local inputs = {}
    for step=1,rho do
        table.insert(inputs, trainset.data[{{i},{(step-1)*20+1, step*20}}])
    end
    local prediction = rnn:forward(inputs)
    for d=10,20 do
        guess = prediction[d][{1,{}}]
        local confidences, indices = torch.sort(guess, true)
        -- if i%100 == 1 then print(answer, guess, indices[1]) end
        if (answer == indices[1]) then
            correction[d] = correction[d] + 1
        end
    end
end
for i=10,20 do
    print(i, " = ", correction[i], 100*correction[i]/trainset:size() .. '%')
end









    Out[12]:





10	






    Out[12]:





 = 	1331	26.62%	
11	 = 	1111	22.22%	
12	 = 	1149	22.98%	
13	 = 	1383	27.66%	
14	 = 	1540	30.8%	
15	 = 	1642	32.84%	
16	 = 	1684	33.68%	
17	 = 	1664	33.28%	
18	 = 	1658	33.16%	
19	 = 	1692	33.84%	
20	 = 	1519	30.38%



In [ ]: