Addmiting different image size in model

[gonzalolc]

Hello I'm new in lua and i'm trying to input in the text-classification model a different image size since my dataset is not 1024 length. how can I do this Excuse me for my ignorance Thank you

[gheinrich]

Hello, you can adjust the feature length there. If you do this you may also have to update the model description there. Make note of the comments in the model description which are showing the shape of the activations at each layer. For a better understanding of this, check the Torch documentation for temporal convolutions.

[gonzalolc]

Thank you. I'll start working based on this info

[gonzalolc]

Hello @gheinrich and @lukeyeager . Thanks for your help I've been trying this model for my 64x64 image and when I run it in digits it gets stucked and does nothing. I have not modified the temporal convolution values, only the reshape and linear modules to fit the 4096 feature lenght. What can I do to fix it? Thanks

assert(pcall(function() require('dpnn') end), 'dpnn module required: luarocks install dpnn')

-- return function that returns network definition
return function(params)
    -- get number of classes from external parameters (default to 14)
    local nclasses = params.nclasses or 3

    if pcall(function() require('cudnn') end) then
       print('Using CuDNN backend')
       backend = cudnn
       convLayer = cudnn.SpatialConvolution
       convLayerName = 'cudnn.SpatialConvolution'
    else
       print('Failed to load cudnn backend (is libcudnn.so in your library path?)')
       if pcall(function() require('cunn') end) then
           print('Falling back to legacy cunn backend')
       else
           print('Failed to load cunn backend (is CUDA installed?)')
           print('Falling back to legacy nn backend')
       end
       backend = nn -- works with cunn or nn
       convLayer = nn.SpatialConvolutionMM
       convLayerName = 'nn.SpatialConvolutionMM'
    end

    local feature_len = 1
    if params.inputShape then
        assert(params.inputShape[1]==1, 'Network expects 1xHxW images')
        params.inputShape:apply(function(x) feature_len=feature_len*x end)
    end

    local alphabet_len = 71 -- max index in input samples

    local net = nn.Sequential()
    -- feature_len x 1 x 1
    net:add(nn.View(-1,feature_len))
    -- feature_len
    net:add(nn.OneHot(alphabet_len))
    -- feature_len x alphabet_len
    net:add(backend.TemporalConvolution(alphabet_len, 256, 7))
    -- those shapes are assuming feature_len=4096
    -- [4096-6=4090] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(4090-3)/3+1=1363] x 256
    net:add(backend.TemporalConvolution(256, 256, 7))
    -- [1363-6=1357] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(1357-3)/3+1=452] x 256
    net:add(backend.TemporalConvolution(256, 256, 3))
    -- [452-2=450] x 256
    net:add(nn.Threshold())
    net:add(backend.TemporalConvolution(256, 256, 3))
    -- [450-2=448] x 256
    net:add(nn.Threshold())
    net:add(backend.TemporalConvolution(256, 256, 3))
    -- [448-2=446] x 256
    net:add(nn.Threshold())
    net:add(backend.TemporalConvolution(256, 256, 3))
    -- [446-2=444] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(444-3)/3+1=148] x 256
    net:add(nn.Reshape(37888))
    -- 37888
    net:add(nn.Linear(37888, 4096))
    net:add(nn.Threshold())
    net:add(nn.Dropout(0.5))
    -- 4096
    net:add(nn.Linear(4096, 4096))
    net:add(nn.Threshold())
    net:add(nn.Dropout(0.5))
    -- 4096
    net:add(nn.Linear(4096, nclasses))
    net:add(backend.LogSoftMax())

    -- weight initialization
    local w,dw = net:getParameters()
    w:normal():mul(5e-2)

    return {
        model = net,
        loss = nn.ClassNLLCriterion(),
        trainBatchSize = 128,
        validationBatchSize = 128
    }
end

[gheinrich]

Have you waited long enough? You can check the GPU utilization to see if the GPU is busy. If you train this model on CPU it is going to take forever...

[gonzalolc]

Yes I waited long enough and yes I'm using a Nvidia getforce 950gtx. The problem is that digits just freezes and abort itself without an actual error

[gheinrich]

Closing due to inactivity. Let us know if you find a solution, thanks.