Segmentation Fault

[shawnjhenry]

Hi,

I'm attempting to train a network consisting of a recurrent layer on top of a time delay conv net using nn.Recurrent. It runs fine, albeit slowly, on my Macbook, but when I try to run it on my much faster Ubuntu desktop I get a segmentation fault error with a core dump message. Where does torch dump the error message?

In case you want to try reproducing the error, the network looks like this:

memorySize = 10 windowSize = 5 nSamples = 2693 nFeatures = 100 nClasses = 7

TDCNN = nn.Sequential() TDCNN:add(nn.LookupTable(nSamples, nFeatures)) TDCNN:add(nn.Padding(1, windowSize-1)) TDCNN:add(nn.Padding(1, -windowSize+1)) TDCNN:add(nn.TemporalConvolution(nFeatures, nFeatures, windowSize)) TDCNN:add(nn.Max(1))

RNN = nn.Recurrent(nn.Identity(), TDCNN, nn.Linear(nFeatures, nFeatures), nn.Tanh(), memorySize)

model = nn.Sequential() model:add(RNN) model:add(nn.Linear(nFeatures, nClasses)) model:add(nn.LogSoftMax())

loss = nn.ClassNLLCriterion()

Inputs to the network are torch.LongTensors of varying length.

Thanks, Shawn

[nicholas-leonard]

Hi, the nn.Recurrent in this repository is deprecated. Use the rnn repository instead. For examples, you can checkout dp. If you still have issues, let me know.