Using nn.PReLU with Sequencer leads to SegFault on CPU or cuda runtime error (77) on GPU

[spurra]

I've been playing around with the RNN library and got the following code:

cnn = nn.Sequential()

-- Creates a conv_block: [SpatConv, SpatMaxPool, PReLU, Dropout]
function add_conv_block(model, nFilterIn, nFilterOut, filterH, filterW, dOut, 
    preluLeak)  
    dOut = dOut or default_dropout
    preluLeak = preluLeak or default_preluLeak
    return model:add(nn.SpatialConvolution(nFilterIn, nFilterOut, filterW, 
        filterH))
        :add(nn.SpatialBatchNormalization(nFilterOut))
        :add(nn.SpatialDropout(dOut))
        :add(nn.PReLU(preluLeak))
end


cnn:add(nn.Reshape(1,8,24))
add_conv_block(cnn, 1, 32, 2,6)
cnn:add(nn.Reshape(4256))
cnn:add(nn.Linear(4256, 2048))
-------------------------------------------------------

-- Construct the complete model

model = nn.Sequencer(
	nn.Sequential()
		:add(cnn)
	)

model:cuda()

-------------------------------------------------------

-- Test with dummy input
x = torch.CudaTensor(10, 128, 192)

y = model:forward(x)

Running this on CPU gives me a segmentation fault. Running it on the GPU gives me the following error: cuda runtime error (77) : an illegal memory access was encountered.

After iteratively leaving out parts of the architecture, I found the culprit: nn.PReLU. Uncommenting it leads to normal behaviour. I've replaced it with LeakyReLU, which works fine. I'd like to report this issue in case it can be fixed.