DIGITS Tutorial based on this project about crepe

[gheinrich]

This isn't an issue so I am closing now.

[aichemzee]

Do you have a public aws ami with digits installed?

[gheinrich]

Sorry I don't have that. We have a docker image for DIGITS which makes installation very straightforward.

However you need a more recent version of DIGITS than available in the Docker image to run the Text Classification tutorial. @flx42 is it conceivable to publish a Dockerfile to allow users to create images off the latest DIGITS code from Github?

[darwinzer0]

Thank you for implementing this, it works much faster. Because the input is 1024 characters instead of the 1014 in the original version, should the sizes of the layers be slightly different? e.g. 341 x 256 after the first TemporalMaxPooling, and so forth.

[gheinrich]

Oh yes I should have updated this comment. Or perhaps it should be removed since the idea is to have the number of input characters a parameter.

With feature_len=1024 don't we end up with (1024-6-3)/3+1=339 features after the first max pooling operation? So the successive shapes would be:

    -- those shapes are assuming feature_len==1024
    -- 1024 x alphabet_len
    net:add(backend.TemporalConvolution(alphabet_len, 256, 7))
    -- [1024-6=1018] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(1018-3)/3+1=339] x 256
    net:add(backend.TemporalConvolution(256, 256, 7))
    -- [339-6=333] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(333-3)/3+1=111] x 256
    net:add(backend.TemporalConvolution(256, 256, 3))
    net:add(nn.Threshold())
    -- [111-2=109] x 256
    net:add(backend.TemporalConvolution(256, 256, 3))
    net:add(nn.Threshold())
    -- [109-2=107] x 256
    net:add(backend.TemporalConvolution(256, 256, 3))
    net:add(nn.Threshold())
    -- [107-2=105] x 256
    net:add(backend.TemporalConvolution(256, 256, 3))
    -- [105-2=103] x 256
    net:add(nn.Threshold())
    net:add(nn.TemporalMaxPooling(3, 3))
    -- [(103-3)/3+1=34] x 256
    net:add(nn.Reshape(8704))

We still end up with 8704 features at the input of the fully-connected layers.

[darwinzer0]

Ah yes, I had forgotten the -6 in the first convolution. Thanks, this helps a lot. I was making some changes to the network and wanted make sure I was calculating everything correctly.

[zhangxiangxiao]

This is wonderful! I see the pull request is in DIGITS already. If you do not mind, I will probably brag about it on Facebook a bit :P

Thanks for the great contribution!

[gheinrich]

If you do not mind, I will probably brag about it on Facebook a bit :P

You are most welcome to do so :-)