Questions regarding differences between the implementation and the experiment details in the research paper about dilated-cnn-ner

[strubell]

Yes, the configuration files should implement the same models reported in the paper.

I think you've answered your own question: the 4 layers for the sentence-level dilated CNN model are the 3 layers defined in the layers json in dilated-cnn.conf, plus the initial projection.

[nikhilrayaprolu]

Okay so does this mean, there is nowhere in the experiments where a dilation rate of 4 is used? Then does that affect the total receptive field of the CNN? since using a dilation of 2 doesn't give us a good level of receptive field(effective input width as per the paper).

[strubell]

We experimented with models w/ dilation rate 4, but we found iterating smaller dilation widths to work better. This could partly have been an artifact of TensorFlow's old, naive (bad) implementation of dilated convolutions, which was very memory and time intensive, relative to what it should be. TF versions 1.5+ should have fixed this, so you may find that you can get better results with wider dilations / deeper networks, which should train faster and require less GPU memory now.

Our models still have a wide receptive field due to stacking w/ the dilation-2 layers, certainly much wider than the non-dilated equivalent number of layers.

[nikhilrayaprolu]

I am planning to make a pytorch equivalent of the research paper, and try few extensions for it. Thank You, for your quick replies. your quick replies would surely help me do it better and quicker.

[strubell]

Great, let me know if you have further questions!