Multi task learning use Neon

[tianxiangchen2015]

Hi guys,

How can I use neon to do multi task learning? First, I tired to use layer container to handle multiple inputs and multiple outputs. But the problem is "Arrayiterator" can not handle multiple inputs with multiple number of classes. For example, I have 3 datasets, first and second dataset has 2 classes, the last one has 4 classes. How can I do it?

I want to build a model with 3 inputs and one shared hidden layer, then splits to 3 outputs.

Thanks

[hanlint]

Hello,

You will first want to use the MergeBroadcast container (see examples/image_caption.py for an example where you have two inputs) to build your model. For the data side, you can either write a custom iterator that yields a tuple of inputs. In your case something like:

yield (x1, x2, x3), (y1, y2, y3)

Alternatively, if you pack your data into hdf5 datasets, you may also be able to use the HDF5 iterator, but we haven't really tested this.