Ludwig integration on custom platform

Hello, I am trying to integrate some ludwig features inside our platform and I would ask you basically two things:

QUESTION 1

The first consists in the numpy version: why in requirements.txt numpy>=1.15,<1.16 is specified? what is the main reason? Because in our platform we use latest version of numpy (1.16.1 with pip) and if I’m not wrong, this version creates problem with ludwig:

ImportError: Missing required dependencies ['numpy']

Installing numpy>=1.15,<1.16 is ok but it could make conflicts with other features that are in our platform.

---

QUESTION 2 This is a technical question: I am trying to implement a regression predictor that has text in input and its output is a two value tuple between 0 and 1 (x , y where 0<x<1 and 0<y<1, formally ). This is my model definition:

 'input_features': [ {'name': 'text', 
                            'type': 'text' , 
                            'encoder': 'rnn', 
                            'cell': 'lstm', 
                            'bidirectional': 'true'} ], 
'output_features': [
                                {'name': 'x', 
                                'type': 'numerical'    }, 
                                {'name': 'y', 
                                'type': 'numerical'}
                            ]

But in this way I get sometimes outputs (in the inference phase) < 0 or > 1. How can I tell output_features to produce numerical numbers between 0 and 1 ( sigmoid layer ) ?

Thanks for your support!

[w4nderlust]

As it was a pretty quick thing to do, I ended up adding a clip parameter to numerical output features too. now you can have:

output_features:
-
  name: x
  type: numerical
  clip: [0, 1]

Enjoy 😃 Closing it but feel free to comment further if needed.

[simonefrancia]

Ok I think I solved. What I did is to use the configuration of the packages of out platform, and with this configuration of package I trained the model, in order to be compatible with it in inference mode. Hope that it can be helpful for something in the future!