The csne from aida-ugent

Compatibility with python 3.7+

Python 2.7 is way past end-of-life and Python 3.6 will be at end of life at the end of the year. Are there specific reasons why you pinned the repository to older versions of python as well as pretty old versions of its dependencies?

It says on the README now that the code is extensively tested but I'm not really sure what this means without having something like unittests that can be run programatically. If testing were possible programatically, you could show that it's compatible with more modern versions of python and the dependent libraries

Search for alpha results in overflow error

I'm getting an overflow error when fitting the prior because of v = np.exp(v). Exact output:

[iter 0] Gradient norm: 5491631.49                                                                                                     
[iter 0] Objective: 1321041.74                                                                                                         
[iter 1] Gradient norm: 9906547.80                                                                                                     
[iter 1] Objective: 1285034.30                                                                                                         
Fitting prior:  20%|████████████████▊                                                                   | 2/10 [00:00<00:03,  2.44it/s]
/Users/cthoyt/dev/CSNE/src/csne/maxent_comb.py:191: RuntimeWarning: overflow encountered in exp
  v = np.exp(v)
[iter 2] Gradient norm: 10061996.41                                                                                                    
[iter 2] Objective: 1284818.77                                                                                                         
[iter 3] Gradient norm: 10061021.36                                                                                                    
[iter 3] Objective: 1284812.61                                                                                                         
[iter 4] Gradient norm: 10060726.78                                                                                                    
[iter 4] Objective: 1284806.45                                                                                                         
[iter 5] Gradient norm: 10060432.22                                                                                                    
[iter 5] Objective: 1284800.29                                                                                                         
[iter 6] Gradient norm: 10060137.66                                                                                                    
[iter 6] Objective: 1284794.13                                                                                                         
[iter 7] Gradient norm: 10059843.11                                                                                                    
[iter 7] Objective: 1284787.98                                                                                                         
[iter 8] Gradient norm: 10059548.56                                                                                                    
[iter 8] Objective: 1284781.82                                                                                                         
[iter 9] Gradient norm: 10059254.03                                                                                                    
[iter 9] Objective: 1284775.66

It's happening in this code:

CSNE/src/csne/maxent_comb.py

Lines 188 to 200 in 27f9d10

 while True: 

 x_test = x - alpha * delta 

 v = x_test[r_idx] + x_test[c_idx] 

 for fi in range(self.__nfuncs): 

 v += self.__F[fi].maskdF * x_test[self.__2n + fi] 

 v = np.exp(v) 

 obj_test = np.sum(np.log(1 + v)) - np.dot(x_test, self.__cs) 

 if obj_test <= obj - (alpha * imp): 

 x = x_test 

 aux = v 

 obj = obj_test 

 break 

 alpha = alpha / 2.0

Any ideas? If you look at the norms, they jump up to being much higher after this error. Maybe this math should be done in log space? Or alternatively, perhaps these numbers are nonsensically high and everything should be normed.

Make pip installable

Thanks for the nice package and paper - it would be much more useful if it were pip installable. Would you be willing to accept a PR to enable this?

aida-ugent / csne Goto Github PK

csne's People

Contributors

Stargazers

Watchers

Forkers

csne's Issues

Compatibility with python 3.7+

Search for alpha results in overflow error

Make pip installable

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

Jobs

	while True:
	x_test = x - alpha * delta
	v = x_test[r_idx] + x_test[c_idx]
	for fi in range(self.__nfuncs):
	v += self.__F[fi].maskdF * x_test[self.__2n + fi]
	v = np.exp(v)
	obj_test = np.sum(np.log(1 + v)) - np.dot(x_test, self.__cs)
	if obj_test <= obj - (alpha * imp):
	x = x_test
	aux = v
	obj = obj_test
	break
	alpha = alpha / 2.0