In [1]:

    

# This model is used for wordsim evaluation for results in the paper Multimodal Word Distributions


    

In [2]:

    

%matplotlib inline
from word2gm_loader import Word2GM
from quantitative_eval import *


    

In [3]:

    

model_dir = 'modelfiles/w2gm-k2-d50'


    

In [4]:

    

w2gm_2s = Word2GM(model_dir)


    

    




('Using the latest checkpoint file', u'modelfiles/w2gm-k2-d50/model.ckpt-481392835')
('Number of mixtures = ', 2)

In [5]:

    

w2gm_2s.show_nearest_neighbors('rock', 0)
w2gm_2s.show_nearest_neighbors('rock', 1)


    

    













    




<ggplot: (8787654437793)>
Top 10 highest similarity
['rock:0', 'basalt:1', 'boulder:1', 'boulders:0', 'stalagmites:0', 'stalactites:0', 'rocks:1', 'sand:0', 'quartzite:1', 'bedrock:0']
Top 10 lowest variance of top 20 highest similarity
['breccia:0', 'gneiss:0', 'tuff:0', 'outcropping:1', 'quartzite:1', 'stalagmites:0', 'stalactites:0', 'eroded:0', 'sedimentary:1', 'basalt:1', 'bedrock:0', 'boulders:0', 'boulder:1', 'pebbles:1', 'rocks:1', 'cave:0', 'caves:1', 'limestone:0', 'rock:0', 'sand:0']






    













    




<ggplot: (8787587078329)>
Top 10 highest similarity
['rock:1', 'rock/:1', 'ska:0', 'funk:1', 'pop-rock:1', 'punk:1', 'indie-rock:0', 'band:0', 'indie:0', 'pop:1']
Top 10 lowest variance of top 20 highest similarity
['blues-rock:0', 'psychobilly:0', 'indie-rock:0', 'pop-rock:1', 'pop/:1', 'folk-rock:1', 'psychedelia:0', 'rockabilly:1', 'five-piece:1', 'rock/:1', 'electronica:0', 'ska:0', 'reggae:0', 'techno:1', 'funk:1', 'punk:1', 'indie:0', 'rock:1', 'pop:1', 'band:0']

In [6]:

    

quantitative_eval(model_names=[('w2gm-2s-k50', model_dir)])


    

    
Out[6]:






  

    

      

      
Dataset
      
w2gm-2s-k50/max
      
w2gm-2s-k50/dis
    
  
  

    

      
0
      
SL
      
29.310622
      
26.017090
    
    

      
1
      
WS
      
73.469769
      
62.846685
    
    

      
2
      
WS-S
      
76.729902
      
70.077696
    
    

      
3
      
WS-R
      
71.748623
      
57.983979
    
    

      
4
      
MEN
      
73.553575
      
68.499809
    
    

      
5
      
MC
      
79.083224
      
76.746775
    
    

      
6
      
RG
      
74.511261
      
71.549869
    
    

      
7
      
YP
      
45.074084
      
39.178062
    
    

      
8
      
MT-287
      
66.596475
      
57.241792
    
    

      
9
      
MT-771
      
60.817616
      
57.258203
    
    

      
10
      
RW
      
28.618603
      
31.641952
    
  

In [7]:

    

# the number is slightly different from the paper result for SCWS which uses window size = 5
quantitative_scws_df(model_dir)


    

    
Out[7]:






  

    

      

      
method
      
spearman
    
  
  

    

      
0
      
SCWS_maxdot
      
62.136451