###########################################################
#Computational Linguistics 17-18
#DISSECT tutorial
#How to: PLAY WITH cDSMs
#sandro.pezzelle@unitn.it
###########################################################

# 1. APPLY WEIGHTED ADDITIVE MODEL

###########################################################

#let's see how weigthing differently the consituents
#affects the composed vector

my_comp = WeightedAdditive(alpha = 1, beta = 1)

composed_space2 = my_comp.compose([("sword", "fish", "sword_fish"),
                                  ("bull", "dog", "bull_dog")], 
                                 my_space)
print composed_space2.get_neighbours("bull_dog", 10, CosSimilarity(), 
                              space2 = my_space)

#'bulldog' is not among the top 10 neighbors. Let's weight 'bull' less...
my_comp = WeightedAdditive(alpha = 0.2, beta = 0.8)

composed_space2 = my_comp.compose([("sword", "fish", "sword_fish"),
                                  ("bull", "dog", "bull_dog")], 
                                 my_space)
print composed_space2.get_neighbours("bull_dog", 10, CosSimilarity(), 
                              space2 = my_space)

#now 'bulldog' appears in the 10th position. Slightly better! :)
#you'd be asking yourself whether there's any way to 'learn'
#the weights to be applied in order to get a composed vector
#that approximates the original one... the answer is YES!

###########################################################

# 2. PLAY WITH FULL ADDITIVE MODEL

###########################################################

print composed_space_fadd_test.get_neighbours("volley_ball", 10, CosSimilarity(), 
                              space2 = arg_space)
print composed_space_fadd_test.get_neighbours("wind_mill", 10, CosSimilarity(), 
                              space2 = arg_space)

#it seems working, right?
#note that the model gets better and better as the number of
#training examples increase!

###########################################################

# 3. TRAIN AND TEST YOUR FULL-ADD MODEL

###########################################################

from composes.utils import io_utils, log_utils
from composes.semantic_space.space import Space
from composes.composition.full_additive import FullAdditive
from composes.similarity.cos import CosSimilarity

my_space = io_utils.load("vectors.pkl")
compound_space = io_utils.load("composed_space_compounds.pkl")
print rows_observed = compound_space.get_row2id()

my_comp = FullAdditive()

train_data = [("tennis", "table", "tennis_table"),
		("dog", "house", "dog_house"),
		("training", "shoes", "training_shoes"),
		("pencil", "case", "pencil_case"),
		("volley", "ball", "volley_ball"),
		("wind", "mill", "wind_mill")]

my_comp.train(train_data, my_space, compound_space)

test_data_ym3 = [("bike", "girl", "bike_girl"),
              ("city", "bike", "city_bike"),
	      ("thursday", "class", "thursday_class"),
              ("tv", "series", "tv_series")]

composed_ym3 = my_comp.compose(test_data_ym3, my_space)

print composed_ym3.get_neighbours("bike_girl", 10, CosSimilarity(), 
                              space2 = my_space)
print composed_ym3.get_neighbours("city_bike", 10, CosSimilarity(), 
                              space2 = my_space)
print composed_ym3.get_neighbours("thursday_class", 10, CosSimilarity(), 
                              space2 = my_space)
print composed_ym3.get_neighbours("tv_series", 10, CosSimilarity(), 
                              space2 = my_space)