def stage3():
with open("../working/Y.dat", 'rb') as picklefile:
Y = cPickle.load(picklefile)
# Load hierarchy (parents & children indices)
parents_index = preproc.extract_parents(Y, "../raw_data/hierarchy.txt")
children_index = preproc.inverse_index(parents_index)
##Save state
with open("../working/parents_index.dat", 'wb') as picklefile:
cPickle.dump(parents_index, picklefile, -1)
with open("../working/children_index.dat", 'wb') as picklefile:
cPickle.dump(children_index, picklefile, -1)
def stage3():
with open("../working/Y.dat", 'rb') as picklefile:
Y = cPickle.load(picklefile)
# Load hierarchy (parents & children indices)
parents_index = preproc.extract_parents(Y, "../raw_data/hierarchy.txt")
children_index = preproc.inverse_index(parents_index)
##Save state
with open("../working/parents_index.dat", 'wb') as picklefile:
cPickle.dump(parents_index, picklefile, -1)
with open("../working/children_index.dat", 'wb') as picklefile:
cPickle.dump(children_index, picklefile, -1)
def onego_main():
# Test on toyset.
preproc.subset("../raw_data/train.csv", "../data/train.csv", 1, 200000)
# Load toyset .csv -> X & Y
X, Y = preproc.extract_XY("../data/train.csv")
# Prune corpora
label_counter = pruning.LabelCounter(Y)
word_counter = pruning.WordCounter(X)
label_counter.prune(no_below=2, no_above=1.0, max_n=None)
word_counter.prune(no_below=2, no_above=0.4, max_n=None) # assume balanced
pruning.prune_corpora(X, Y, label_counter, word_counter)
del word_counter # free up memory
# Transform X to tf-idf
bin_word_counter = pruning.WordCounter(X, binary=True)
similarity.transform_tfidf(X, bin_word_counter)
del bin_word_counter # free up memory
# Load hierarchy (parents & children indices)
parents_index = preproc.extract_parents(Y, "../raw_data/hierarchy.txt")
children_index = preproc.inverse_index(parents_index)
# CV-split X & Y (using default params)
v_X, v_Y, t_X, t_Y = cv.prop_sample_CV(X=X, Y=Y)
del X, Y # free up memory
# Obtain k-NN scores & pscores, predict, and calculate F1!
k = 70
w1, w2, w3, w4 = 3.4, 0.6, 0.8, 0.2
alpha = 0.9
cat_pns = evaluation.CategoryPNCounter()
for d_i, labels_i in izip(v_X, v_Y):
scores, pscores = similarity.cossim(d_i, t_X, k, t_Y, parents_index,
children_index)
ranks = similarity.optimized_ranks(scores, pscores, label_counter, w1,
w2, w3, w4)
predicted_labels = similarity.predict(ranks, alpha)
cat_pns.fill_pns(predicted_labels, labels_i)
cat_pns.calculate_cat_pr()
MaF = cat_pns.calculate_MaF()
print "MaF:", MaF
def onego_main():
# Test on toyset.
preproc.subset("../raw_data/train.csv", "../data/train.csv", 1, 200000)
# Load toyset .csv -> X & Y
X, Y = preproc.extract_XY("../data/train.csv")
# Prune corpora
label_counter = pruning.LabelCounter(Y)
word_counter = pruning.WordCounter(X)
label_counter.prune(no_below=2, no_above=1.0, max_n=None)
word_counter.prune(no_below=2, no_above=0.4, max_n=None) # assume balanced
pruning.prune_corpora(X, Y, label_counter, word_counter)
del word_counter # free up memory
# Transform X to tf-idf
bin_word_counter = pruning.WordCounter(X, binary=True)
similarity.transform_tfidf(X, bin_word_counter)
del bin_word_counter # free up memory
# Load hierarchy (parents & children indices)
parents_index = preproc.extract_parents(Y, "../raw_data/hierarchy.txt")
children_index = preproc.inverse_index(parents_index)
# CV-split X & Y (using default params)
v_X, v_Y, t_X, t_Y = cv.prop_sample_CV(X=X, Y=Y)
del X, Y # free up memory
# Obtain k-NN scores & pscores, predict, and calculate F1!
k = 70
w1, w2, w3, w4 = 3.4, 0.6, 0.8, 0.2
alpha = 0.9
cat_pns = evaluation.CategoryPNCounter()
for d_i, labels_i in izip(v_X, v_Y):
scores, pscores = similarity.cossim(d_i, t_X, k, t_Y, parents_index,
children_index)
ranks = similarity.optimized_ranks(scores, pscores, label_counter,
w1, w2, w3, w4)
predicted_labels = similarity.predict(ranks, alpha)
cat_pns.fill_pns(predicted_labels, labels_i)
cat_pns.calculate_cat_pr()
MaF = cat_pns.calculate_MaF()
print "MaF:", MaF
