def intersection_align(embed1, embed2):
"""
Get the intersection of two embeddings.
Returns embeddings with common vocabulary and indices.
"""
common_vocab = filter(set(embed1.iw).__contains__, embed2.iw)
newvecs1 = np.empty((len(common_vocab), embed1.m.shape[1]))
newvecs2 = np.empty((len(common_vocab), embed2.m.shape[1]))
for i in xrange(len(common_vocab)):
newvecs1[i] = embed1.m[embed1.wi[common_vocab[i]]]
newvecs2[i] = embed2.m[embed2.wi[common_vocab[i]]]
return Embedding(newvecs1, common_vocab), Embedding(newvecs2, common_vocab)
def get_dim_reduced_fixed(self, dim=2, normalize=False, basis=2000):
pca = PCA(n_components=2)
pca.fit(self.embeds[basis].m)
year_reduced_embeds = collections.OrderedDict()
for year,embed in self.embeds.iteritems():
proj_embed_mat = pca.transform(embed.m)
year_reduced_embeds[year] = Embedding(proj_embed_mat, embed.iw, normalize=normalize)
return SequentialEmbedding.from_ordered_dict(year_reduced_embeds)
def get_word_path(self, word, n=3, dim=2, num_rand=None, word_list=None, basis_year=2000):
subembeds = self.get_word_subembeds(word, n=n, num_rand=num_rand, word_list=word_list)
pca = PCA(n_components=dim)
basis = pca.fit_transform(subembeds.embeds[basis_year].m)
basis = Embedding(basis, subembeds.embeds[basis_year].iw, normalize=False)
word_path = collections.OrderedDict()
for year, embed in self.embeds.iteritems():
word_path[year] = pca.transform(embed.represent(word)).flatten()
return word_path, basis
def linear_align(base_embed, other_embed):
"""
Align other embedding to base embedding using best linear transform.
NOTE: Assumes indices are aligned
"""
basevecs = base_embed.m
othervecs = other_embed.m
fixedvecs = othervecs.dot(np.linalg.pinv(othervecs)).dot(basevecs)
return Embedding(fixedvecs, other_embed.iw)
def smart_procrustes_align(base_embed, other_embed, post_normalize=False):
in_base_embed, in_other_embed = intersection_align(base_embed, other_embed)
base_vecs = in_base_embed.m
other_vecs = in_other_embed.m
m = other_vecs.T.dot(base_vecs)
u, _, v = np.linalg.svd(m)
ortho = u.dot(v)
return Embedding((other_embed.m).dot(ortho),
other_embed.iw,
normalize=post_normalize)
def procrustes_align(base_embed, other_embed):
"""
Align other embedding to base embeddings via Procrustes.
Returns best distance-preserving aligned version of other_embed
NOTE: Assumes indices are aligned
"""
basevecs = base_embed.m - base_embed.m.mean(0)
othervecs = other_embed.m - other_embed.m.mean(0)
m = othervecs.T.dot(basevecs)
u, _, v = np.linalg.svd(m)
ortho = u.dot(v)
fixedvecs = othervecs.dot(ortho)
return Embedding(fixedvecs, other_embed.iw)
def get_dim_reduced_avg(self, dim=2, normalize=False):
first_iter = True
for year,embed in self.embeds.iteritems():
if first_iter:
align_words = set(embed.iw)
first_iter = False
else:
align_words = align_words.intersection(set(embed.iw))
first_iter = True
for year,embed in self.embeds.iteritems():
if first_iter:
avg_mat = embed.get_subembed(list(align_words)).m
first_iter = False
else:
avg_mat += embed.get_subembed(list(align_words)).m
avg_mat /= float(len(self.embeds))
pca = PCA(n_components=2)
pca.fit(avg_mat)
year_reduced_embeds = collections.OrderedDict()
for year,embed in self.embeds.iteritems():
proj_embed_mat = pca.transform(embed.m)
year_reduced_embeds[year] = Embedding(proj_embed_mat, embed.iw, normalize=normalize)
return SequentialEmbedding.from_ordered_dict(year_reduced_embeds)
def write_vecs(finname, foutname):
og_embed = Embedding.load(finname, normalize=False)
red_embed = og_embed.get_subembed(TOP_WORDS)
np.save(foutname+".npy", red_embed.m)
with file(foutname+".vocab","w") as outf:
print >> outf, " ".join(red_embed.iw)
def __init__(self, years, top_freq=None):
self.embeds = collections.OrderedDict()
for year in years:
self.embeds[year] = Embedding.load(INPUT_PATH.format(year=year))
def reduce_dim(embedding, dim=2, post_normalize=False):
pca = PCA(n_components=2)
reduced_vecs = pca.fit_transform(embedding.m)
return Embedding(reduced_vecs, embedding.iw, normalize=post_normalize)
