def do_prepare(self, params, prepare):
if similarity in params:
self.similarity = similarity
else: # Default similarity is cosine
self.similarity = lambda s1, s2: cosine(np.nan_to_num(s1),
np.nan_to_num(s2))
return prepare(params, self.samples)
def run(self, batcher, params):
results = {}
for dataset in self.datasets:
sys_scores = []
input1, input2, gs_scores = self.data[dataset]
for ii in range(0, len(gs_scores), params.batch_size):
batch1 = input1[ii:ii + params.batch_size]
batch2 = input2[ii:ii + params.batch_size]
# we assume that the get_batch function already throws out the faulty ones
if len(batch1) == len(batch2) and len(batch1) > 0:
enc1 = batcher(batch1, params)
enc2 = batcher(batch2, params)
for kk in range(enc2.shape[0]):
sys_score = cosine(np.nan_to_num(enc1[kk]), np.nan_to_num(enc2[kk]))
sys_scores.append(sys_score)
results[dataset] = {'pearson': pearsonr(sys_scores, gs_scores), 'spearman': spearmanr(sys_scores, gs_scores)}
logging.debug('{0} : pearson = {1}, spearman = {2}'.format(dataset, results[dataset]['pearson'], results[dataset]['spearman']))
avg_pearson = np.mean([results[dset]['pearson'][0] for dset in results.keys()])
avg_spearman = np.mean([results[dset]['spearman'][0] for dset in results.keys()])
results['all'] = {'pearson': avg_pearson, 'spearman': avg_spearman}
logging.debug('Results (all) : Pearson = {0}, Spearman = {1}\n'.format(results['all']['pearson'], results['all']['spearman']))
return results
def scoring_fn(i, j):
"""
Calculate how similar the entities are
:param i: triple index
:param j: triple index
:return: cosine similarity
"""
# Get verb entities
ents = (marg[i], marg[j])
# Get cosine similarity
return cosine(*ents)
def getdistinctivefeatures(lang1, lang2, phonemeMap):
"""
Contrast this with getF1.
I can't get this to work correctly.
:param lang1: a set of Phonemes
:param lang2: a set of Phonemes
:return: the Distinctive Features score for these languages.
"""
if len(lang1) == 0:
print "ERROR: first lang is empty or doesn't exist"
return -1
if len(lang2) == 0:
print "ERROR: second lang is empty or doesn't exist"
return -1
# loop over all pairs.
scores = {}
total = 0
for p in lang1:
# get closest in lang2
maxsim = 0 # just a small number...
maxp = None # max phoneme associate with maxsim
for p2 in lang2:
pu1 = p.Phoneme
pu2 = p2.Phoneme
if pu1 in phonemeMap and pu2 in phonemeMap:
ps = tuple(sorted([pu1, pu2]))
if ps in phonedist:
sim = phonedist[ps]
else:
sim = 1-utils.cosine(phonemeMap[pu1], phonemeMap[pu2])
phonedist[ps] = sim
else:
# not there...?
#print "SHOULD NEVER HAPPEN!",
#if pu1 not in phonemeMap:
# print "missing ", pu1
#if pu2 not in phonemeMap:
#print "missing ", pu2
sim = 0
scores[(pu1,pu2)] = sim
total += sim
total /= float(len(lang1) * len(lang2))
return total
def score_fn(term, description, verb_weight=1, head_weight=1):
"Score using vector addition and cosine similarity"
which, triple = description
if which == 'SBJ':
weights = [verb_weight, head_weight, 1]
elif which == 'OBJ':
weights = [verb_weight, 1, head_weight]
else:
raise ValueError(which)
combined = sum(vec[token] * wei for token, wei in zip(triple, weights))
target = vec[term]
return cosine(target, combined)
def compare(lang1, lang2):
"""
Given two language names, get distance according to phonology scores.
:param lang1: name of first lang (eg English)
:param lang2: name of second lang
:return: the distance of the languages, or -1 if one or both langs not found.
"""
l1, l2 = comparefeats(lang1, lang2)
if l1 and l2:
return utils.cosine(l1.phon_feats(), l2.phon_feats())
else:
print "One or both langs not found: {0}, {1}".format(l1, l2)
return -1
def scoring_fn(i, j):
"Add triples and take cosine"
# Get tokens
v1 = raw_triples[i][0]
v2, s, o = raw_triples[j]
if with_lookup:
v1 = pred_name[lookup['v'][v1]]
v2 = pred_name[lookup['v'][v2]]
s = pred_name[lookup['n'][s]]
o = pred_name[lookup['n'][o]]
# Lookup vectors
v1, v2, s, o = [vec[model.vocab[x].index] for x in [v1, v2, s, o]]
comp1 = v1 + s + o
comp2 = v2 + s + o
return cosine(comp1, comp2)
def calculate_distance_offset(pairs, df, offset, offset2=None):
# CONVERTED TO calculate_distance_mode
"""
From a set of pairs, calculate the distance of the offset of the
pair in relation to the global offset.
Parameters:
-----------
pairs: list
list containing tuples of IDs of norms [(id1,id2),(id3,id4)...]
df: pandas.dataframe
dataframe containing ids and embeddings of sentences
offset: np.array
vector containing the global offset (offset of all conflicts)
"""
label = 0
vdist = []
pb = progressbar.ProgressBar(len(pairs))
for i, arr in enumerate(pairs):
emb1 = df.id2embed(arr[0])
emb2 = df.id2embed(arr[1])
local_offset = emb1 - emb2
# cosine (similar:0->2:not_similar)
cos = utils.cosine(local_offset, offset)
# euclidean distance (similar:0->inf:not_similar)
euc = utils.euclidean(local_offset, offset)
if len(offset2) > 0:
cos2 = utils.cosine(local_offset, offset2)
euc2 = utils.euclidean(local_offset, offset2)
vdist.append([cos, euc, cos2, euc2])
else:
vdist.append((cos, euc))
pb.update()
#if i == 1000: break
return vdist
def compare(lang1, lang2):
"""
Given two language names, get distance according to phonology scores.
:param lang1: name of first lang (eg English)
:param lang2: name of second lang
:return: the distance of the languages, or -1 if one or both langs not found.
"""
l1, l2 = comparefeats(lang1, lang2)
if l1 and l2:
return utils.cosine(l1.phon_feats(),l2.phon_feats())
else:
print "One or both langs not found: {0}, {1}".format(l1, l2)
return -1
def run(self, batcher, params):
results = {}
for dataset in self.datasets:
sys_scores = []
input1, input2, gs_scores = self.data[dataset]
for ii in range(0, len(gs_scores), params.batch_size):
batch1 = input1[ii:ii + params.batch_size]
batch2 = input2[ii:ii + params.batch_size]
# we assume that the get_batch function already throws out the faulty ones
if len(batch1) == len(batch2) and len(batch1) > 0:
enc1 = batcher(batch1, params)
enc2 = batcher(batch2, params)
for kk in range(enc2.shape[0]):
sys_score = cosine(np.nan_to_num(enc1[kk]),
np.nan_to_num(enc2[kk]))
sys_scores.append(sys_score)
results[dataset] = {'pearson': pearsonr(sys_scores, gs_scores), 'spearman': spearmanr(sys_scores, gs_scores),\
'nsamples': len(sys_scores)}
logging.debug('%s : pearson = %.4f, spearman = %.4f' %
(dataset, results[dataset]['pearson'][0],
results[dataset]['spearman'][0]))
weights = [results[dset]['nsamples'] for dset in results.keys()]
list_prs = np.array(
[results[dset]['pearson'][0] for dset in results.keys()])
list_spr = np.array(
[results[dset]['spearman'][0] for dset in results.keys()])
avg_pearson = np.average(list_prs)
avg_spearman = np.average(list_spr)
wavg_pearson = np.average(list_prs, weights=weights)
wavg_spearman = np.average(list_spr, weights=weights)
results['all'] = {'pearson': {'mean':avg_pearson, 'wmean':wavg_pearson},\
'spearman': {'mean':avg_spearman, 'wmean':wavg_spearman}}
logging.debug(
'ALL (weighted average) : Pearson = %.4f, Spearman = %.4f' %
(wavg_pearson, wavg_spearman))
logging.debug('ALL (average) : Pearson = %.4f, Spearman = %.4f\n' %
(avg_pearson, avg_spearman))
return results
def neighbours(self, word, size = 10):
"""
Get nearest words with KDTree, ranking by cosine distance
"""
word = word.strip()
v = self.word_vec(word)
[distances], [points] = self.kdt.query(array([v]), k = size, return_distance = True)
assert len(distances) == len(points), "distances and points should be in same shape."
words, scores = [], {}
for (x,y) in zip(points, distances):
w = self.index2word[x]
if w == word: s = 1.0
else: s = utils.cosine(v, self.syn0[x])
if s < 0: s = abs(s)
words.append(w)
scores[w] = min(s, 1.0)
for x in sorted(words, key=scores.get, reverse=True):
yield x, scores[x]
def calc_word_similarity(self, test_file, embed_vec):
"""Calculate Word Similarity
Arguments:
test_file {str} -- similarity test file e.g wordsim-240 and wordsim-297
embed_vec {Keyedvectors} -- A pre-load gensim word vectors
"""
pred, label, found = [], [], 0
with open(test_file, 'r', encoding='utf8') as fr:
lines = fr.readlines()
for line in lines:
w1, w2, score = line.split()
if w1 in embed_vec and w2 in embed_vec:
found += 1
pred.append(cosine(embed_vec[w1], embed_vec[w2]))
label.append(float(score))
file_name = test_file.split("/")[-1].replace('.txt', '')
print(f"Test File: {file_name}")
print(f"Numbers of words Found: {found}")
print(f"Numbers of words Not Found: {len(lines) - found}")
print(f"Spearman's Rank Coeficient: {rho(label, pred)}")
# construction of the query vector
parsed_dir = "C:\\Users\\nbonardo\\movies_parsed"
n_files = len([name for name in os.listdir(parsed_dir) if os.path.isfile(os.path.join(parsed_dir, name))])
queryVector = dict()
nWordsQ = len(term_ids)
for term_id in term_ids:
term_idf = 1 + math.log(n_files / len(idx[str(term_id)]))
queryVector[term_id] = term_idf / nWordsQ
#print("Query vector", queryVector)
# calculation of the cosine similarity between each document vector of the result and the quere vector
# storing the result (as tuple) in a heap structure
h = []
for doc in final:
cos = utils.cosine(queryVector, docVectors[doc])
#print("Cosine similarity of doc", doc, "=", cos, utils.getMovieTitle(doc))
heapq.heappush(h, (cos, doc))
# output of the top-K movies according to cosine similarity
K = 5
K = min(K, len(final))
print("*** TOP-", K, " results ***")
print(" Id | Title | Cosine Similarity")
print("-----+------------------------------------------+------------------")
for _ in range(K):
movieTup = heapq.heappop(h)
#print(movieTup[1], utils.getMovieTitle(movieTup[1]), movieTup[0])
print("%4s | %40s | %.6s" % (movieTup[1]+1, utils.getMovieTitle(movieTup[1]), movieTup[0]))
elif search_engine == '3':
def relevance(self, title, content, sentences):
"""计算各个句子和标题以及正文的相关性"""
c = []
for s in sentences:
c.append(self.weight * cosine(s, content) + (1 - self.weight) * cosine(s, title))
return c
# F6: Proper Noun
for Si in sents:
Si_propnouns = np.intersect1d(Si, propernoun)
F6.append(len(Si_propnouns) / len(Si))
# F7: Similarities Between Sentences
vocab = sorted(set(flat))
TF = get_TF(sents, vocab)
sim_SiSj = []
for i, Si in enumerate(TF):
temp = []
for j, Sj in enumerate(TF):
if i == j: continue
temp.append(cosine(Si, Sj))
sim_SiSj.append(sum(temp))
max_simSiSj = max(sim_SiSj)
for sim_Si in sim_SiSj:
F7.append(sim_Si / max_simSiSj)
# F8: Term Weight
TFIDF = get_TFIDF(sents, vocab)
sum_TFIDF = []
for tfidf in TFIDF:
sum_TFIDF.append(sum(tfidf))
max_sum_TFIDF = max(sum_TFIDF)
for sum_tfidf in sum_TFIDF:
ids = torch.tensor(query_tokens['input_ids']).unsqueeze(0)
mask = torch.tensor(query_tokens['attention_mask']).unsqueeze(0)
pred = classifier(ids, mask)
top_val, top_idx = torch.topk(pred[0], 3, dim=1)
pred_categories = model_classes[top_idx].tolist()[0]
topics = [cat_map[cat] for cat in pred_categories]
### Encode query to embedding and return top item
# encode user query
query_embedding = sentence_transformer.encode(query)
# filter df to relevant categories and grab embeddings
relevant_docs = df[df['categories'].isin(pred_categories)]
relevant_embeddings = embeddings[relevant_docs.index]
# Calculate cosine similarity of user query and relevant embeddings
sims = []
for doc in relevant_embeddings:
doc_sim = cosine(query_embedding, doc)
sims.append(doc_sim)
top_matches = np.argsort(sims)[::-1]
top_item = relevant_docs.iloc[top_matches[0]]
print(f'User Query: {query}\n')
print(f'Predicted Topics {topics}\n')
print(f'Recomended Paper:\n {top_item.title}\n')
print(f'Abstract:\n{top_item.abstract}\n')
def compute_loss(self, input_image, pca_render, gcn_render, pca_texture, gcn_texture, proj_color,
pca_color, gcn_color, input_feat, gcn_feat, regularization, get_inter=False):
"""Adds to the inference model the layers required to generate loss."""
with tf.name_scope('loss'):
with tf.name_scope('data_loss'):
skin_mask = self._erosion2d(input_image[..., 3:])
gcn_render_mask = tf.round(gcn_render[..., 3:]) * skin_mask
# pca_render_loss = tf.losses.mean_squared_error(
pca_render_loss = tf.losses.absolute_difference(
predictions=pca_render[..., :3] * gcn_render_mask, labels=input_image[..., :3] *
gcn_render_mask, reduction=tf.losses.Reduction.SUM) / tf.reduce_sum(gcn_render_mask)
# gcn_render_loss = tf.losses.mean_squared_error(
gcn_render_loss = tf.losses.absolute_difference(
predictions=gcn_render[..., :3] * gcn_render_mask, labels=input_image[..., :3] *
gcn_render_mask, reduction=tf.losses.Reduction.SUM) / tf.reduce_sum(gcn_render_mask)
# project_loss_image = tf.losses.mean_squared_error(
project_loss_image = tf.losses.absolute_difference(
predictions=gcn_color * proj_color[..., 3:],
labels=proj_color[..., :3] * proj_color[..., 3:], reduction=tf.losses.Reduction.MEAN)
# project_loss_pca = tf.losses.mean_squared_error(
project_loss_pca = tf.losses.absolute_difference(
predictions=gcn_color * (1 - proj_color[..., 3:]),
labels=pca_color * (1 - proj_color[..., 3:]), reduction=tf.losses.Reduction.MEAN)
project_loss = project_loss_image + 0.3 * project_loss_pca
# refine_loss = tf.losses.mean_squared_error(
refine_loss = tf.losses.absolute_difference(predictions=gcn_texture, labels=pca_texture,
reduction=tf.losses.Reduction.MEAN)
perception_loss = 1 - tf.reduce_mean(utils.cosine(input_feat, gcn_feat))
var_losses = []
gcn_skin_texture = tf.gather(gcn_texture, self.bfm.skin_index, axis=1)
for i in range(3):
_, variance = tf.nn.moments(gcn_skin_texture[..., i], axes=1)
var_losses.append(variance)
var_loss = tf.reduce_mean(var_losses)
sym_diff = tf.gather(gcn_texture, self.bfm.left_index, axis=1) - tf.gather(
gcn_texture, self.bfm.right_index, axis=1)
sym_loss = tf.reduce_mean(tf.sqrt(tf.reduce_sum(tf.square(sym_diff) + 1e-16, axis=-1)))
# adj_tensor = tf.constant(self.adjacent.reshape(
# [1, self.num_vert, self.num_vert, 1]),
# dtype=tf.int32,
# shape=[1, self.num_vert, self.num_vert, 1])
# coo = self.adjacent.tocoo()
# indices = np.mat([0, self.adjacent.row, self.adjacent.col, 0]).transpose()
# values = np.ones_like(self.adjacent.data, np.float32)
# adj_tensor = tf.SparseTensor(indices, values, self.adjacent.shape)
# # adj_tensor = tf.SparseTensor(self.adjacent.indices,
# # np.clip(self.adjacent.data, 0, 1),
# # self.adjacent.shape)
# expand = tf.ones([1, self.num_vert, self.num_vert, 3], dtype=tf.float32)
# expand = expand * tf.expand_dims(gcn_texture, axis=1)
# exp_trans = tf.transpose(expand, [0, 2, 1, 3])
# # vertical = tf.ones([self.num_vert, self.num_vert, 3], dtype=tf.float32)
# # vertical = vertical * tf.expand_dims(gcn_texture, axis=2)
# smooth_loss = tf.abs((expand - exp_trans) * adj_tensor)
# test = tf.sparse_to_dense(smooth_loss.indices, )
#TODO: need attention
# data_loss = self.ph_ref_lambda * refine_loss + self.ph_ren_lambda * (
# gcn_render_loss + 0.2 * project_loss +
# 0.2 * perception_loss) + 0.1 * sym_loss
data_loss = self.ph_ref_lambda * refine_loss + self.ph_ren_lambda * (
project_loss + 0.2 * perception_loss + 0.5 * sym_loss + 0.01 * var_loss)
# if not get_inter:
# self.skin_mask = skin_mask
# self.gcn_render_mask = gcn_render_mask
# self.gcn_render_image = gcn_render[..., :3]
# self.input_image_rgb = input_image[..., :3]
# self.pca_render_image = pca_render[..., :3]
with tf.name_scope('regularization'):
regularization *= tf.add_n(self.regularizers)
loss = data_loss + regularization
tf.summary.scalar('loss/data_loss', data_loss)
tf.summary.scalar('loss/pca_render_loss', pca_render_loss)
tf.summary.scalar('loss/gcn_render_loss', gcn_render_loss)
tf.summary.scalar('loss/project_loss', project_loss)
tf.summary.scalar('loss/refine_loss', refine_loss)
tf.summary.scalar('loss/perception_loss', perception_loss)
tf.summary.scalar('loss/var_loss', var_loss)
tf.summary.scalar('loss/sym_loss', sym_loss)
tf.summary.scalar('loss/regularization', regularization)
logger.info('Successfully Computed Losses')
return loss, pca_render_loss, gcn_render_loss, project_loss, refine_loss, perception_loss, var_loss, sym_loss
def sim(a, b):
"Cosine similarity of vectors"
return cosine(vec[a], vec[b])