def write_performance(self, model_type, metrics, train_time):
def get_header_string():
header = [
'acc', 'mr_f', 'mrr_f', 'mr_p', 'mrr_p', 'pre', 'rec', 'f1',
'S', 'D', 'B', 'E', 'l_rate', 'data_set', 't_sec', 'model_type'
]
return format_list_to_string(header, '\t')
def get_perf_string(metrics, train_time):
content = []
content.extend(metrics[:-1])
content.extend([self.n_sense, self.dim, self.batch_size, self.n_epoch, self.learning_rate,\
self.regu_strength, self.dropout, self.dataset, train_time, model_type])
return format_list_to_string(content, '\t')
# quantitative analysis
header_string = get_header_string()
perf_string = get_perf_string(metrics, train_time)
print header_string + '\n' + perf_string
file_header = read_first_line(self.perf_file)
with open(self.perf_file, 'a') as fout:
if file_header != header_string:
fout.write(header_string + '\n')
fout.write(perf_string + '\n')
# error analysis
error_indicator = metrics[-1]
error_indicator_file = self.instance_analysis_path + str(
model_type) + '.txt'
ensure_directory_exist(error_indicator_file)
with open(error_indicator_file, 'w') as fout:
for element in error_indicator:
fout.write(str(element) + '\n')
def write_to_file(instances, x_vocab, y_vocab, output_file):
ensure_directory_exist(output_file)
with open(output_file, 'w') as fout:
for instance in instances:
instance_string = format_instance_to_string(
instance, x_vocab, y_vocab)
fout.write(instance_string + '\n')
def write_one_case(self, idx, neighbor_ids, scores):
neighbor_info = [idx]
description = self.vocab.get_description(idx)
neighbor_info.append(description)
for (neighbor_idx, score) in zip(neighbor_ids, scores):
description = self.vocab.get_description(neighbor_idx)
neighbor_info.append([description, score])
neighbor_info = format_list_to_string(neighbor_info, ' ')
# print neighbor_info
ensure_directory_exist(self.case_output_file)
with open(self.case_output_file, 'a') as fp:
fp.write(neighbor_info + '\n')
def write_performance(pd, model_type, metrics, train_time):
header = ['time', 'acc', 'mr_f', 'mrr_f', 'mr_p', 'mrr_p',\
'n_sense', 'dim', 'n_epoch', 'data_dir', 'model_type']
content = [train_time, metrics[0], metrics[1], metrics[2], metrics[3], metrics[4],\
pd['n_sense'], pd['embedding_dim'], pd['n_epoch'], pd['data_dir'], model_type]
header_string = format_list_to_string(header, '\t')
content_string = format_list_to_string(content, '\t')
print header_string + '\n' + content_string
# write to file
perf_file = pd['performance_file']
ensure_directory_exist(perf_file)
file_header = read_first_line(perf_file)
with open(perf_file, 'a') as fout:
if file_header != header_string:
fout.write(header_string + '\n')
fout.write(content_string + '\n')
def __init__(self, opt):
self.opt = opt
self.n_sense = opt['n_sense']
self.dim = opt['embedding_dim']
self.n_epoch = opt['n_epoch']
self.dataset = opt['data_dir'].split('/')[-2]
self.batch_size = opt['batch_size']
self.learning_rate = opt['learning_rate']
self.regu_strength = opt['regu_strength']
self.dropout = opt['dropout']
self.perf_file = opt['data_dir'] + 'output/performance.txt'
self.instance_analysis_path = opt['data_dir'] + 'instance/'
ensure_directory_exist(self.perf_file)
self.test_data_file = opt['data_dir'] + 'input/test.txt'
self.x_vocab_file = opt['data_dir'] + 'input/words.txt'
self.criterion = nn.NLLLoss()
def find_duplicate_sense_words(self):
if self.n_sense != 2:
print 'Similar sense detection is only supported for n_sense = 2.'
return
print 'Start detecting duplicate sense for %s words.' % self.n_words
words = []
for i in xrange(self.n_words):
idx = i * self.n_sense + 1
vec_a = self.emb_matrix[idx]
vec_b = self.emb_matrix[idx + 1]
if self.is_similar(vec_a, vec_b, mode='norm'):
word = self.vocab.get_description(idx)
words.append(word)
ensure_directory_exist(self.dup_sense_file)
with open(self.dup_sense_file, 'w') as fout:
for w in words:
fout.write(w + '\n')
def write_cluster_members(self, clus, cluster_file, parent_dir):
n_cluster = clus.n_cluster
clusters = clus.clusters # a dict: cluster id -> keywords
with open(cluster_file, 'w') as fout:
for clus_id in range(n_cluster):
members = clusters[clus_id]
for keyword_id in members:
keyword = self.keywords[keyword_id]
fout.write(str(clus_id) + '\t' + keyword + '\n')
# write the cluster for each sub-folder
clus_centers = clus.center_ids
for clus_id, center_keyword_id in clus_centers:
center_keyword = self.keywords[center_keyword_id]
output_file = parent_dir + center_keyword + '/seed_keywords.txt'
ensure_directory_exist(output_file)
members = clusters[clus_id]
with open(output_file, 'w') as fout:
for keyword_id in members:
keyword = self.keywords[keyword_id]
fout.write(keyword + '\n')
def write_document_membership(self, clus, output_file, parent_dir):
n_cluster = clus.n_cluster
keyword_membership = clus.membership # an array containing the membership of the keywords
cluster_document_map = defaultdict(list) # key: cluster id, value: document list
with open(output_file, 'w') as fout:
for idx, doc in zip(self.original_doc_ids, self.documents):
doc_membership = self.get_doc_membership(n_cluster, doc, keyword_membership)
cluster_id = self.assign_document(doc_membership)
cluster_document_map[cluster_id].append(idx)
fout.write(str(idx) + '\t' + str(cluster_id) + '\n')
# write the document ids for each sub-folder
clus_centers = clus.center_ids
for clus_id, center_keyword_id in clus_centers:
center_keyword = self.keywords[center_keyword_id]
output_file = parent_dir + center_keyword + '/doc_ids.txt'
ensure_directory_exist(output_file)
doc_ids = cluster_document_map[clus_id]
with open(output_file, 'w') as fout:
for doc_id in doc_ids:
fout.write(str(doc_id) + '\n')
def train_neg(train_data, model, criterion, optimizer, model_type, pd):
forward_time, backward_time = 0, 0
n_epoch = pd['n_epoch']
train_log_file = pd['train_log_file']
ensure_directory_exist(train_log_file)
with open(train_log_file, 'a') as fout:
# train
for epoch in xrange(n_epoch):
running_loss = 0.0
for i in xrange(len(train_data)):
# get the input
inputs, labels = train_data[i]
inputs = Variable(torch.LongTensor(inputs))
noise_labels = train_data.sample_negatives(5, labels[0])
labels.extend(noise_labels)
labels = Variable(torch.LongTensor(labels))
f_start_time = time.time()
output = model(inputs, labels)
f_end_time = time.time()
forward_time += f_end_time - f_start_time
loss = criterion(output)
# zero the parameter gradients
optimizer.zero_grad()
# backward + optimize
b_start_time = time.time()
loss.backward()
optimizer.step()
b_end_time = time.time()
backward_time += (b_end_time - b_start_time)
# print statistics
running_loss += loss.data[0]
if (i + 1) % 2000 == 0:
print('%20s [%d, %5d] training loss: %.3f' % (model_type, epoch+1, i+1, running_loss/2000))
fout.write('%20s [%d, %5d] training loss: %.3f\n' % (model_type, epoch+1, i+1, running_loss/20))
running_loss = 0.0
print 'forward time:', forward_time
print 'backward time:', backward_time
def save_model(self, model, model_type):
model_name = self.get_model_name(model_type)
file_name = self.model_path + model_name
ensure_directory_exist(file_name)
torch.save(model.state_dict(), file_name)