def __init__(self, args, reduced_size=None, info={}):
super(CNN, self).__init__()
# disc_type=DISC_TYPE_MATRIX
self.disc_type = disc_type = args.disc_type
self.layer1 = nn.Sequential(
nn.Conv2d(1, 4, kernel_size=2, padding=0),
nn.ReLU())
# 1,4,3,3
self.layer2 = nn.Sequential(
nn.Conv2d(4, 8, kernel_size=2),
nn.ReLU())
# 1,8,2,2
## but for 5 lines, it is 1,8,3,3
if args.data_type == "sonnet_endings":
self.scorer = nn.Linear(2 * 2 * 8, 1)
elif args.data_type == "limerick":
self.scorer = nn.Linear(3 * 3 * 8, 1)
self.predictor = nn.Sigmoid()
self.args = args
self.use_cuda = args.use_cuda
##
self.g_indexer = Indexer(args)
self.g_indexer.load('tmp/tmp_' + args.g2p_model_name + '/solver_g_indexer')
self.g2pmodel = Model(H=info['H'], args=args, i_size=self.g_indexer.w_cnt, o_size=self.g_indexer.w_cnt,
start_idx=self.g_indexer.w2idx[utils.START])
if not args.learn_g2p_encoder_from_scratch:
print("=====" * 7, "LOADING g2p ENCODER PRETRAINED")
model_dir = 'tmp/tmp_' + args.g2p_model_name + '/'
state_dict_best = torch.load(model_dir + 'model_best')
self.g2pmodel.load_state_dict(state_dict_best)
if not args.trainable_g2p:
assert not args.learn_g2p_encoder_from_scratch
for param in self.g2pmodel.parameters():
param.requires_grad = False
class Crawler:
def __init__(self, list_file):
self.logger = Logger.get_logger(utils.get_fullname(self))
self.list_file = list_file
self.indexer = Indexer()
self._client = None
# return raw list in list format
def parse_list(self, list_file):
try:
self.logger.info('Opening RSS file: %s' % list_file)
f = open(list_file, 'r')
except IOError:
self.logger.error('Cannot read file: %s' % list_file)
return -1
self.feeds_list = []
line = f.readline()
while line:
self.logger.debug('Reading: %s' % line)
feeds = feedparser.parse(line)
try:
# default only get the latest entry
raw_f = feeds['entries'][0]
feed_item = EzrssFeed(raw_f.link)
feed_item.parse_name(raw_f.summary or raw_f.value)
feed_item.parse_season(raw_f.summary or raw_f.value)
self.feeds_list.append(feed_item)
except IndexError:
pass
line = f.readline()
@property
def client(self):
return self._client
@client.setter
def client(self, c):
self._client = c
def run(self):
self.parse_list(self.list_file)
self.logger.info('Start checking latest RSS feeds')
for feed in self.feeds_list:
if feed.name:
if not self.indexer.episode_exists(feed.name, feed.url):
save_path = os.path.join(getattr(settings, 'SAVE_DIR'),
feed.name,
feed.season)
self.client.start_from_url(feed.url, save_path)
self.indexer.save(feed.name, feed.url)
else:
continue
self.logger.info('Exiting application')
def from_word_vectors(cls, word_vectors, unique_labels):
"""Instantiate the vectorizer"""
review_vocab = word_vectors
rating_vocab = Indexer()
# Add ratings
for l in unique_labels:
rating_vocab.add_and_get_index(l)
return cls(review_vocab, rating_vocab)
def __init__(self, max_word_length):
vocab = [chr(ord('a') + i) for i in range(0, 26)] + [' ']
self.char_vocab_index = Indexer()
self.char_vocab_index.add_and_get_index(PAD_TOKEN) # PAD is 0
self.char_vocab_index.add_and_get_index(
UNK_TOKEN) # Unknown token is 1
for char in vocab:
self.char_vocab_index.add_and_get_index(char)
self.max_word_length = max_word_length
def __init__(self, args):
super().__init__()
self.args = args
self.pad_token_id = args.pad_token_id
# Initialize embedding layer (1)
self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
# Initialize char embedding layer
self.char_embedding = nn.Embedding(args.char_vocab_size,
args.char_embedding_dim)
# Initialize Context2Query (2)
self.aligned_att = AlignedAttention(args.embedding_dim,
args.char_embedding_dim)
rnn_cell = nn.LSTM if args.rnn_cell_type == 'lstm' else nn.GRU
# Initialize passage encoder (3)
self.passage_rnn = rnn_cell(
args.embedding_dim * 2,
args.hidden_dim,
bidirectional=args.bidirectional,
batch_first=True,
)
# Initialize question encoder (4)
self.question_rnn = rnn_cell(
args.embedding_dim,
args.hidden_dim,
bidirectional=args.bidirectional,
batch_first=True,
)
self.dropout = nn.Dropout(self.args.dropout)
# Adjust hidden dimension if bidirectional RNNs are used
_hidden_dim = (args.hidden_dim *
2 if args.bidirectional else args.hidden_dim)
# Initialize attention layer for question attentive sum (5)
self.question_att = SpanAttention(_hidden_dim)
# Initialize bilinear layer for start positions (6)
self.start_output = BilinearOutput(_hidden_dim, _hidden_dim)
# Initialize bilinear layer for end positions (7)
self.end_output = BilinearOutput(_hidden_dim, _hidden_dim)
# Initialize char indexer
vocab = [chr(ord('a') + i) for i in range(0, 26)] + [' ']
self.char_vocab_index = Indexer()
for char in vocab:
self.char_vocab_index.add_and_get_index(char)
def read_word_embeddings(embeddings_file: str) -> WordEmbeddings:
"""
Loads the given embeddings (ASCII-formatted) into a WordEmbeddings object. Augments this with an UNK embedding
that is the 0 vector. Reads in all embeddings with no filtering -- you should only use this for relativized
word embedding files.
:param embeddings_file: path to the file containing embeddings
:return: WordEmbeddings object reflecting the words and their embeddings
"""
f = open(embeddings_file)
word_indexer = Indexer()
vectors = []
# Make position 0 a PAD token, which can be useful if you
word_indexer.add_and_get_index("PAD")
# Make position 1 the UNK token
word_indexer.add_and_get_index("UNK")
for line in f:
if line.strip() != "":
space_idx = line.find(' ')
word = line[:space_idx]
numbers = line[space_idx + 1:]
float_numbers = [
float(number_str) for number_str in numbers.split()
]
vector = np.array(float_numbers)
word_indexer.add_and_get_index(word)
# Append the PAD and UNK vectors to start. Have to do this weirdly because we need to read the first line
# of the file to see what the embedding dim is
if len(vectors) == 0:
vectors.append(np.zeros(vector.shape[0]))
vectors.append(np.zeros(vector.shape[0]))
vectors.append(vector)
f.close()
# Turn vectors into a 2-D numpy array
return WordEmbeddings(word_indexer, np.array(vectors))
class CharTokenizer:
"""
Class to create char tokens
"""
def __init__(self, max_word_length):
vocab = [chr(ord('a') + i) for i in range(0, 26)] + [' ']
self.char_vocab_index = Indexer()
self.char_vocab_index.add_and_get_index(PAD_TOKEN) # PAD is 0
self.char_vocab_index.add_and_get_index(
UNK_TOKEN) # Unknown token is 1
for char in vocab:
self.char_vocab_index.add_and_get_index(char)
self.max_word_length = max_word_length
def convert_words_to_charids(self, words):
word_charids = []
for w in words:
charids = []
for c in w:
charids.append(self.char_vocab_index.index_of(c))
charids = charids[:self.max_word_length]
if len(charids) < self.max_word_length:
charids.extend([0] * (self.max_word_length - len(charids)))
word_charids.append(charids)
return word_charids
def load(self, specialTokenList=None):
indexer = Indexer(specialTokenList)
print "... loading training data."
trainPairs,trainLens = self._load_pairs(indexer,
self.dataDict['train_source'],
self.dataDict['train_target'])
print "... loading test data."
testPairs,testLens = self._load_pairs(indexer,
self.dataDict['test_source'],
self.dataDict['test_source'])
print "Done!\n"
return indexer,trainPairs,trainLens,testPairs,testLens
def generate_indexer(usr_dataset, usr_bm_tg, feature_begin, feature_end):
logging.info('generating indexer ...')
indexer = Indexer(['user', 'tag', 'bookmark'])
min_time = 1e30
max_time = -1
for line in usr_dataset[1:]:
line_items = line.split('\t')
contact_timestamp = float(line_items[2]) / 1000
min_time = min(min_time, contact_timestamp)
max_time = max(max_time, contact_timestamp)
if feature_begin < contact_timestamp <= feature_end:
indexer.index('user', line_items[0])
indexer.index('user', line_items[1])
for line in usr_bm_tg[1:]:
line_items = line.split('\t')
tag_timestamp = float(line_items[3]) / 1000
if feature_begin < tag_timestamp <= feature_end:
indexer.index('user', line_items[0])
indexer.index('bookmark', line_items[1])
indexer.index('tag', line_items[2])
with open('delicious/data/metadata.txt', 'w') as output:
output.write('Nodes:\n')
output.write('-----------------------------\n')
output.write('#Users: %d\n' % indexer.indices['user'])
output.write('#Tags: %d\n' % indexer.indices['tag'])
output.write('#Bookmarks: %d\n' % indexer.indices['bookmark'])
output.write('\nEdges:\n')
output.write('-----------------------------\n')
output.write('#Contact: %d\n' % len(usr_dataset))
output.write('#Save : %d\n' % len(usr_bm_tg))
output.write('#Attach: %d\n' % len(usr_bm_tg))
output.write('\nTime Span:\n')
output.write('-----------------------------\n')
output.write('From: %s\n' % datetime.fromtimestamp(min_time))
output.write('To: %s\n' % datetime.fromtimestamp(max_time))
return indexer
class FeatureExtractor():
def __init__(self):
self.indexer = Indexer()
def get_indexer(self):
return self.indexer
def extract_features(self, ex):
feature_vector = np.zeros(len(self.indexer))
for word in ex.text:
index = self.indexer.index_of(word)
feature_vector[index] += 1
return feature_vector
contains a header line and 45463 data lines,
each line includes a mId and its overview (some sentences).
'''
movies.to_csv("processed_data/overviews.csv",
columns=['mId', 'overview'],
index=False)
movies.to_csv("processed_data/mId2Title.csv",
columns=['mId', 'tmdbId', 'title'],
index=False)
''' create genres
mId2Genre: 45463 lines, each line includes (mId, num of genres, gIds)
Genre2Id: 20 lines, each line includes (gId, genre name)
gId ranges from 45843 to 45862
'''
f = open("processed_data/mId2Genre.txt", "w")
genreIdx = Indexer()
for idx, row in movies.iterrows():
mId, raw_genres = row['mId'], row['genres']
raw_genres = raw_genres.replace("\'", "\"")
genres_l = json.loads(raw_genres)
f.write("%d %d" % (mId, len(genres_l)))
for g in genres_l:
f.write(" %d" % (genreIdx.add_and_get_index(g['name']) + id_base))
f.write("\n")
f.close()
f = open("processed_data/Genre2Id.txt", "w")
num_genres = len(genreIdx)
for i in range(num_genres):
f.write("%d %s\n" % (i + id_base, genreIdx.get_object(i)))
f.close()
labels, emoji_labels = get_labels(tweet, indexer)
label = get_most_recent_label(tweet, emoji_labels, indexer)
except:
continue
cleaned_text = clean_tweet(tweet)
datapoint = DataPoint(cleaned_text, label)
dataset.append(datapoint)
label_counter[indexer.get_object(label)] += 1
count += 1
if count % 500000 == 0:
print("created", count, "datapoints")
return dataset
indexer = Indexer()
label_counter = Counter()
dataset = create_dataset(tweets, indexer, label_counter
print ("length of dataset: ", len(dataset))
from tokenizer import tokenizer as vinay
v = vinay.TweetTokenizer(regularize=True, preserve_len=False)
word_cnts = Counter()
def count_words(text):
words = v.tokenize(text)
for word in words:
word_cnts[word] += 1
class CharBaselineReader(nn.Module):
"""
Baseline QA Model
[Architecture]
0) Inputs: passages and questions
1) Embedding Layer: converts words to vectors
2) Context2Query: computes weighted sum of question embeddings for
each position in passage.
3) Passage Encoder: LSTM or GRU.
4) Question Encoder: LSTM or GRU.
5) Question Attentive Sum: computes weighted sum of question hidden.
6) Start Position Pointer: computes scores (logits) over passage
conditioned on the question vector.
7) End Position Pointer: computes scores (logits) over passage
conditioned on the question vector.
Args:
args: `argparse` object.
Inputs:
batch: a dictionary containing batched tensors.
{
'passages': LongTensor [batch_size, p_len],
'questions': LongTensor [batch_size, q_len],
'start_positions': Not used in `forward`,
'end_positions': Not used in `forward`,
}
Returns:
Logits for start positions and logits for end positions.
Tuple: ([batch_size, p_len], [batch_size, p_len])
"""
def __init__(self, args):
super().__init__()
self.args = args
self.pad_token_id = args.pad_token_id
# Initialize embedding layer (1)
self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
# Initialize char embedding layer
self.char_embedding = nn.Embedding(args.char_vocab_size,
args.char_embedding_dim)
# Initialize Context2Query (2)
self.aligned_att = AlignedAttention(args.embedding_dim,
args.char_embedding_dim)
rnn_cell = nn.LSTM if args.rnn_cell_type == 'lstm' else nn.GRU
# Initialize passage encoder (3)
self.passage_rnn = rnn_cell(
args.embedding_dim * 2,
args.hidden_dim,
bidirectional=args.bidirectional,
batch_first=True,
)
# Initialize question encoder (4)
self.question_rnn = rnn_cell(
args.embedding_dim,
args.hidden_dim,
bidirectional=args.bidirectional,
batch_first=True,
)
self.dropout = nn.Dropout(self.args.dropout)
# Adjust hidden dimension if bidirectional RNNs are used
_hidden_dim = (args.hidden_dim *
2 if args.bidirectional else args.hidden_dim)
# Initialize attention layer for question attentive sum (5)
self.question_att = SpanAttention(_hidden_dim)
# Initialize bilinear layer for start positions (6)
self.start_output = BilinearOutput(_hidden_dim, _hidden_dim)
# Initialize bilinear layer for end positions (7)
self.end_output = BilinearOutput(_hidden_dim, _hidden_dim)
# Initialize char indexer
vocab = [chr(ord('a') + i) for i in range(0, 26)] + [' ']
self.char_vocab_index = Indexer()
for char in vocab:
self.char_vocab_index.add_and_get_index(char)
def load_pretrained_embeddings(self, vocabulary, path):
"""
Loads GloVe vectors and initializes the embedding matrix.
Args:
vocabulary: `Vocabulary` object.
path: Embedding path, e.g. "glove/glove.6B.300d.txt".
"""
if self.args.embedding == 'glove':
embedding_map = load_cached_embeddings(path)
# Create embedding matrix. By default, embeddings are randomly
# initialized from Uniform(-0.1, 0.1).
embeddings = torch.zeros(
(len(vocabulary),
self.args.embedding_dim)).uniform_(-0.1, 0.1)
# Initialize pre-trained embeddings.
num_pretrained = 0
for (i, word) in enumerate(vocabulary.words):
if word in embedding_map:
#embeddings[i] = torch.tensor(embedding_map[word])
num_pretrained += 1
# Place embedding matrix on GPU.
self.embedding.weight.data = cuda(self.args, embeddings)
else:
#####################
# Loads Fasttext embeddings
embedding_map = load_fasttext_embeddings(path)
# Create embedding matrix. By default, embeddings are randomly
# initialized from Uniform(-0.1, 0.1).
embeddings = torch.zeros(
(len(vocabulary),
self.args.embedding_dim)).uniform_(-0.1, 0.1)
# Initialize pre-trained embeddings.
num_pretrained = 0
for (i, word) in enumerate(vocabulary.words):
embeddings[i] = torch.tensor(
embedding_map.get_word_vector(word))
num_pretrained += 1
# Place embedding matrix on GPU.
self.embedding.weight.data = cuda(self.args, embeddings)
return num_pretrained
def sorted_rnn(self, sequences, sequence_lengths, rnn):
"""
Sorts and packs inputs, then feeds them into RNN.
Args:
sequences: Input sequences, [batch_size, len, dim].
sequence_lengths: Lengths for each sequence, [batch_size].
rnn: Registered LSTM or GRU.
Returns:
All hidden states, [batch_size, len, hid].
"""
# Sort input sequences
sorted_inputs, sorted_sequence_lengths, restoration_indices = _sort_batch_by_length(
sequences, sequence_lengths)
# Pack input sequences
packed_sequence_input = pack_padded_sequence(
sorted_inputs,
sorted_sequence_lengths.data.long().tolist(),
batch_first=True)
# Run RNN
packed_sequence_output, _ = rnn(packed_sequence_input, None)
# Unpack hidden states
unpacked_sequence_tensor, _ = pad_packed_sequence(
packed_sequence_output, batch_first=True)
# Restore the original order in the batch and return all hidden states
return unpacked_sequence_tensor.index_select(0, restoration_indices)
def forward(self, batch):
# Obtain masks and lengths for passage and question.
passage_mask = (batch['passages'] != self.pad_token_id
) # [batch_size, p_len]
question_mask = (batch['questions'] != self.pad_token_id
) # [batch_size, q_len]
passage_lengths = passage_mask.long().sum(-1) # [batch_size]
question_lengths = question_mask.long().sum(-1) # [batch_size]
# 1) Embedding Layer: Embed the passage and question.
passage_embeddings = self.embedding(
batch['passages']) # [batch_size, p_len, p_dim]
question_embeddings = self.embedding(
batch['questions']) # [batch_size, q_len, q_dim]
passage_char_embeddings = self.char_embedding(
batch['char_passages']
) # [batch_size, p_len, word_length, word_dim] [64, 168, 16, 64]
question_char_embeddings = self.char_embedding(
batch['char_questions']
) # [batch_size, q_len, word_length, word_dim]
if self.args.char_embedding_type == 'average':
# Average char embeddings baseline
passage_char_embeddings_avg = passage_char_embeddings.mean(
dim=2).squeeze(0)
question_char_embeddings_avg = question_char_embeddings.mean(
dim=2).squeeze(0)
passage_final_embeddings = torch.cat(
[passage_embeddings, passage_char_embeddings_avg], dim=2)
question_final_embeddings = torch.cat(
[question_embeddings, question_char_embeddings_avg], dim=2)
#print('passage_char_embeddings ', passage_char_embeddings.shape)
#print('question_char_embeddings ', question_char_embeddings.shape)
else:
# Conv 1D char embeddings
passage_char_embeddings_conv1d_input = passage_char_embeddings.reshape(
(-1, passage_char_embeddings.shape[3],
passage_char_embeddings.shape[2]))
question_char_embeddings_conv1d_input = question_char_embeddings.reshape(
(-1, question_char_embeddings.shape[3],
question_char_embeddings.shape[2]))
conv1d = torch.nn.Conv1d(self.args.char_embedding_dim,
self.args.char_embedding_dim, 3)
relu = torch.nn.ReLU()
if torch.cuda.is_available():
conv1d.cuda()
relu.cuda()
passage_char_embeddings_tmp1 = relu(
conv1d(passage_char_embeddings_conv1d_input))
# Last dimension of conv1d output we want to collapse using global max pooling
passage_char_embeddings_final = torch.nn.functional.max_pool1d(
passage_char_embeddings_tmp1,
passage_char_embeddings_tmp1.shape[2]).squeeze(2).reshape(
passage_char_embeddings.shape[0],
passage_char_embeddings.shape[1], -1)
question_char_embeddings_tmp1 = relu(
conv1d(question_char_embeddings_conv1d_input))
# Last dimension of conv1d output we want to collapse using global max pooling
question_char_embeddings_final = torch.nn.functional.max_pool1d(
question_char_embeddings_tmp1,
question_char_embeddings_tmp1.shape[2]).squeeze(2).reshape(
question_char_embeddings.shape[0],
question_char_embeddings.shape[1], -1)
passage_final_embeddings = torch.cat(
[passage_embeddings, passage_char_embeddings_final], dim=2)
question_final_embeddings = torch.cat(
[question_embeddings, question_char_embeddings_final], dim=2)
# 2) Context2Query: Compute weighted sum of question embeddings for
# each passage word and concatenate with passage embeddings.
aligned_scores = self.aligned_att(
passage_final_embeddings, question_final_embeddings,
~question_mask) # [batch_size, p_len, q_len]
aligned_embeddings = aligned_scores.bmm(
question_embeddings) # [batch_size, p_len, q_dim]
passage_embeddings = cuda(
self.args,
torch.cat((passage_embeddings, aligned_embeddings), 2),
) # [batch_size, p_len, p_dim + q_dim]
# 3) Passage Encoder
passage_hidden = self.sorted_rnn(
passage_embeddings, passage_lengths,
self.passage_rnn) # [batch_size, p_len, p_hid]
passage_hidden = self.dropout(
passage_hidden) # [batch_size, p_len, p_hid]
# 4) Question Encoder: Encode question embeddings.
question_hidden = self.sorted_rnn(
question_embeddings, question_lengths,
self.question_rnn) # [batch_size, q_len, q_hid]
# 5) Question Attentive Sum: Compute weighted sum of question hidden
# vectors.
question_scores = self.question_att(question_hidden, ~question_mask)
question_vector = question_scores.unsqueeze(1).bmm(
question_hidden).squeeze(1)
question_vector = self.dropout(question_vector) # [batch_size, q_hid]
# 6) Start Position Pointer: Compute logits for start positions
start_logits = self.start_output(passage_hidden, question_vector,
~passage_mask) # [batch_size, p_len]
# 7) End Position Pointer: Compute logits for end positions
end_logits = self.end_output(passage_hidden, question_vector,
~passage_mask) # [batch_size, p_len]
return start_logits, end_logits # [batch_size, p_len], [batch_size, p_len]
from solver import Solver
from preprocess.tacotron.utils import spectrogram2wav
#from preprocess.tacotron.audio import inv_spectrogram, save_wav
from scipy.io.wavfile import write
from preprocess.tacotron.mcep import mc2wav
if __name__ == '__main__':
feature = 'sp'
hps = Hps()
hps.load('./hps/v19.json')
hps_tuple = hps.get_tuple()
solver = Solver(hps_tuple, None)
solver.load_model('/storage/model/voice_conversion/v19/model.pkl-59999')
if feature == 'mc':
# indexer to extract data
indexer = Indexer()
src_mc = indexer.index(speaker_id='225',
utt_id='366',
dset='test',
feature='norm_mc')
tar_mc = indexer.index(speaker_id='226',
utt_id='366',
dset='test',
feature='norm_mc')
expand_src_mc = np.expand_dims(src_mc, axis=0)
expand_tar_mc = np.expand_dims(tar_mc, axis=0)
src_mc_tensor = torch.from_numpy(expand_src_mc).type(torch.FloatTensor)
tar_mc_tensor = torch.from_numpy(expand_tar_mc).type(torch.FloatTensor)
c1 = Variable(torch.from_numpy(np.array([0]))).cuda()
c2 = Variable(torch.from_numpy(np.array([1]))).cuda()
results = [src_mc]
from keras.models import Sequential
from keras.layers import Dense, Embedding, LSTM, Dropout
from sklearn.model_selection import train_test_split
from keras.utils.np_utils import to_categorical
import re
import json
import pandas as pd
from utils import get_train_data_from_csv, get_dev_data_from_csv, get_test_data_from_csv, Indexer, get_indexer
from nltk.tokenize import TweetTokenizer
from sklearn.metrics import classification_report
include_test = True
tknr = TweetTokenizer()
indexer = get_indexer('indexer_15_dups.csv')
word_indexer = Indexer()
word_indexer.add_and_get_index("UNK")
train_data = get_train_data_from_csv('data/train_15_ds.csv')[0:1000]
dev_data = get_dev_data_from_csv('data/dev_15_ds.csv')[:200]
test_data = get_test_data_from_csv('data/test_15_ds.csv')[0:200]
X_train = []
Y_train = []
X_dev = []
Y_dev = []
Y_dev_true = []
X_test = []
Y_test = []
Y_test_true = []
def __init__(self):
self.indexer = Indexer()
def generate_papers(datafile, feature_begin, feature_end, observation_begin,
observation_end, conf_list):
logging.info('generating papers ...')
# try:
# result = pickle.load(open('dblp/data/papers_%s.pkl' % path, 'rb'))
# return result
# except IOError:
# pass
indexer = Indexer(['author', 'paper', 'term', 'venue'])
index, authors, title, year, venue = None, None, None, None, None
references = []
write = 0
cite = 0
include = 0
published = 0
min_year = 3000
max_year = 0
papers_feature_window = []
papers_observation_window = []
with open(datafile) as file:
dataset = file.read().splitlines()
for line in dataset:
if not line:
if year and venue:
year = int(year)
if year > 0 and authors and venue in conf_list:
min_year = min(min_year, year)
max_year = max(max_year, year)
authors = authors.split(',')
terms = parse_term(title)
write += len(authors)
cite += len(references)
include += len(terms)
published += 1
p = Paper(year)
if feature_begin < year <= feature_end:
p.id = indexer.index('paper', index)
p.terms = [
indexer.index('term', term) for term in terms
]
p.references = [
indexer.index('paper', paper_id)
for paper_id in references
]
p.authors = [
indexer.index('author', author_name)
for author_name in authors
]
p.venue = indexer.index('venue', venue)
bisect.insort(papers_feature_window, p)
elif observation_begin < year <= observation_end:
p.references = references
p.authors = authors
papers_observation_window.append(p)
index, authors, title, year, venue = None, None, None, None, None
references = []
else:
begin = line[1]
if begin == '*':
title = line[2:]
elif begin == '@':
authors = line[2:]
elif begin == 't':
year = line[2:]
elif begin == 'c':
venue = line[2:]
elif begin == 'i':
index = line[6:]
elif begin == '%':
references.append(line[2:])
for p in papers_observation_window:
authors = []
references = []
for author in p.authors:
author_id = indexer.get_index('author', author)
if author_id is not None:
authors.append(author_id)
for ref in p.references:
paper_id = indexer.get_index('paper', ref)
if paper_id is not None:
references.append(paper_id)
p.authors = authors
p.references = references
with open('dblp/data/metadata_%s.txt' % path, 'w') as output:
output.write('Nodes:\n')
output.write('-----------------------------\n')
output.write('#Authors: %d\n' % indexer.indices['author'])
output.write('#Papers: %d\n' % indexer.indices['paper'])
output.write('#Venues: %d\n' % indexer.indices['venue'])
output.write('#Terms: %d\n\n' % indexer.indices['term'])
output.write('\nEdges:\n')
output.write('-----------------------------\n')
output.write('#Write: %d\n' % write)
output.write('#Cite: %d\n' % cite)
output.write('#Publish: %d\n' % published)
output.write('#Contain: %d\n' % include)
output.write('\nTime Span:\n')
output.write('-----------------------------\n')
output.write('From: %s\n' % min_year)
output.write('To: %s\n' % max_year)
result = papers_feature_window, papers_observation_window, indexer.indices
# pickle.dump(result, open('dblp/data/papers_%s.pkl' % path, 'wb'))
return result
def generate_indexer(user_rates_movies_ds, user_tags_movies_ds, movie_actor_ds,
movie_director_ds, movie_genre_ds, movie_countries_ds,
feature_begin, feature_end):
logging.info('generating indexer ...')
min_time = 1e30
max_time = -1
indexer = Indexer(
['user', 'tag', 'movie', 'actor', 'director', 'genre', 'country'])
for line in user_rates_movies_ds[1:]:
line_items = line.split('\t')
rating_timestamp = float(line_items[3]) / 1000
min_time = min(min_time, rating_timestamp)
max_time = max(max_time, rating_timestamp)
rating = float(line_items[2])
if feature_begin < rating_timestamp <= feature_end and rating > rating_threshold:
indexer.index('user', line_items[0])
indexer.index('movie', line_items[1])
for line in user_tags_movies_ds[1:]:
line_items = line.split('\t')
tag_timestamp = float(line_items[3]) / 1000
if feature_begin < tag_timestamp <= feature_end:
indexer.index('user', line_items[0])
indexer.index('movie', line_items[1])
indexer.index('tag', line_items[2])
for line in movie_actor_ds[1:]:
line_items = line.split('\t')
ranking = int(line_items[3])
if ranking < actor_threshold and line_items[0] in indexer.mapping[
'movie']:
# indexer.index('movie', line_items[0])
indexer.index('actor', line_items[1])
for line in movie_director_ds[1:]:
line_items = line.split('\t')
if line_items[0] in indexer.mapping['movie']:
# indexer.index('movie', line_items[0])
indexer.index('director', line_items[1])
for line in movie_genre_ds[1:]:
line_items = line.split('\t')
if line_items[0] in indexer.mapping['movie']:
# indexer.index('movie', line_items[0])
indexer.index('genre', line_items[1])
for line in movie_countries_ds[1:]:
line_items = line.split('\t')
if line_items[0] in indexer.mapping['movie']:
# indexer.index('movie', line_items[0])
indexer.index('country', line_items[1])
with open('movielens/data/metadata.txt', 'w') as output:
output.write('Nodes:\n')
output.write('-----------------------------\n')
output.write('#Users: %d\n' % indexer.indices['user'])
output.write('#Tags: %d\n' % indexer.indices['tag'])
output.write('#Movies: %d\n' % indexer.indices['movie'])
output.write('#Actors: %d\n' % indexer.indices['actor'])
output.write('#Director: %d\n' % indexer.indices['director'])
output.write('#Genre: %d\n' % indexer.indices['genre'])
output.write('#Countriy: %d\n' % indexer.indices['country'])
output.write('\nEdges:\n')
output.write('-----------------------------\n')
output.write('#Rate: %d\n' % len(user_rates_movies_ds))
output.write('#Attach: %d\n' % len(user_tags_movies_ds))
output.write('#Played_by: %d\n' % len(movie_actor_ds))
output.write('#Directed_by : %d\n' % len(movie_director_ds))
output.write('#Has: %d\n' % len(movie_genre_ds))
output.write('#Produced_in: %d\n' % len(movie_countries_ds))
output.write('\nTime Span:\n')
output.write('-----------------------------\n')
output.write('From: %s\n' % datetime.fromtimestamp(min_time))
output.write('To: %s\n' % datetime.fromtimestamp(max_time))
return indexer
class CNN(nn.Module):
def __init__(self, args, reduced_size=None, info={}):
super(CNN, self).__init__()
# disc_type=DISC_TYPE_MATRIX
self.disc_type = disc_type = args.disc_type
self.layer1 = nn.Sequential(
nn.Conv2d(1, 4, kernel_size=2, padding=0),
nn.ReLU())
# 1,4,3,3
self.layer2 = nn.Sequential(
nn.Conv2d(4, 8, kernel_size=2),
nn.ReLU())
# 1,8,2,2
## but for 5 lines, it is 1,8,3,3
if args.data_type == "sonnet_endings":
self.scorer = nn.Linear(2 * 2 * 8, 1)
elif args.data_type == "limerick":
self.scorer = nn.Linear(3 * 3 * 8, 1)
self.predictor = nn.Sigmoid()
self.args = args
self.use_cuda = args.use_cuda
##
self.g_indexer = Indexer(args)
self.g_indexer.load('tmp/tmp_' + args.g2p_model_name + '/solver_g_indexer')
self.g2pmodel = Model(H=info['H'], args=args, i_size=self.g_indexer.w_cnt, o_size=self.g_indexer.w_cnt,
start_idx=self.g_indexer.w2idx[utils.START])
if not args.learn_g2p_encoder_from_scratch:
print("=====" * 7, "LOADING g2p ENCODER PRETRAINED")
model_dir = 'tmp/tmp_' + args.g2p_model_name + '/'
state_dict_best = torch.load(model_dir + 'model_best')
self.g2pmodel.load_state_dict(state_dict_best)
if not args.trainable_g2p:
assert not args.learn_g2p_encoder_from_scratch
for param in self.g2pmodel.parameters():
param.requires_grad = False
def display_params(self):
print("=" * 44)
print("[CNN]: model parametrs")
for name, param in self.named_parameters():
print("name=", name, " || grad:", param.requires_grad, "| size = ", param.size())
print("=" * 44)
def _compute_word_reps(self, words_str, deb=False):
if deb:
print("words_str = ", words_str)
use_eow_marker = self.args.use_eow_in_enc
assert not use_eow_marker, "Not yet tested"
word_reps = [self.g_indexer.w_to_idx(s1) for s1 in words_str]
if self.args.use_eow_in_enc:
x_end = self.g_indexer.w2idx[utils.END]
word_reps = [x_i + [x_end] for x_i in word_reps]
word_reps = [self.g2pmodel.encode(w) for w in word_reps]
return word_reps
def _compute_pairwise_dot(self, measure_encodings_b):
ret = []
sz = len(measure_encodings_b)
for measure_encodings_b_t in measure_encodings_b:
for measure_encodings_b_t2 in measure_encodings_b:
t1 = torch.sum(measure_encodings_b_t * measure_encodings_b_t2)
t2 = torch.sqrt(torch.sum(measure_encodings_b_t * measure_encodings_b_t))
t3 = torch.sqrt(torch.sum(measure_encodings_b_t2 * measure_encodings_b_t2))
assert t2 > 0
assert t3 > 0, "t3=" + str(t3)
ret.append(t1 / (t2 * t3))
ret = torch.stack(ret)
ret = ret.view(sz, sz)
return ret
def _score_matrix(self, x, deb=False):
x = x[0].unsqueeze(0).unsqueeze(0) # -> 1,1,ms,ms
if deb:
print("---x.shape = ", x.size())
out = self.layer1(x)
if deb:
print("---out = ", out.size(), out)
out = self.layer2(out)
if deb:
print("---out = ", out.size(), out)
out = out.view(out.size(0), -1) # arrange by bsz
score = self.scorer(out)
if deb:
print("---out sum = ", torch.sum(out))
print("---score = ", score)
prob = self.predictor(score)
return {'prob': prob, 'out': out, 'score': score}
def _compute_rhyming_matrix(self, words_str, deb=False):
word_reps = self._compute_word_reps(words_str)
rhyming_matrix = self._compute_pairwise_dot(word_reps)
return rhyming_matrix, words_str
def _compute_rnn_on_word_reps(self, word_reps):
h = torch.zeros(1, self.linear_rep_H), torch.zeros(1, self.linear_rep_H)
if self.use_cuda:
h = h[0].cuda(), h[1].cuda()
for w in word_reps:
h = self.linear_rep_encoder(w, h)
out, c = h
return c
def _run_discriminator(self, words_str, deb):
rhyming_matrix, words_str = self._compute_rhyming_matrix(words_str, deb)
vals = self._score_matrix([rhyming_matrix])
vals.update({'rhyming_matrix': rhyming_matrix, 'linear_rep': None, 'words_str': words_str})
return vals
def update_discriminator(self, line_endings_gen, line_endings_train, deb=False, word_idx_to_str_dict=None):
eps = 0.0000000001
ret = {}
dump_info = {}
words_str_train = [word_idx_to_str_dict[word_idx.data.cpu().item()] for word_idx in line_endings_train]
words_str_gen = [word_idx_to_str_dict[word_idx.data.cpu().item()] for word_idx in line_endings_gen]
disc_real = self._run_discriminator(words_str_train, deb)
if deb:
print("rhyming_matrix_trai = ", disc_real['rhyming_matrix'], "|| prob = ", disc_real['prob'])
if self.args.disc_type == DISC_TYPE_MATRIX:
dump_info['rhyming_matrix_trai'] = disc_real['rhyming_matrix'].data.cpu().numpy()
dump_info['real_prob'] = disc_real['prob'].data.cpu().item()
dump_info['real_words_str'] = disc_real['words_str']
disc_gen = self._run_discriminator(words_str_gen, deb)
if deb:
print("rhyming_matrix_gen = ", disc_gen['rhyming_matrix'], "|| prob = ", disc_gen['prob'])
if self.args.disc_type == DISC_TYPE_MATRIX:
dump_info['rhyming_matrix_gen'] = disc_gen['rhyming_matrix'].data.cpu().numpy()
dump_info['gen_prob'] = disc_gen['prob'].data.cpu().item()
dump_info['gen_words_str'] = disc_gen['words_str']
prob_real = disc_real['prob']
prob_gen = disc_gen['prob']
loss = -torch.log(prob_real + eps) - torch.log(1.0 - prob_gen + eps)
reward = prob_gen
if self.args.use_score_as_reward:
reward = disc_gen['score']
ret.update({'loss': loss, 'reward': reward, 'dump_info': dump_info})
return ret
def __init__(self, list_file): self.logger = Logger.get_logger(utils.get_fullname(self)) self.list_file = list_file self.indexer = Indexer() self._client = None
reload(sys)
sys.setdefaultencoding('utf-8')
app = Flask(__name__)
app.secret_key = b'_5#y2L"F4Qas5nb113@&B#(V!*#8z\n\xec]/'
db = Database()
recommender = Recommender()
if not db.checkConnectivity():
print 'Unable to connect to database'
sys.exit(-1)
indexer = Indexer()
@app.before_request
def authenticateUser():
if request.endpoint != 'search' and request.endpoint != 'signIn' and 'userid' not in session:
return redirect(url_for('signIn'))
@app.route('/signout')
def signOut():
session.clear()
return redirect(url_for('signIn'))
@app.route('/signin', methods=['GET', 'POST'])