vocab.update(line.strip() for line in f)
# read embeddings file
if options.w2v_format:
words, embs = read_text_embs(options.vectors)
else:
words, embs = read_pickle_embs(options.vectors)
dim = len(embs[0])
word_to_ix = {w: i for (i, w) in enumerate(words)}
in_vocab = 0
for word, emb in zip(words, embs):
if word == POLYGLOT_UNK or word == W2V_UNK: continue
if word in vocab:
in_vocab += 1
training_instances.append(Instance(charseq(word, c2i), emb))
training_char_count = len(c2i)
print "Total in Embeddings vocabulary:", len(words)
print "Training set character count: ", training_char_count
# Test: Will be doing inference on these words using mimick
if len(vocab) > 0:
total = len(vocab)
for v in vocab:
if v not in words:
test_instances.append(
Instance(charseq(v, c2i), np.array([0.0] * dim)))
print "Total Number of output words:", total
print "Total in Training Vocabulary:", in_vocab
print "Percentage in-vocab:", in_vocab / total
print "Total character count: ", len(c2i)
training_instances = dataset["training_instances"]
test_instances = dataset["test_instances"]
populate_test_insts_from_vocab = len(test_instances) == 0
emb_dim = len(training_instances[0].word_emb)
# Load words to write
vocab_words = {}
if populate_test_insts_from_vocab:
train_words = [wordify(w, i2c) for w in training_instances]
with codecs.open(options.vocab, "r", "utf-8") as vocab_file:
for vw in vocab_file.readlines():
vw = vw.strip()
vocab_words[vw] = np.zeros(emb_dim)
if populate_test_insts_from_vocab and vw not in train_words:
test_instances.append(
Instance(charseq(vw, c2i), np.zeros(emb_dim)))
if populate_test_insts_from_vocab:
# need to update i2c if saw new characters
i2c = {i: c for c, i in c2i.items()}
if not options.cnn:
model = LSTMMimick(c2i, options.num_lstm_layers, options.char_dim,
options.hidden_dim, emb_dim)
else:
model = CNNMimick(c2i, options.num_conv_layers, options.char_dim, options.hidden_dim,\
options.window_width, options.pooling_maxk, options.w_stride, emb_dim)
trainer = dy.MomentumSGDTrainer(model.model, options.learning_rate, 0.9)
root_logger.info("Training Algorithm: {}".format(type(trainer)))
test_instances = dataset["test_instances"]
populate_test_insts_from_vocab = len(test_instances) == 0
emb_dim = len(training_instances[0].word_emb)
# Load words to write
vocab_words = {}
if populate_test_insts_from_vocab:
train_words = [wordify(w, i2c) for w in training_instances]
for filename in options.vocab:
with codecs.open(filename, "r", "utf-8") as vocab_file:
for vw in vocab_file.readlines():
vw = vw.strip()
if vw in vocab_words: continue
vocab_words[vw] = np.zeros(emb_dim)
if populate_test_insts_from_vocab and vw not in train_words:
test_instances.append(Instance(charseq(vw, c2i), np.zeros(emb_dim)))
if populate_test_insts_from_vocab:
# need to update i2c if saw new characters
i2c = { i: c for c, i in c2i.items() }
if not options.cnn:
model = LSTMMimick(c2i, options.num_lstm_layers, options.char_dim, options.hidden_dim, emb_dim)
else:
model = CNNMimick(c2i, options.num_conv_layers, options.char_dim, options.hidden_dim,\
options.window_width, options.pooling_maxk, options.w_stride, emb_dim)
trainer = dy.MomentumSGDTrainer(model.model, options.learning_rate, 0.9)
root_logger.info("Training Algorithm: {}".format(type(trainer)))
root_logger.info("Number training instances: {}".format(len(training_instances)))
