def test_read_word_embeddings_with_special_tokens(embedding_file):
file_name, gold_embeddings = embedding_file
cm_vocab, cm_embeddings = vocab.read_word_embeddings(file_name, EMBED_DIM)
assert cm_vocab == (vocab.SPECIAL_TOKENS + VOCAB)
for i, embed in enumerate(cm_embeddings[len(vocab.SPECIAL_TOKENS):, :]):
assert list(gold_embeddings[i]) == list(embed)
def test_read_word_embeddings_without_special_tokens(embedding_file):
file_name, gold_embeddings = embedding_file
cm_vocab, cm_embeddings = vocab.read_word_embeddings(
file_name, EMBED_DIM, include_special_tokens=False)
assert cm_vocab == (VOCAB)
for i, embed in enumerate(cm_embeddings):
assert list(gold_embeddings[i]) == list(embed)
def test_read_word_embeddings_with_special_tokens_custom_vocab_size(
embedding_file):
file_name, gold_embeddings = embedding_file
vocab_size = len(VOCAB) // 2
cm_vocab, cm_embeddings = vocab.read_word_embeddings(
file_name, EMBED_DIM, vocab_size)
assert cm_vocab == (vocab.SPECIAL_TOKENS + VOCAB[:vocab_size])
for i, embed in enumerate(cm_embeddings[len(vocab.SPECIAL_TOKENS):, :]):
assert list(gold_embeddings[i]) == list(embed)
def test_read_word_embeddings_without_special_tokens_custom_vocab_size(
embedding_file):
file_name, gold_embeddings = embedding_file
vocab_size = len(VOCAB) // 2
cm_vocab, cm_embeddings = vocab.read_word_embeddings(
file_name, EMBED_DIM, vocab_size, include_special_tokens=False)
assert cm_vocab == (VOCAB[:vocab_size])
for i, embed in enumerate(cm_embeddings):
assert list(gold_embeddings[i]) == list(embed)
def test_lookup_ops(embedding_file):
fn, gold_embeds = embedding_file
cm_vocab, _ = vocab.read_word_embeddings(fn, EMBED_DIM)
oov_words = ['dsjii', 'disjfi']
test_strings = vocab.SPECIAL_TOKENS + VOCAB + oov_words
tf.enable_eager_execution()
test_strings = tf.constant(test_strings)
table = vocab.get_vocab_lookup(cm_vocab)
ids = table.lookup(test_strings)
assert list(ids.numpy()) == list(range(
len(vocab.SPECIAL_TOKENS +
VOCAB))) + [vocab.OOV_TOKEN_ID] * len(oov_words)
def eval(config, data_dir, checkpoint_path=None, my_model_fn=model_fn):
V, embed_matrix = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size
)
estimator = get_estimator(config, embed_matrix, my_model_fn)
output = estimator.evaluate(
input_fn=lambda: eval_input_fn(config, data_dir, vocab.create_vocab_lookup_tables(V), num_examples=1e10),
checkpoint_path=checkpoint_path
)
return output
def augment_debug(config, debug_dataset, data_dir, checkpoint_path=None, my_model_fn=model_fn):
V, embed_matrix = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size
)
config.put('optim.batch_size', 1)
estimator = get_estimator(config, embed_matrix, my_model_fn)
with open(str(data_dir / debug_dataset), 'rb') as f:
debug_examples = pickle.load(f)
debugged = augment_debug_dataset(debug_examples, estimator, checkpoint_path, V)
with open("%s_debugged" % debug_dataset, 'w', encoding='utf8') as f:
json.dump(debugged, f)
def test_rnn_encoder(dataset_file, embedding_file):
with tf.Graph().as_default():
d_fn, gold_dataset = dataset_file
e_fn, gold_embeds = embedding_file
v, embed_matrix = vocab.read_word_embeddings(e_fn, EMBED_DIM)
vocab_lookup = vocab.get_vocab_lookup(v)
dataset = neural_editor.input_fn(d_fn, vocab_lookup, BATCH_SIZE,
NUM_EPOCH)
embedding = tf.get_variable(
'embeddings',
shape=embed_matrix.shape,
initializer=tf.constant_initializer(embed_matrix))
iter = dataset.make_initializable_iterator()
(src, tgt, iw, dw), _ = iter.get_next()
EDIT_DIM = 8
output = ev.rnn_encoder(tf.nn.embedding_lookup(embedding, src),
tf.nn.embedding_lookup(embedding, tgt),
tf.nn.embedding_lookup(embedding, iw),
tf.nn.embedding_lookup(embedding, dw),
sequence.length_pre_embedding(src),
sequence.length_pre_embedding(tgt),
sequence.length_pre_embedding(iw),
sequence.length_pre_embedding(dw), 256, 2, 256,
1, EDIT_DIM, 100.0, 0.1, 14.0, 0.8)
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer(),
tf.tables_initializer()
])
sess.run(iter.initializer)
while True:
try:
oeo = sess.run(output)
assert oeo.shape == (BATCH_SIZE, EDIT_DIM)
except:
break
def augment_meta_test(config, meta_test_path, data_dir, checkpoint_path=None, my_model_fn=model_fn):
V, embed_matrix = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size
)
config.put('optim.batch_size', 1)
estimator = get_estimator(config, embed_matrix, my_model_fn)
with open(str(data_dir / config.dataset.path / 'train.tsv'), encoding='utf8') as f:
train_examples = []
for l in tqdm(f, total=util.get_num_total_lines(data_dir / config.dataset.path / 'train.tsv')):
l = l[:-1]
src, tgt = l.split('\t')
train_examples.append((l, util.jaccard(
set([w.lower() for w in src.split(' ')]),
set([w.lower() for w in tgt.split(' ')]),
)))
train_examples = list(filter(lambda x: 0.6 < x[1] < 0.8, train_examples))
with open(meta_test_path, 'rb') as f:
meta_test = pickle.load(f)
# for i, m in enumerate(tqdm(meta_test)):
# dtrain = augment_dataset(train_examples, estimator, checkpoint_path, meta_test[0], V)
# assert len(dtrain) == len(meta_test[0]) * (1 + NUM_CANDIDATES)
# print("DTRRAAAAAAAAAAAAAIIIIIIIIIIIINNNNNNNN:", len(dtrain))
# meta_test[i] = (dtrain, meta_test[i][1], meta_test[i][2])
new_meta_test = []
all_dtrain = augment_dataset(train_examples, estimator, checkpoint_path, meta_test, V)
for episode_id in all_dtrain:
new_meta_test.append((
all_dtrain[episode_id],
meta_test[episode_id][1],
meta_test[episode_id][2],
))
with open(meta_test_path + '_augmented.pkl', 'wb') as f:
pickle.dump(new_meta_test, f)
def test_wtf():
with tf.Graph().as_default():
V, embed_matrix = vocab.read_word_embeddings(
Path('../data') / 'word_vectors' / 'glove.6B.300d_yelp.txt',
300,
10000
)
table = vocab.create_vocab_lookup_tables(V)
vocab_s2i = table[vocab.STR_TO_INT]
vocab_i2s = table[vocab.INT_TO_STR]
dataset = input_fn('../data/yelp_dataset_large_split/train.tsv', table, 64, 1)
iter = dataset.make_initializable_iterator()
(src, tgt, iw, dw), _ = iter.get_next()
src_len = length_pre_embedding(src)
tgt_len = length_pre_embedding(tgt)
iw_len = length_pre_embedding(iw)
dw_len = length_pre_embedding(dw)
dec_inputs = decoder.prepare_decoder_inputs(tgt, vocab.get_token_id(vocab.START_TOKEN, vocab_s2i))
dec_output = decoder.prepare_decoder_output(tgt, tgt_len, vocab.get_token_id(vocab.STOP_TOKEN, vocab_s2i),
vocab.get_token_id(vocab.PAD_TOKEN, vocab_s2i))
t_src = vocab_i2s.lookup(src)
t_tgt = vocab_i2s.lookup(tgt)
t_iw = vocab_i2s.lookup(iw)
t_dw = vocab_i2s.lookup(dw)
t_do = vocab_i2s.lookup(dec_output)
t_di = vocab_i2s.lookup(dec_inputs)
with tf.Session() as sess:
sess.run([tf.global_variables_initializer(), tf.local_variables_initializer(), tf.tables_initializer()])
sess.run(iter.initializer)
while True:
try:
# src, tgt, iw, dw = sess.run([src, tgt, iw, dw])
ts, tt, tiw, tdw, tdo, tdi = sess.run([t_src, t_tgt, t_iw, t_dw, t_do, t_di])
except:
break
def test_num_examples():
V, embed_matrix = vocab.read_word_embeddings(
'data/word_vectors/glove.6B.300d_dbpedia.txt',
300,
10000
)
vocab_table = vocab.create_vocab_lookup_tables(V)
dataset = input_fn('data/quora_naug/train.tsv', vocab_table, 10, 1)
for (features, _) in dataset:
base_words, extended_base_words, \
output_words, extended_output_words, \
src_words, tgt_words, \
inserted_words, deleted_words, \
oov = features
oov_len = sequence.length_string(oov, vocab.PAD_TOKEN)
print(features)
def predict(config, data_dir, checkpoint_path=None, my_model_fn=model_fn):
V, embed_matrix = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size
)
config.put('optim.batch_size', 1)
estimator = get_estimator(config, embed_matrix, my_model_fn)
output = estimator.predict(
input_fn=lambda: eval_input_fn(config, data_dir, vocab.create_vocab_lookup_tables(V), num_examples=10),
checkpoint_path=checkpoint_path
)
for p in output:
print(p['joined'])
return output
def predict_cmd(config, data_dir, checkpoint_path=None):
V, embed_matrix = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size
)
config.put('optim.batch_size', 1)
estimator = get_estimator(config, embed_matrix)
output = estimator.predict(
input_fn=lambda: input_fn_cmd(vocab.create_vocab_lookup_tables(V)),
checkpoint_path=checkpoint_path
)
for p in output:
print('\nResult:')
print(p['joined'])
return output
def get_vocab_embedding_matrix(config, data_dir):
if config.editor.get('use_sub_words', False):
if config.editor.get('use_t2t_sub_words'):
output = vocab.read_t2t_subword_embeddings(config)
else:
output = vocab.read_subword_embeddings(config)
else:
if config.editor.get('word_vocab_file_path', None):
vocab_path = str(config.local_data_dir / config.dataset.path / config.editor.word_vocab_file_path)
else:
vocab_path = None
output = vocab.read_word_embeddings(
data_dir / 'word_vectors' / config.editor.wvec_path,
config.editor.word_dim,
config.editor.vocab_size,
random_initialization=(not config.editor.get('use_pretrained_embeddings', True)),
vocab_file=vocab_path
)
return output
def test_context_encoder(dataset_file, embedding_file):
with tf.Graph().as_default():
d_fn, gold_dataset = dataset_file
e_fn, gold_embeds = embedding_file
v, embed_matrix = vocab.read_word_embeddings(e_fn, EMBED_DIM)
vocab_lookup = vocab.get_vocab_lookup(v)
dataset = neural_editor.input_fn(d_fn, vocab_lookup, BATCH_SIZE,
NUM_EPOCH)
embedding = tf.get_variable(
'embeddings',
shape=embed_matrix.shape,
initializer=tf.constant_initializer(embed_matrix))
iter = dataset.make_initializable_iterator()
(_, _, src, _), _ = iter.get_next()
src_len = sequence.length_pre_embedding(src)
src_embd = tf.nn.embedding_lookup(embedding, src)
output = ev.context_encoder(src_embd, src_len, HIDDEN_DIM, NUM_LAYER)
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer(),
tf.tables_initializer()
])
sess.run(iter.initializer)
while True:
try:
oeo, o_src, o_src_len, o_src_embd = sess.run(
[output, src, src_len, src_embd])
assert oeo.shape == (BATCH_SIZE, o_src_len.max(),
HIDDEN_DIM)
except:
break
def test_encoder(dataset_file, embedding_file):
d_fn, gold_dataset = dataset_file
e_fn, gold_embeds = embedding_file
v, embed_matrix = vocab.read_word_embeddings(e_fn, EMBED_DIM)
vocab_lookup = vocab.get_vocab_lookup(v)
dataset = neural_editor.input_fn(d_fn, vocab_lookup, BATCH_SIZE, NUM_EPOCH)
embedding = tf.get_variable(
'embeddings',
shape=embed_matrix.shape,
initializer=tf.constant_initializer(embed_matrix))
iter = dataset.make_initializable_iterator()
(src, _, _, _), _ = iter.get_next()
src_len = sequence.length_pre_embedding(src)
src_embd = tf.nn.embedding_lookup(embedding, src)
encoder_output, _ = encoder.bidirectional_encoder(src_embd, src_len,
HIDDEN_DIM, NUM_LAYER,
0.9)
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer(),
tf.tables_initializer()
])
sess.run(iter.initializer)
oeo, o_src, o_src_len, o_src_embd = sess.run(
[encoder_output, src, src_len, src_embd])
for i in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
print(i)
assert oeo.shape == (BATCH_SIZE, o_src_len.max(), HIDDEN_DIM)
def test_decoder_train(dataset_file, embedding_file):
with tf.Graph().as_default():
d_fn, gold_dataset = dataset_file
e_fn, gold_embeds = embedding_file
v, embed_matrix = vocab.read_word_embeddings(e_fn, EMBED_DIM)
vocab_lookup = vocab.get_vocab_lookup(v)
stop_token = tf.constant(bytes(vocab.STOP_TOKEN, encoding='utf8'),
dtype=tf.string)
stop_token_id = vocab_lookup.lookup(stop_token)
start_token = tf.constant(bytes(vocab.START_TOKEN, encoding='utf8'),
dtype=tf.string)
start_token_id = vocab_lookup.lookup(start_token)
pad_token = tf.constant(bytes(vocab.PAD_TOKEN, encoding='utf8'),
dtype=tf.string)
pad_token_id = vocab_lookup.lookup(pad_token)
dataset = neural_editor.input_fn(d_fn, vocab_lookup, BATCH_SIZE,
NUM_EPOCH)
iter = dataset.make_initializable_iterator()
(src, tgt, inw, dlw), _ = iter.get_next()
src_len = sequence.length_pre_embedding(src)
tgt_len = sequence.length_pre_embedding(tgt)
dec_inputs = decoder.prepare_decoder_inputs(tgt, start_token_id)
dec_outputs = decoder.prepare_decoder_output(tgt, tgt_len,
stop_token_id,
pad_token_id)
dec_inputs_len = sequence.length_pre_embedding(dec_inputs)
dec_outputs_len = sequence.length_pre_embedding(dec_outputs)
batch_size = tf.shape(src)[0]
edit_vector = edit_encoder.random_noise_encoder(
batch_size, EDIT_DIM, 14.0)
embedding = tf.get_variable(
'embeddings',
shape=embed_matrix.shape,
initializer=tf.constant_initializer(embed_matrix))
src_embd = tf.nn.embedding_lookup(embedding, src)
src_sent_embeds, final_states = encoder.source_sent_encoder(
src_embd, src_len, 20, 3, 0.8)
agn = agenda.linear(final_states, edit_vector, 4)
dec_out = decoder.train_decoder(agn, embedding, dec_inputs,
src_sent_embeds,
tf.nn.embedding_lookup(embedding, inw),
tf.nn.embedding_lookup(embedding, dlw),
dec_inputs_len, src_len,
sequence.length_pre_embedding(inw),
sequence.length_pre_embedding(dlw), 5,
20, 3, False)
# eval_dec_out = decoder.greedy_eval_decoder(
# agn, embedding,
# start_token_id, stop_token_id,
# src_sent_embeds,
# tf.nn.embedding_lookup(embedding, inw),
# tf.nn.embedding_lookup(embedding, dlw),
# src_len, sequence.length_pre_embedding(inw), sequence.length_pre_embedding(dlw),
# 5, 20, 3, 40
# )
eval_dec_out = decoder.beam_eval_decoder(
agn, embedding, start_token_id, stop_token_id, src_sent_embeds,
tf.nn.embedding_lookup(embedding, inw),
tf.nn.embedding_lookup(embedding, dlw), src_len,
sequence.length_pre_embedding(inw),
sequence.length_pre_embedding(dlw), 5, 20, 3, 40)
# saver = tf.train.Saver(write_version=tf.train.SaverDef.V1)
# s = tf.summary.FileWriter('data/an')
# s.add_graph(g)
#
# all_print = tf.get_collection('print')
an, final_states, len = dec_out
stacked = decoder.attention_score(dec_out)
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer(),
tf.tables_initializer()
])
sess.run(iter.initializer)
print(sess.run([eval_dec_out]))
def test_decoder_prepares(dataset_file, embedding_file):
with tf.Graph().as_default():
d_fn, gold_dataset = dataset_file
e_fn, gold_embeds = embedding_file
v, embed_matrix = vocab.read_word_embeddings(e_fn, EMBED_DIM)
vocab_lookup = vocab.get_vocab_lookup(v)
stop_token = tf.constant(bytes(vocab.STOP_TOKEN, encoding='utf8'),
dtype=tf.string)
stop_token_id = vocab_lookup.lookup(stop_token)
start_token = tf.constant(bytes(vocab.START_TOKEN, encoding='utf8'),
dtype=tf.string)
start_token_id = vocab_lookup.lookup(start_token)
pad_token = tf.constant(bytes(vocab.PAD_TOKEN, encoding='utf8'),
dtype=tf.string)
pad_token_id = vocab_lookup.lookup(pad_token)
dataset = neural_editor.input_fn(d_fn, vocab_lookup, BATCH_SIZE,
NUM_EPOCH)
iter = dataset.make_initializable_iterator()
(_, tgt, _, _), _ = iter.get_next()
tgt_len = sequence.length_pre_embedding(tgt)
dec_inputs = decoder.prepare_decoder_inputs(tgt, start_token_id)
dec_outputs = decoder.prepare_decoder_output(tgt, tgt_len,
stop_token_id,
pad_token_id)
dec_inputs_len = sequence.length_pre_embedding(dec_inputs)
dec_outputs_len = sequence.length_pre_embedding(dec_outputs)
dec_outputs_last = sequence.last_relevant(
tf.expand_dims(dec_outputs, 2), dec_outputs_len)
dec_outputs_last = tf.squeeze(dec_outputs_last)
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer(),
tf.tables_initializer()
])
sess.run(iter.initializer)
while True:
try:
dec_inputs, dec_outputs, tgt_len, dil, dol, start_token_id, stop_token_id, dec_outputs_last, tgt = sess.run(
[
dec_inputs, dec_outputs, tgt_len, dec_inputs_len,
dec_outputs_len, start_token_id, stop_token_id,
dec_outputs_last, tgt
])
assert list(dil) == list(dol) == list(tgt_len + 1)
assert list(dec_inputs[:, 0]) == list(
np.ones_like(dec_inputs[:, 0]) * start_token_id)
assert list(dec_outputs_last) == list(
np.ones_like(dec_outputs_last) * stop_token_id)
except:
break
