def query_model(sess,
input_node,
predictions,
vocab,
rev_vocab,
max_seq_len,
dropout_keep_prob,
saver=None,
embs=None,
out_form="cosine"):
while True:
sys.stdout.write("Type a definition: ")
sys.stdout.flush()
sentence = sys.stdin.readline()
sys.stdout.write("Number of candidates: ")
sys.stdout.flush()
top = int(sys.stdin.readline())
# Get token-ids for the input gloss.
token_ids = data_utils.sentence_to_token_ids(sentence, vocab)
# Pad out (or truncate) the input gloss ids.
padded_ids = np.asarray(data_utils.pad_sequence(
token_ids, max_seq_len))
input_data = np.asarray([padded_ids])
# Single vector encoding the input gloss.
model_preds = sess.run(predictions,
feed_dict={
input_node: input_data,
dropout_keep_prob: 1.0
})
# Softmax already provides scores over the vocab.
if out_form == "softmax":
# Exclude padding and _UNK tokens from the top-k calculation.
candidate_ids = np.squeeze(
model_preds)[2:].argsort()[-top:][::-1] + 2
# Replace top-k ids with corresponding words.
candidates = [rev_vocab[idx] for idx in candidate_ids]
# Cosine requires sim to be calculated for each vocab word.
else:
sims = 1 - np.squeeze(
dist.cdist(model_preds, embs, metric="cosine"))
# replace nans with 0s.
sims = np.nan_to_num(sims)
candidate_ids = sims.argsort()[::-1][:top]
candidates = [rev_vocab[idx] for idx in candidate_ids]
# get baseline candidates from the raw embedding space.
base_rep = np.asarray([np.mean(embs[token_ids], axis=0)])
sims_base = 1 - np.squeeze(dist.cdist(base_rep, embs, metric="cosine"))
sims_base = np.nan_to_num(sims_base)
candidate_ids_base = sims_base.argsort()[::-1][:top]
candidates_base = [rev_vocab[idx] for idx in candidate_ids_base]
print("Top %s baseline candidates:" % top)
for ii, cand in enumerate(candidates_base):
print("%s: %s" % (ii + 1, cand))
print("\n Top %s candidates from the model:" % top)
for ii, cand in enumerate(candidates):
print("%s: %s" % (ii + 1, cand))
old_model_preds = model_preds
sys.stdout.flush()
sentence = sys.stdin.readline()
def get_feed_dict(self, sentences, labels=None, learning_rate=None, dropout=None):
sentences_ids, sequence_lengths = pad_sequence(sentences, 0)
feed = {
self.sentence_ids:sentences_ids,
self.sequence_lengths:sequence_lengths
}
if labels is not None:
labels, _ = pad_sequence(labels, 0)
feed[self.labels] = labels
if dropout is not None:
feed[self.dropout] = dropout
if learning_rate is not None:
feed[self.learning_rate] = learning_rate
return feed, sequence_lengths
def pad_and_bucket_dataset(in_dataset, in_vocabulary, in_config):
BUCKETS = in_config['buckets']
bucket_stats = collect_bucket_stats(in_dataset, BUCKETS)
bucketed_encoder_inputs = [
np.zeros((bucket_stats[bucket_id], input_length), dtype=np.uint32)
for bucket_id, (input_length, output_length) in enumerate(BUCKETS)
]
bucketed_decoder_inputs = [
np.zeros((bucket_stats[bucket_id], output_length), dtype=np.uint32)
for bucket_id, (input_length, output_length) in enumerate(BUCKETS)
]
bucket_cursors = [0 for _ in BUCKETS]
for row in in_dataset.itertuples():
encoder_input_ids, decoder_input_ids = row[1:]
bucket_id = find_bucket(
len(encoder_input_ids),
len(decoder_input_ids),
BUCKETS
)
if bucket_id is None:
continue
bucket_cursor = bucket_cursors[bucket_id]
input_length, output_length = BUCKETS[bucket_id]
padded_encoder_input = pad_sequence(
encoder_input_ids,
input_length,
padding='pre'
)
padded_decoder_input = pad_sequence(decoder_input_ids, output_length)
bucketed_encoder_inputs[bucket_id][bucket_cursor] = padded_encoder_input
bucketed_decoder_inputs[bucket_id][bucket_cursor] = padded_decoder_input
bucket_cursors[bucket_id] += 1
return {
'encoder': [
np.asarray(input_bucket)
for input_bucket in bucketed_encoder_inputs
],
'decoder': [
np.asarray(input_bucket)
for input_bucket in bucketed_decoder_inputs
],
}
def live_decode(in_vocabulary, in_embeddings, in_config):
logging.info('Loading the trained model')
model = create_model(in_vocabulary,
in_vocabulary,
in_embeddings,
in_config['buckets'][BUCKET_ID][0],
in_config['buckets'][BUCKET_ID][1],
in_config,
mode=Mode.TEST)
MODEL_FILE = in_config['model_weights']
model.load_weights(MODEL_FILE)
vocabulary_map = {
token: index
for index, token in enumerate(in_vocabulary)
}
print 'go'
while True:
user_input = stdin.readline().lower().strip()
if user_input == 'q':
break
token_ids = [vocabulary_map[token]
for token in user_input.split()] + [GO_ID]
BUCKETS = in_config['buckets']
bucket_id = find_bucket(len(token_ids), 0, BUCKETS)
decoder_inputs = pad_sequence(token_ids,
BUCKETS[bucket_id][0],
padding='pre')
decoder_input_matrix = np.asarray(decoder_inputs)
decoder_input_matrix = decoder_input_matrix.reshape(
[1] + list(decoder_input_matrix.shape))
decoder_outputs = model.predict(decoder_input_matrix)
argmaxes = np.argmax(decoder_outputs[0], axis=1)
decoded_ids = truncate_decoded_sequence(argmaxes)
print ' '.join(
[in_vocabulary[decoded_token] for decoded_token in decoded_ids])
