def get_q(self, context, ans, ans_pos):
# Process and create a batch of 1
ctxt_feats = preprocessing.process_squad_context(
self.model.vocab, context_as_set=FLAGS.context_as_set)(context)
ans_feats = preprocessing.process_squad_answer(
self.model.vocab,
context_as_set=FLAGS.context_as_set)(ans, ans_pos, context)
ctxt_feats[0] = np.array(ctxt_feats[0], dtype=bytes)
ans_feats[0] = np.array(ans_feats[0], dtype=bytes)
ctxt_dict = tuple([np.asarray([x]) for x in ctxt_feats])
ans_dict = tuple([np.asarray([x]) for x in ans_feats])
q, q_len = self.sess.run(
[self.model.q_hat_beam_string, self.model.q_hat_beam_lens],
feed_dict={
self.model.context_in: ctxt_dict,
self.model.answer_in: ans_dict
})
q_str = " ".join([
w.decode().replace('>', '>').replace('<', '<')
for w in q[0][:q_len[0] - 1]
])
return q_str
def get_q_batch(self, contexts, answers, ans_pos):
# Process and create a batch
ctxt_feats = [[x for x in preprocessing.process_squad_context(self.model.vocab, context_as_set=FLAGS.context_as_set)(contexts[i])] for i in range(len(contexts))]
ans_feats = [[x for x in preprocessing.process_squad_answer(self.model.vocab, context_as_set=FLAGS.context_as_set)(answers[i], ans_pos[i], contexts[i])] for i in range(len(contexts))]
# Now zip to get batches of features, not batches of examples
ctxt_feats = list(zip(*ctxt_feats))
ans_feats = list(zip(*ans_feats))
# pad
ctxt_pad = ['<PAD>', 0, 0, 0, 0]
ans_pad = ['<PAD>', 0, 0, 0]
for i in range(len(ctxt_feats)):
if i in [3,4]: # skip length feature
continue
max_len = max(len(feat) for feat in ctxt_feats[i])
ctxt_feats[i] = [list(feat) + [ctxt_pad[i] for j in range(max_len - len(feat))] for feat in ctxt_feats[i]]
for i in range(len(ans_feats)):
if i in [2]: # skip length feature
continue
max_len = max(len(feat) for feat in ans_feats[i])
ans_feats[i] = [list(feat) + [ans_pad[i] for j in range(max_len - len(feat))] for feat in ans_feats[i]]
# Needed to handle weird unicode stuff properly
ctxt_feats[0] = np.array(ctxt_feats[0], dtype=bytes)
ans_feats[0] = np.array(ans_feats[0], dtype=bytes)
ctxt_dict = tuple([np.array(x) for i, x in enumerate(ctxt_feats)])
ans_dict = tuple([np.array(x) for i, x in enumerate(ans_feats)])
# ctxt_dict = ctxt_feats
# ans_dict = ans_feats
qs, q_lens = self.sess.run([self.model.q_hat_beam_string,self.model.q_hat_beam_lens], feed_dict={self.model.context_in: ctxt_dict, self.model.answer_in: ans_dict})
q_str = [" ".join([w.decode().replace('>','>').replace('<','<') for w in qs[i][:q_lens[i]-1]]) for i in range(len(qs))]
return q_str
def build_data_pipeline(self, batch_size):
with tf.device('/cpu:*'):
self.context_ph = tf.placeholder(tf.string, [None])
self.qs_ph = tf.placeholder(tf.string, [None])
self.as_ph = tf.placeholder(tf.string, [None])
self.a_pos_ph = tf.placeholder(tf.int32, [None])
self.ix = tf.placeholder(tf.int32, [None])
dataset = tf.data.Dataset.from_tensor_slices(
(self.context_ph, self.qs_ph, self.as_ph, self.a_pos_ph,
self.ix))
if self.shuffle:
dataset = dataset.shuffle(buffer_size=100000)
# processing pipeline
dataset = dataset.map(lambda context, q, a, a_pos, ix: (
tuple(
tf.py_func(
preprocessing.process_squad_context(
self.vocab, context_as_set=FLAGS.context_as_set),
[context],
[tf.string, tf.int32, tf.int32, tf.int32, tf.int32])),
tuple(
tf.py_func(
preprocessing.process_squad_question(
self.vocab,
max_copy_size=FLAGS.max_copy_size,
context_as_set=FLAGS.context_as_set,
copy_priority=FLAGS.copy_priority,
smart_copy=FLAGS.smart_copy,
latent_switch=FLAGS.latent_switch), [
q, context, a_pos
], [tf.string, tf.int32, tf.float32, tf.int32])),
tuple(
tf.py_func(
preprocessing.process_squad_answer(
self.vocab, context_as_set=FLAGS.context_as_set), [
a, a_pos, context
], [tf.string, tf.int32, tf.int32, tf.int32])), ix
# q,a
))
# pad out to batches
dataset = dataset.padded_batch(
batch_size,
padded_shapes=(
(
tf.TensorShape([None
]), # source vectors of unknown size
tf.TensorShape([None
]), # source vectors of unknown size
tf.TensorShape([None]),
tf.TensorShape([]), # size(source)
tf.TensorShape([])), # size(source vocab)
(
tf.TensorShape([None
]), # target vectors of unknown size
tf.TensorShape([None
]), # target vectors of unknown size
tf.TensorShape([None, None
]), # target vectors of unknown size
tf.TensorShape([])), # size(source)
(
tf.TensorShape([None
]), # target vectors of unknown size
tf.TensorShape([None
]), # target vectors of unknown size
tf.TensorShape([]),
tf.TensorShape([None])), # size(target)
tf.TensorShape([])), #ix
padding_values=(
(
PAD,
self.vocab[
PAD], # source vectors padded on the right with src_eos_id
0,
len(self.vocab),
0), # size(source) -- unused
(
PAD,
self.vocab[
PAD], # target vectors padded on the right with tgt_eos_id
0.0,
0), # size(source) -- unused
(
PAD,
self.vocab[
PAD], # target vectors padded on the right with tgt_eos_id
0, # answer len
0), # answer locs
0)) # ix
dataset = dataset.repeat(self.num_epochs)
dataset = dataset.prefetch(buffer_size=batch_size * 4)
self.iterator = dataset.make_initializable_iterator()
self.batch_as_nested_tuple = self.iterator.get_next()
self.this_context, self.this_question, self.this_answer, self.this_ix = self.batch_as_nested_tuple
(self.context_raw, self.context_ids, self.context_copy_ids,
self.context_length, self.context_vocab_size) = self.this_context
(self.question_raw, self.question_ids, self.question_oh,
self.question_length) = self.this_question
(self.answer_raw, self.answer_ids, self.answer_length,
self.answer_locs) = self.this_answer
self.batch_len = tf.shape(self.context_raw)[0]
