def test_stacked_bilstm_compatibility(self):
checkpoint_dir = tempfile.mkdtemp(prefix="checkpoint_dir")
checkpoint_path = os.path.join(checkpoint_dir, "model.ckpt")
hidden_size = 10
num_layers = 3
dropout_ratio = 0.0
input_emb = np.random.uniform(size=[3, 5, 9]).astype(np.float32)
input_len = [4, 5, 2]
# Make sure we fail explicitly if the specified devices can't be used.
config = tf.ConfigProto(allow_soft_placement=False,
log_device_placement=True)
with tf.Graph().as_default():
with tf.device("/gpu:0"):
output_emb = cudnn_layers.stacked_bilstm(
input_emb=input_emb,
input_len=input_len,
hidden_size=hidden_size,
num_layers=num_layers,
dropout_ratio=dropout_ratio,
mode=tf.estimator.ModeKeys.TRAIN,
use_cudnn=True)
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
gpu_output_emb = sess.run(output_emb)
saver.save(sess, checkpoint_path)
with tf.Graph().as_default():
with tf.device("/cpu:0"):
output_emb = cudnn_layers.stacked_bilstm(
input_emb=input_emb,
input_len=input_len,
hidden_size=hidden_size,
num_layers=num_layers,
dropout_ratio=dropout_ratio,
mode=tf.estimator.ModeKeys.TRAIN,
use_cudnn=False)
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
saver.restore(sess, checkpoint_path)
cpu_output_emb = sess.run(output_emb)
for c, g, l in zip(cpu_output_emb, gpu_output_emb, input_len):
self.assertAllClose(c[:l], g[:l])
def test_stacked_bilstm(self):
with tf.Graph().as_default():
input_emb = tf.random_uniform([3, 5, 8])
input_len = tf.constant([4, 5, 2])
output_emb = cudnn_layers.stacked_bilstm(
input_emb=input_emb,
input_len=input_len,
hidden_size=10,
num_layers=3,
dropout_ratio=0.2,
mode=tf.estimator.ModeKeys.TRAIN)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
actual_output_emb = sess.run(output_emb)
self.assertAllEqual(actual_output_emb.shape, [3, 5, 10 * 2])
def apply_lstm(x, seq_len):
"""Run a bi-directional LSTM over the `x`.
Args:
x: <tf.float32>[batch, seq_len, dim]
seq_len: <tf.int32>[batch] for None, sequence lengths of `seq2`
Returns:
out, <tf.float32>[batch, seq_len, out_dim]
"""
return cudnn_layers.stacked_bilstm(input_emb=x,
input_len=seq_len,
hidden_size=FLAGS.lstm_dim,
num_layers=1,
dropout_ratio=0.0,
mode=tf_estimator.ModeKeys.TRAIN,
use_cudnn=None)
def score_endpoints(question_emb,
question_len,
context_emb,
context_len,
hidden_size,
num_layers,
dropout_ratio,
mode,
use_cudnn=None):
"""Compute two scores over context words based on the input embeddings.
Args:
question_emb: <float32> [batch_size, max_question_len, hidden_size]
question_len: <int32> [batch_size]
context_emb: <float32>[batch_size, max_context_len, hidden_size]
context_len: <int32> [batch_size]
hidden_size: Size of hidden layers.
num_layers: Number of LSTM layers.
dropout_ratio: The probability of dropping out hidden units.
mode: Object of type tf.estimator.ModeKeys.
use_cudnn: Specify the use of cudnn. `None` denotes automatic selection.
Returns:
start_scores: <float32> [batch_size, max_context_words]
end_scores: <float32> [batch_size, max_context_words]
"""
# [batch_size, max_question_len]
question_mask = tf.sequence_mask(question_len,
tensor_utils.shape(question_emb, 1),
dtype=tf.float32)
# [batch_size, max_context_len, hidden_size]
attended_emb = _attend_to_question(context_emb=context_emb,
question_emb=question_emb,
question_mask=question_mask,
hidden_size=hidden_size)
# [batch_size, max_context_len, hidden_size * 2]
context_emb = tf.concat([context_emb, attended_emb], -1)
with tf.variable_scope("contextualize_context"):
# [batch_size, max_context_len, hidden_size]
contextualized_context_emb = cudnn_layers.stacked_bilstm(
input_emb=context_emb,
input_len=context_len,
hidden_size=hidden_size,
num_layers=num_layers,
dropout_ratio=dropout_ratio,
mode=mode,
use_cudnn=use_cudnn)
with tf.variable_scope("contextualize_question"):
# [batch_size, max_question_len, hidden_size]
contextualized_question_emb = cudnn_layers.stacked_bilstm(
input_emb=question_emb,
input_len=question_len,
hidden_size=hidden_size,
num_layers=num_layers,
dropout_ratio=dropout_ratio,
mode=mode,
use_cudnn=use_cudnn)
if mode == tf_estimator.ModeKeys.TRAIN:
contextualized_context_emb = tf.nn.dropout(contextualized_context_emb,
1.0 - dropout_ratio)
contextualized_question_emb = tf.nn.dropout(
contextualized_question_emb, 1.0 - dropout_ratio)
# [batch_size, hidden_size]
pooled_question_emb = _attention_pool(contextualized_question_emb,
question_mask)
if mode == tf_estimator.ModeKeys.TRAIN:
pooled_question_emb = tf.nn.dropout(pooled_question_emb,
1.0 - dropout_ratio)
# [batch_size, max_context_len]
with tf.variable_scope("start_scores"):
start_scores = _bilinear_score(contextualized_context_emb,
pooled_question_emb)
with tf.variable_scope("end_scores"):
end_scores = _bilinear_score(contextualized_context_emb,
pooled_question_emb)
context_log_mask = tf.log(
tf.sequence_mask(context_len,
tensor_utils.shape(context_emb, 1),
dtype=tf.float32))
start_scores += context_log_mask
end_scores += context_log_mask
return start_scores, end_scores
