def _testDecoderTraining(self, decoder):
batch_size = 4
vocab_size = 10
time_dim = 5
depth = 6
inputs = tf.placeholder_with_default(np.random.randn(
batch_size, time_dim, depth).astype(np.float32),
shape=(None, None, depth))
# NOTE: max(sequence_length) may be less than time_dim when num_gpus > 1
sequence_length = [1, 3, 4, 2]
memory_sequence_length = [3, 7, 5, 4]
memory_time = max(memory_sequence_length)
memory = tf.placeholder_with_default(np.random.randn(
batch_size, memory_time, depth).astype(np.float32),
shape=(None, None, depth))
outputs, _, _ = decoder.decode(
inputs,
sequence_length,
vocab_size=vocab_size,
memory=memory,
memory_sequence_length=memory_sequence_length)
output_time_dim = tf.shape(outputs)[1]
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
with self.test_session() as sess:
output_time_dim_val = sess.run(output_time_dim)
self.assertEqual(time_dim, output_time_dim_val)
def _testDecoderTraining(self,
decoder,
initial_state_fn=None,
num_sources=1,
dtype=tf.float32):
batch_size = 4
vocab_size = 10
time_dim = 5
depth = 6
inputs = tf.placeholder_with_default(np.random.randn(
batch_size, time_dim, depth).astype(dtype.as_numpy_dtype()),
shape=(None, None, depth))
# NOTE: max(sequence_length) may be less than time_dim when num_gpus > 1
sequence_length = [1, 3, 4, 2]
initial_state, memory, memory_sequence_length = _generate_source_context(
batch_size,
depth,
initial_state_fn=initial_state_fn,
num_sources=num_sources,
dtype=dtype)
outputs, _, _, attention = decoder.decode(
inputs,
sequence_length,
vocab_size=vocab_size,
initial_state=initial_state,
memory=memory,
memory_sequence_length=memory_sequence_length,
return_alignment_history=True)
self.assertEqual(outputs.dtype, dtype)
output_time_dim = tf.shape(outputs)[1]
if decoder.support_alignment_history and num_sources == 1:
self.assertIsNotNone(attention)
else:
self.assertIsNone(attention)
saver = tf.train.Saver(var_list=tf.global_variables())
with self.test_session(graph=tf.get_default_graph()) as sess:
sess.run(tf.global_variables_initializer())
output_time_dim_val = sess.run(output_time_dim)
self.assertEqual(time_dim, output_time_dim_val)
if decoder.support_alignment_history and num_sources == 1:
attention_val, memory_time = sess.run(
[attention, tf.shape(memory)[1]])
self.assertAllEqual([batch_size, time_dim, memory_time],
attention_val.shape)
return saver.save(sess,
os.path.join(self.get_temp_dir(), "model.ckpt"))
def _testDecoderTraining(self,
decoder,
support_alignment_history=False,
dtype=tf.float32):
batch_size = 4
vocab_size = 10
time_dim = 5
depth = 6
inputs = tf.placeholder_with_default(np.random.randn(
batch_size, time_dim, depth).astype(dtype.as_numpy_dtype()),
shape=(None, None, depth))
# NOTE: max(sequence_length) may be less than time_dim when num_gpus > 1
sequence_length = [1, 3, 4, 2]
memory_sequence_length = [3, 7, 5, 4]
memory_time = max(memory_sequence_length)
memory = tf.placeholder_with_default(np.random.randn(
batch_size, memory_time, depth).astype(dtype.as_numpy_dtype()),
shape=(None, None, depth))
outputs, _, _, attention = decoder.decode(
inputs,
sequence_length,
vocab_size=vocab_size,
memory=memory,
memory_sequence_length=memory_sequence_length,
return_alignment_history=True)
self.assertEqual(outputs.dtype, dtype)
output_time_dim = tf.shape(outputs)[1]
if support_alignment_history:
self.assertIsNotNone(attention)
else:
self.assertIsNone(attention)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
with self.test_session() as sess:
output_time_dim_val = sess.run(output_time_dim)
self.assertEqual(time_dim, output_time_dim_val)
if support_alignment_history:
attention_val = sess.run(attention)
self.assertAllEqual([batch_size, time_dim, memory_time],
attention_val.shape)