def universal_transformer_tiny():
hparams = transformer.transformer_tiny()
hparams = update_hparams_for_universal_transformer(hparams)
# anywhere that we get good scores for the tiny model, it's with "sepconv"
hparams.transformer_ffn_type = "sepconv"
hparams.num_rec_steps = 8
return hparams
def transformer_extra_tiny():
hparams = transformer.transformer_tiny()
hparams.num_hidden_layers = 1
hparams.hidden_size = 32
hparams.filter_size = 128
hparams.num_heads = 2
return hparams
def testSlowVsFast(self):
model, features = get_model(transformer.transformer_tiny())
decode_length = 30
out_logits, _ = model(features)
out_logits = tf.squeeze(out_logits, axis=[2, 3])
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=tf.reshape(out_logits, [-1, VOCAB_SIZE]),
labels=tf.reshape(features["targets"], [-1]))
loss = tf.reduce_mean(loss)
apply_grad = tf.train.AdamOptimizer(0.001).minimize(loss)
with self.test_session():
tf.global_variables_initializer().run()
for _ in range(10):
apply_grad.run()
model.set_mode(tf.estimator.ModeKeys.PREDICT)
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
greedy_result = model._slow_greedy_infer(features,
decode_length)["outputs"]
greedy_result = tf.squeeze(greedy_result, axis=[2, 3])
fast_result = model._greedy_infer(features,
decode_length)["outputs"]
with self.test_session():
greedy_res = greedy_result.eval()
fast_res = fast_result.eval()
self.assertEqual(fast_res.shape,
(BATCH_SIZE, INPUT_LENGTH + decode_length))
self.assertAllClose(greedy_res, fast_res)
def get_model(hparams=None, mode=tf.estimator.ModeKeys.TRAIN,
has_input=True, model_cls=transformer.Transformer):
if hparams is None:
hparams = transformer.transformer_tiny()
hparams.hidden_size = 8
hparams.filter_size = 32
hparams.num_heads = 1
hparams.layer_prepostprocess_dropout = 0.0
p_hparams = problem_hparams.test_problem_hparams(VOCAB_SIZE,
VOCAB_SIZE,
hparams)
if not has_input:
del p_hparams.modality["inputs"]
hparams.problem_hparams = p_hparams
inputs = -1 + np.random.random_integers(
VOCAB_SIZE, size=(BATCH_SIZE, INPUT_LENGTH, 1, 1))
targets = -1 + np.random.random_integers(
VOCAB_SIZE, size=(BATCH_SIZE, TARGET_LENGTH, 1, 1))
features = {
"targets": tf.constant(targets, dtype=tf.int32, name="targets"),
"target_space_id": tf.constant(1, dtype=tf.int32)
}
if has_input:
features["inputs"] = tf.constant(inputs, dtype=tf.int32, name="inputs")
return model_cls(hparams, mode, p_hparams), features
def _create_greedy_infer_model(self):
"""Creates model for greedy inference testing.
Returns:
model: A t2t model.
features: An map of string to tensor.
"""
model, features = get_model(transformer.transformer_tiny())
out_logits, _ = model(features)
out_logits = tf.squeeze(out_logits, axis=[2, 3])
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=tf.reshape(out_logits, [-1, VOCAB_SIZE]),
labels=tf.reshape(features["targets"], [-1]))
loss = tf.reduce_mean(loss)
apply_grad = tf.train.AdamOptimizer(0.001).minimize(loss)
with self.test_session():
tf.global_variables_initializer().run()
for _ in range(10):
apply_grad.run()
model.set_mode(tf.estimator.ModeKeys.PREDICT)
return model, features
def m2m_m_transformer_hparams():
hparams = transformer_tiny()
hparams.num_hidden_layers = 4
hparams.hidden_size = 128
hparams.filter_size = 512
hparams.num_heads = 4
return hparams
def _testTransformer(self, net):
batch_size = 3
input_length = 5
target_length = 7
vocab_size = 9
hparams = transformer.transformer_tiny()
p_hparams = problem_hparams.test_problem_hparams(
hparams, vocab_size, vocab_size)
inputs = -1 + np.random.random_integers(
vocab_size, size=(batch_size, input_length, 1, 1))
targets = -1 + np.random.random_integers(
vocab_size, size=(batch_size, target_length, 1, 1))
with self.test_session() as session:
features = {
"inputs": tf.constant(inputs, dtype=tf.int32),
"targets": tf.constant(targets, dtype=tf.int32),
"target_space_id": tf.constant(1, dtype=tf.int32),
}
model = net(hparams, p_hparams)
shadred_logits, _, _ = model.model_fn(features, True)
logits = tf.concat(shadred_logits, 0)
session.run(tf.global_variables_initializer())
res = session.run(logits)
self.assertEqual(res.shape,
(batch_size, target_length, 1, 1, vocab_size))
def get_model(hparams=None, mode=tf.estimator.ModeKeys.TRAIN,
has_input=True, model_cls=transformer.Transformer):
if hparams is None:
hparams = transformer.transformer_tiny()
hparams.hidden_size = 8
hparams.filter_size = 32
hparams.num_heads = 1
hparams.layer_prepostprocess_dropout = 0.0
p_hparams = problem_hparams.test_problem_hparams(VOCAB_SIZE,
VOCAB_SIZE,
hparams)
if not has_input:
del p_hparams.modality["inputs"]
hparams.problem_hparams = p_hparams
inputs = -1 + np.random.random_integers(
VOCAB_SIZE, size=(BATCH_SIZE, INPUT_LENGTH, 1, 1))
targets = -1 + np.random.random_integers(
VOCAB_SIZE, size=(BATCH_SIZE, TARGET_LENGTH, 1, 1))
features = {
"targets": tf.constant(targets, dtype=tf.int32, name="targets"),
"target_space_id": tf.constant(1, dtype=tf.int32)
}
if has_input:
features["inputs"] = tf.constant(inputs, dtype=tf.int32, name="inputs")
return model_cls(hparams, mode, p_hparams), features
def transformer_extra_tiny_agg():
hparams = transformer.transformer_tiny()
hparams.batch_size = 8
hparams.num_hidden_layers = 1
hparams.hidden_size = 64
hparams.filter_size = 256
hparams.num_heads = 2
return hparams
def testEvolvedTransformer(self):
model, features = get_model(hparams=transformer.transformer_tiny())
logits, _ = model(features)
with self.test_session() as session:
session.run(tf.global_variables_initializer())
res = session.run(logits)
self.assertEqual(res.shape,
(BATCH_SIZE, TARGET_LENGTH, 1, 1, VOCAB_SIZE))
def dstc_transformer_hparams_v8():
hparams = transformer_tiny()
hparams.num_hidden_layers = 4
hparams.dropout = 0.7
hparams.hidden_size = 128
hparams.filter_size = 512
hparams.num_heads = 4
return hparams
def t_rel_len2048_dropout15_tiny(): """Hparams for LM with relative attention, tiny transformer.""" # hparams = transformer.transformer_base() hparams = transformer.transformer_tiny() update_transformer_hparams_for_music(hparams) update_truncate_length(hparams, 2048) update_dropout(hparams, 0.15) hparams.self_attention_type = "dot_product_relative_v2" # Need to specify num_hidden_layers hparams.attention_key_channels = 512 hparams.num_hidden_layers = 8 return hparams
def get_model():
hparams = transformer.transformer_tiny()
hparams.layer_prepostprocess_dropout = 0.0
p_hparams = problem_hparams.test_problem_hparams(VOCAB_SIZE, VOCAB_SIZE,
hparams)
hparams.problem_hparams = p_hparams
inputs = np.random.randint(VOCAB_SIZE, size=(BATCH_SIZE, INPUT_LENGTH, 1, 1))
targets = np.random.randint(
VOCAB_SIZE, size=(BATCH_SIZE, TARGET_LENGTH, 1, 1))
features = {
"targets": tf.constant(targets, dtype=tf.int32, name="targets"),
"target_space_id": tf.constant(1, dtype=tf.int32),
"inputs": tf.constant(inputs, dtype=tf.int32, name="inputs"),
}
return (evolved_transformer.EvolvedTransformer(
hparams, tf.estimator.ModeKeys.TRAIN, p_hparams), features)
def transformer_encoding(node_seq_input, num_nodes, params, mode):
"""Construct a node-level encoder based on the transformer module.
Args:
node_seq_input : tf.Tensor. A tensor with 3 dimensions.
num_nodes: tf.Tensor. Number of nodes per instance.
params : dict. A parameter dictionary.
mode : tf.estimator.ModeKeys object.
Returns:
node_seq_output: tf.Tensor. A tensor with 3 dimensions.
"""
node_weights = tf.sequence_mask(num_nodes)
hparams = transformer.transformer_tiny()
hparams.hidden_size = params["transformer_hidden_unit"]
hparams.num_heads = params["transformer_head"]
hparams.num_hidden_layers = params["transformer_hidden_layer"]
if hparams.hidden_size % hparams.num_heads != 0:
raise ValueError(
"The hidden_size needs to be divisible by trans_head.")
transformer_encoder = transformer.TransformerEncoder(hparams, mode=mode)
# Input shape [batch_size, sequence_length, 1, hidden_dim].
node_seq_input = tf.layers.dense(node_seq_input, hparams.hidden_size)
node_seq_input_reshape = tf.expand_dims(node_seq_input, 2)
# Targets and target_space_id are required by decoder of transformer,
# are both set as 0 for encoder.
node_seq_output = transformer_encoder(
{
"inputs": node_seq_input_reshape,
"targets": 0,
"target_space_id": 0,
},
nonpadding=node_weights)
node_seq_output = tf.squeeze(node_seq_output[0], 2)
# Construct a residue network by adding up the input and output
node_seq_output = tf.add(node_seq_input, node_seq_output)
return node_seq_output
def transformer_hparams(hidden_size):
"""Creates hyperpameters for autoregressive prior.
Args:
hidden_size: Width of attention layers and neural network output layer.
Returns:
hparams: Hyperpameters with basic presets for a Transformer.
"""
hparams = transformer.transformer_tiny()
hparams.add_hparam("shared_rel", False)
hparams.add_hparam("q_filter_width", 1)
hparams.add_hparam("kv_filter_width", 1)
hparams.hidden_size = hidden_size
hparams.num_layers = 6
hparams.layer_prepostprocess_dropout = 0.
hparams.attention_dropout = 0.
hparams.relu_dropout = 0.
hparams.block_length = 1
hparams.block_width = 1
hparams.ffn_layer = "conv_hidden_relu"
return hparams
def testBeamVsFast(self):
model, features = get_model(transformer.transformer_tiny())
decode_length = 30
out_logits, _ = model(features)
out_logits = tf.squeeze(out_logits, axis=[2, 3])
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=tf.reshape(out_logits, [-1, VOCAB_SIZE]),
labels=tf.reshape(features["targets"], [-1]))
loss = tf.reduce_mean(loss)
apply_grad = tf.train.AdamOptimizer(0.001).minimize(loss)
with self.test_session():
tf.global_variables_initializer().run()
for _ in range(10):
apply_grad.run()
model.set_mode(tf.estimator.ModeKeys.PREDICT)
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
beam_result = model._beam_decode_slow(features,
decode_length,
beam_size=4,
top_beams=1,
alpha=1.0)["outputs"]
fast_result = model._beam_decode(features,
decode_length,
beam_size=4,
top_beams=1,
alpha=1.0)["outputs"]
with self.test_session():
beam_res = beam_result.eval()
fast_res = fast_result.eval()
self.assertAllClose(beam_res, fast_res)
def _testTransformer(self, net):
batch_size = 3
input_length = 5
target_length = 7
vocab_size = 9
hparams = transformer.transformer_tiny()
p_hparams = problem_hparams.test_problem_hparams(hparams, vocab_size,
vocab_size)
inputs = -1 + np.random.random_integers(
vocab_size, size=(batch_size, input_length, 1, 1))
targets = -1 + np.random.random_integers(
vocab_size, size=(batch_size, target_length, 1, 1))
with self.test_session() as session:
features = {
"inputs": tf.constant(inputs, dtype=tf.int32),
"targets": tf.constant(targets, dtype=tf.int32),
"target_space_id": tf.constant(1, dtype=tf.int32),
}
model = net(hparams, p_hparams)
shadred_logits, _, _ = model.model_fn(features, True)
logits = tf.concat(shadred_logits, 0)
session.run(tf.global_variables_initializer())
res = session.run(logits)
self.assertEqual(res.shape, (batch_size, target_length, 1, 1, vocab_size))
def universal_transformer_tiny1(): hparams = transformer.transformer_tiny() hparams = update_hparams_for_universal_transformer(hparams) hparams.num_rec_steps = 8 return hparams
def transformer_aux_tiny():
"""Set of hyperparameters."""
hparams = transformer.transformer_tiny()
hparams.shared_embedding_and_softmax_weights = False
hparams.add_hparam("shift_values", "1,2")
return hparams
def transformer_tiny_bs3():
hparams = transformer.transformer_tiny()
hparams.add_hparam("block_size", 3)
return hparams
def r_transformer_tiny():
hparams = transformer.transformer_tiny()
hparams = update_hparams_for_r_transformer(hparams)
hparams.num_rec_steps = 8
return hparams
def evolved_transformer_tiny():
"""Base parameters for Evolved Transformer model."""
hparams = add_evolved_transformer_hparams(transformer.transformer_tiny())
hparams.learning_rate_schedule = ("constant*single_cycle_cos_decay")
return hparams
def transformer_aux_tiny():
"""Set of hyperparameters."""
hparams = transformer.transformer_tiny()
hparams.shared_embedding_and_softmax_weights = False
hparams.add_hparam("shift_values", "1,2")
return hparams
def universal_transformer_tiny(): hparams = transformer.transformer_tiny() hparams = update_hparams_for_universal_transformer(hparams) hparams.num_rec_steps = 8 return hparams
