def single_test(l, model, sess, task, nprint, batch_size, print_out=True,
offset=None):
"""Test model on test data of length l using the given session."""
inpt, target = data.get_batch(l, batch_size, False, task, offset)
_, res, _, steps = model.step(sess, inpt, target, False)
errors, total, seq_err = data.accuracy(inpt, res, target, batch_size, nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(" %s len %d errors %.2f sequence-errors %.2f"
% (task, l, 100*errors, 100*seq_err))
return errors, seq_err, (steps, inpt, [np.argmax(o, axis=1) for o in res])
def assign_vectors(word_vector_file, embedding_key, vocab_path, sess):
"""Assign the embedding_key variable from the given word vectors file."""
# For words in the word vector file, set their embedding at start.
if not tf.gfile.Exists(word_vector_file):
data.print_out("Word vector file does not exist: %s" % word_vector_file)
sys.exit(1)
vocab, _ = wmt.initialize_vocabulary(vocab_path)
vectors_variable = [v for v in tf.trainable_variables()
if embedding_key == v.name]
if len(vectors_variable) != 1:
data.print_out("Word vector variable not found or too many.")
sys.exit(1)
vectors_variable = vectors_variable[0]
vectors = vectors_variable.eval()
data.print_out("Pre-setting word vectors from %s" % word_vector_file)
with tf.gfile.GFile(word_vector_file, mode="r") as f:
# Lines have format: dog 0.045123 -0.61323 0.413667 ...
for line in f:
line_parts = line.split()
# The first part is the word.
word = line_parts[0]
if word in vocab:
# Remaining parts are components of the vector.
word_vector = np.array(map(float, line_parts[1:]))
if len(word_vector) != FLAGS.vec_size:
data.print_out("Warn: Word '%s', Expecting vector size %d, "
"found %d" % (word, FLAGS.vec_size,
len(word_vector)))
else:
vectors[vocab[word]] = word_vector
# Assign the modified vectors to the vectors_variable in the graph.
sess.run([vectors_variable.initializer],
{vectors_variable.initializer.inputs[1]: vectors})
def print_vectors(embedding_key, vocab_path, word_vector_file):
"""Print vectors from the given variable."""
_, rev_vocab = wmt.initialize_vocabulary(vocab_path)
vectors_variable = [v for v in tf.trainable_variables()
if embedding_key == v.name]
if len(vectors_variable) != 1:
data.print_out("Word vector variable not found or too many.")
sys.exit(1)
vectors_variable = vectors_variable[0]
vectors = vectors_variable.eval()
l, s = vectors.shape[0], vectors.shape[1]
data.print_out("Printing %d word vectors from %s to %s."
% (l, embedding_key, word_vector_file))
with tf.gfile.GFile(word_vector_file, mode="w") as f:
# Lines have format: dog 0.045123 -0.61323 0.413667 ...
for i in xrange(l):
f.write(rev_vocab[i])
for j in xrange(s):
f.write(" %.8f" % vectors[i][j])
f.write("\n")
def single_test(bin_id, model, sess, nprint, batch_size, dev, p, print_out=True,
offset=None, beam_model=None):
"""Test model on test data of length l using the given session."""
if not dev[p][bin_id]:
data.print_out(" bin %d (%d)\t%s\tppl NA errors NA seq-errors NA"
% (bin_id, data.bins[bin_id], p))
return 1.0, 1.0, 0.0
inpt, target = data.get_batch(
bin_id, batch_size, dev[p], FLAGS.height, offset)
if FLAGS.beam_size > 1 and beam_model:
loss, res, new_tgt, scores = m_step(
model, beam_model, sess, batch_size, inpt, target, bin_id,
FLAGS.eval_beam_steps, p)
score_avgs = [sum(s) / float(len(s)) for s in scores]
score_maxs = [max(s) for s in scores]
score_str = ["(%.2f, %.2f)" % (score_avgs[i], score_maxs[i])
for i in xrange(FLAGS.eval_beam_steps)]
data.print_out(" == scores (avg, max): %s" % "; ".join(score_str))
errors, total, seq_err = data.accuracy(inpt, res, target, batch_size,
nprint, new_tgt, scores[-1])
else:
loss, res, _, _ = model.step(sess, inpt, target, False)
errors, total, seq_err = data.accuracy(inpt, res, target, batch_size,
nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(" bin %d (%d)\t%s\tppl %.2f errors %.2f seq-errors %.2f"
% (bin_id, data.bins[bin_id], p, data.safe_exp(loss),
100 * errors, 100 * seq_err))
return (errors, seq_err, loss)
def multi_test(l,
model,
sess,
task,
nprint,
batch_size,
offset=None,
ensemble=None):
"""Run multiple tests at lower batch size to save memory."""
errors, seq_err = 0.0, 0.0
to_print = nprint
low_batch = FLAGS.low_batch_size
low_batch = min(low_batch, batch_size)
for mstep in xrange(batch_size / low_batch):
cur_offset = None if offset is None else offset + mstep * low_batch
err, sq_err, _ = single_test(l,
model,
sess,
task,
to_print,
low_batch,
False,
cur_offset,
ensemble=ensemble)
to_print = max(0, to_print - low_batch)
errors += err
seq_err += sq_err
if FLAGS.mode > 0:
cur_errors = float(low_batch * errors) / ((mstep + 1) * low_batch)
cur_seq_err = float(low_batch * seq_err) / (
(mstep + 1) * low_batch)
data.print_out(
" %s multitest current errors %.2f sequence-errors %.2f" %
(task, 100 * cur_errors, 100 * cur_seq_err))
errors = float(low_batch) * float(errors) / batch_size
seq_err = float(low_batch) * float(seq_err) / batch_size
data.print_out(" %s len %d errors %.2f sequence-errors %.2f" %
(task, l, 100 * errors, 100 * seq_err))
return errors, seq_err
def multi_test(l, model, sess, task, nprint, batch_size, offset=None):
"""Run multiple tests at lower batch size to save memory."""
errors, seq_err = 0.0, 0.0
to_print = nprint
low_batch = FLAGS.low_batch_size
low_batch = min(low_batch, batch_size)
for mstep in xrange(batch_size / low_batch):
cur_offset = None if offset is None else offset + mstep * low_batch
err, sq_err, _ = single_test(l, model, sess, task, to_print, low_batch,
False, cur_offset)
to_print = max(0, to_print - low_batch)
errors += err
seq_err += sq_err
if FLAGS.mode > 0:
cur_errors = float(low_batch * errors) / ((mstep+1) * low_batch)
cur_seq_err = float(low_batch * seq_err) / ((mstep+1) * low_batch)
data.print_out(" %s multitest current errors %.2f sequence-errors %.2f"
% (task, 100*cur_errors, 100*cur_seq_err))
errors = float(low_batch) * float(errors) / batch_size
seq_err = float(low_batch) * float(seq_err) / batch_size
data.print_out(" %s len %d errors %.2f sequence-errors %.2f"
% (task, l, 100*errors, 100*seq_err))
return errors, seq_err
def spec(self, inp, task, nclass):
"""Return the target given the input for some tasks."""
if task == "sort":
return sorted(inp)
elif task == "id":
return inp
elif task == "rev":
return [i for i in reversed(inp)]
elif task == "shuffle": # bit reverse permutation
n_bits = (len(inp) - 1).bit_length()
res = []
for i in range(len(inp)):
i1 = reverse_bit(i, n_bits) % len(inp)
res.append(inp[i1])
return res
elif task == "incr":
carry = 1
res = []
for i in range(len(inp)):
if inp[i] + carry < nclass:
res.append(inp[i] + carry)
carry = 0
else:
res.append(1)
carry = 1
return res
elif task == "left":
return [inp[0]]
elif task == "right":
return [inp[-1]]
elif task == "left-shift":
return [inp[l - 1] for l in range(len(inp))]
elif task == "right-shift":
return [inp[l + 1] for l in range(len(inp))]
else:
data_utils.print_out("Unknown spec for task " + str(task))
sys.exit()
def single_test(l,
model,
sess,
task,
nprint,
batch_size,
print_out=True,
offset=None,
ensemble=None,
get_steps=False):
"""Test model on test data of length l using the given session."""
inpt, target = data.get_batch(l, batch_size, False, task, offset)
_, res, _, steps = model.step(sess,
inpt,
target,
False,
get_steps=get_steps)
errors, total, seq_err = data.accuracy(inpt, res, target, batch_size,
nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(" %s len %d errors %.2f sequence-errors %.2f" %
(task, l, 100 * errors, 100 * seq_err))
# Ensemble eval.
if ensemble:
results = []
for m in ensemble:
model.saver.restore(sess, m)
_, result, _, _ = model.step(sess, inpt, target, False)
m_errors, m_total, m_seq_err = data.accuracy(
inpt, result, target, batch_size, nprint)
m_seq_err = float(m_seq_err) / batch_size
if total > 0:
m_errors = float(m_errors) / m_total
data.print_out(
" %s len %d m-errors %.2f m-sequence-errors %.2f" %
(task, l, 100 * m_errors, 100 * m_seq_err))
results.append(result)
ens = [sum(o) for o in zip(*results)]
errors, total, seq_err = data.accuracy(inpt, ens, target, batch_size,
nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(
" %s len %d ens-errors %.2f ens-sequence-errors %.2f" %
(task, l, 100 * errors, 100 * seq_err))
return errors, seq_err, (steps, inpt, [np.argmax(o, axis=1) for o in res])
def single_test(l, model, sess, task, nprint, batch_size, print_out=True,
offset=None, ensemble=None, get_steps=False):
"""Test model on test data of length l using the given session."""
inpt, target = data.get_batch(l, batch_size, False, task, offset)
_, res, _, steps = model.step(sess, inpt, target, False, get_steps=get_steps)
errors, total, seq_err = data.accuracy(inpt, res, target, batch_size, nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(" %s len %d errors %.2f sequence-errors %.2f"
% (task, l, 100*errors, 100*seq_err))
# Ensemble eval.
if ensemble:
results = []
for m in ensemble:
model.saver.restore(sess, m)
_, result, _, _ = model.step(sess, inpt, target, False)
m_errors, m_total, m_seq_err = data.accuracy(inpt, result, target,
batch_size, nprint)
m_seq_err = float(m_seq_err) / batch_size
if total > 0:
m_errors = float(m_errors) / m_total
data.print_out(" %s len %d m-errors %.2f m-sequence-errors %.2f"
% (task, l, 100*m_errors, 100*m_seq_err))
results.append(result)
ens = [sum(o) for o in zip(*results)]
errors, total, seq_err = data.accuracy(inpt, ens, target,
batch_size, nprint)
seq_err = float(seq_err) / batch_size
if total > 0:
errors = float(errors) / total
if print_out:
data.print_out(" %s len %d ens-errors %.2f ens-sequence-errors %.2f"
% (task, l, 100*errors, 100*seq_err))
return errors, seq_err, (steps, inpt, [np.argmax(o, axis=1) for o in res])
def train():
"""Train the model."""
batch_size = FLAGS.batch_size * FLAGS.num_gpus
(model, beam_model, min_length, max_length, checkpoint_dir,
(train_set, dev_set, en_vocab_path, fr_vocab_path), sv, sess) = initialize()
with sess.as_default():
quant_op = model.quantize_op
max_cur_length = min(min_length + 3, max_length)
prev_acc_perp = [1000000 for _ in xrange(5)]
prev_seq_err = 1.0
is_chief = FLAGS.task < 1
do_report = False
# Main traning loop.
while not sv.ShouldStop():
global_step, max_cur_length, learning_rate = sess.run(
[model.global_step, model.cur_length, model.lr])
acc_loss, acc_l1, acc_total, acc_errors, acc_seq_err = 0.0, 0.0, 0, 0, 0
acc_grad_norm, step_count, step_c1, step_time = 0.0, 0, 0, 0.0
# For words in the word vector file, set their embedding at start.
bound1 = FLAGS.steps_per_checkpoint - 1
if FLAGS.word_vector_file_en and global_step < bound1 and is_chief:
assign_vectors(FLAGS.word_vector_file_en, "embedding:0",
en_vocab_path, sess)
if FLAGS.max_target_vocab < 1:
assign_vectors(FLAGS.word_vector_file_en, "target_embedding:0",
en_vocab_path, sess)
if FLAGS.word_vector_file_fr and global_step < bound1 and is_chief:
assign_vectors(FLAGS.word_vector_file_fr, "embedding:0",
fr_vocab_path, sess)
if FLAGS.max_target_vocab < 1:
assign_vectors(FLAGS.word_vector_file_fr, "target_embedding:0",
fr_vocab_path, sess)
for _ in xrange(FLAGS.steps_per_checkpoint):
step_count += 1
step_c1 += 1
global_step = int(model.global_step.eval())
train_beam_anneal = global_step / float(FLAGS.train_beam_anneal)
train_beam_freq = FLAGS.train_beam_freq * min(1.0, train_beam_anneal)
p = random.choice(FLAGS.problem.split("-"))
train_set = global_train_set[p][-1]
bucket_id = get_bucket_id(train_buckets_scale[p][-1], max_cur_length,
train_set)
# Prefer longer stuff 60% of time if not wmt.
if np.random.randint(100) < 60 and FLAGS.problem != "wmt":
bucket1 = get_bucket_id(train_buckets_scale[p][-1], max_cur_length,
train_set)
bucket_id = max(bucket1, bucket_id)
# Run a step and time it.
start_time = time.time()
inp, target = data.get_batch(bucket_id, batch_size, train_set,
FLAGS.height)
noise_param = math.sqrt(math.pow(global_step + 1, -0.55) *
prev_seq_err) * FLAGS.grad_noise_scale
# In multi-step mode, we use best from beam for middle steps.
state, new_target, scores, history = None, None, None, []
while (FLAGS.beam_size > 1 and
train_beam_freq > np.random.random_sample()):
# Get the best beam (no training, just forward model).
new_target, new_first, new_inp, scores = get_best_beam(
beam_model, sess, inp, target,
batch_size, FLAGS.beam_size, bucket_id, history, p)
history.append(new_first)
# Training step with the previous input and the best beam as target.
_, _, _, state = model.step(sess, inp, new_target, FLAGS.do_train,
noise_param, update_mem=True, state=state)
# Change input to the new one for the next step.
inp = new_inp
# If all results are great, stop (todo: not to wait for all?).
if FLAGS.nprint > 1:
print(scores)
if sum(scores) / float(len(scores)) >= 10.0:
break
# The final step with the true target.
loss, res, gnorm, _ = model.step(
sess, inp, target, FLAGS.do_train, noise_param,
update_mem=True, state=state)
step_time += time.time() - start_time
acc_grad_norm += 0.0 if gnorm is None else float(gnorm)
# Accumulate statistics.
acc_loss += loss
acc_l1 += loss
errors, total, seq_err = data.accuracy(
inp, res, target, batch_size, 0, new_target, scores)
if FLAGS.nprint > 1:
print("seq_err: ", seq_err)
acc_total += total
acc_errors += errors
acc_seq_err += seq_err
# Report summary every 10 steps.
if step_count + 3 > FLAGS.steps_per_checkpoint:
do_report = True # Don't polute plot too early.
if is_chief and step_count % 10 == 1 and do_report:
cur_loss = acc_l1 / float(step_c1)
acc_l1, step_c1 = 0.0, 0
cur_perp = data.safe_exp(cur_loss)
summary = tf.Summary()
summary.value.extend(
[tf.Summary.Value(tag="log_perplexity", simple_value=cur_loss),
tf.Summary.Value(tag="perplexity", simple_value=cur_perp)])
sv.SummaryComputed(sess, summary, global_step)
# Normalize and print out accumulated statistics.
acc_loss /= step_count
step_time /= FLAGS.steps_per_checkpoint
acc_seq_err = float(acc_seq_err) / (step_count * batch_size)
prev_seq_err = max(0.0, acc_seq_err - 0.02) # No noise at error < 2%.
acc_errors = float(acc_errors) / acc_total if acc_total > 0 else 1.0
t_size = float(sum([len(x) for x in train_set])) / float(1000000)
msg = ("step %d step-time %.2f train-size %.3f lr %.6f grad-norm %.4f"
% (global_step + 1, step_time, t_size, learning_rate,
acc_grad_norm / FLAGS.steps_per_checkpoint))
data.print_out("%s len %d ppl %.6f errors %.2f sequence-errors %.2f" %
(msg, max_cur_length, data.safe_exp(acc_loss),
100*acc_errors, 100*acc_seq_err))
# If errors are below the curriculum threshold, move curriculum forward.
is_good = FLAGS.curriculum_ppx > data.safe_exp(acc_loss)
is_good = is_good and FLAGS.curriculum_seq > acc_seq_err
if is_good and is_chief:
if FLAGS.quantize:
# Quantize weights.
data.print_out(" Quantizing parameters.")
sess.run([quant_op])
# Increase current length (until the next with training data).
sess.run(model.cur_length_incr_op)
# Forget last perplexities if we're not yet at the end.
if max_cur_length < max_length:
prev_acc_perp.append(1000000)
# Lower learning rate if we're worse than the last 5 checkpoints.
acc_perp = data.safe_exp(acc_loss)
if acc_perp > max(prev_acc_perp[-5:]) and is_chief:
sess.run(model.lr_decay_op)
prev_acc_perp.append(acc_perp)
# Save checkpoint.
if is_chief:
checkpoint_path = os.path.join(checkpoint_dir, "neural_gpu.ckpt")
model.saver.save(sess, checkpoint_path,
global_step=model.global_step)
# Run evaluation.
bin_bound = 4
for p in FLAGS.problem.split("-"):
total_loss, total_err, tl_counter = 0.0, 0.0, 0
for bin_id in xrange(len(data.bins)):
if bin_id < bin_bound or bin_id % FLAGS.eval_bin_print == 1:
err, _, loss = single_test(bin_id, model, sess, FLAGS.nprint,
batch_size * 4, dev_set, p,
beam_model=beam_model)
if loss > 0.0:
total_loss += loss
total_err += err
tl_counter += 1
test_loss = total_loss / max(1, tl_counter)
test_err = total_err / max(1, tl_counter)
test_perp = data.safe_exp(test_loss)
summary = tf.Summary()
summary.value.extend(
[tf.Summary.Value(tag="test/%s/loss" % p, simple_value=test_loss),
tf.Summary.Value(tag="test/%s/error" % p, simple_value=test_err),
tf.Summary.Value(tag="test/%s/perplexity" % p,
simple_value=test_perp)])
sv.SummaryComputed(sess, summary, global_step)
def __init__(self,
nmaps,
vec_size,
niclass,
noclass,
dropout,
max_grad_norm,
cutoff,
nconvs,
kw,
kh,
height,
mem_size,
learning_rate,
min_length,
num_gpus,
num_replicas,
grad_noise_scale,
sampling_rate,
act_noise=0.0,
do_rnn=False,
atrous=False,
beam_size=1,
backward=True,
do_layer_norm=False,
autoenc_decay=1.0):
#todo Feeds for parameters and ops to update them.
self.nmaps = nmaps
if backward:
self.global_step = tf.Variable(0,
trainable=False,
name="global_step")
self.cur_length = tf.Variable(min_length, trainable=False)
self.cur_length_incr_op = self.cur_length.assign_add(1)
self.lr = tf.Variable(learning_rate, trainable=False)
self.lr_decay_op = self.lr.assign(self.lr * 0.995)
self.do_training = tf.placeholder(tf.float32, name="do_training")
self.update_mem = tf.placeholder(tf.int32, name="update_mem")
self.noise_param = tf.placeholder(tf.float32, name="noise_param")
self.input = tf.placeholder(tf.int32, name="inp")
self.target = tf.placeholder(tf.int32, name="tgt")
self.prev_step = tf.placeholder(tf.float32, name="prev_step")
gpu_input = tf.split(axis=0,
num_or_size_splits=num_gpus,
value=self.input)
gpu_target = tf.split(axis=0,
num_or_size_splits=num_gpus,
value=self.target)
gpu_prev_step = tf.split(axis=0,
num_or_size_splits=num_gpus,
value=self.prev_step)
batch_size = tf.shape(gpu_input[0])[0]
if backward:
adam_lr = 0.005 * self.lr
adam = tf.train.AdamOptimizer(adam_lr, epsilon=1e-3)
def adam_update(grads):
return adam.apply_gradients(zip(grads,
tf.trainable_variables()),
global_step=self.global_step,
name="adam_update")
#todo When switching from Adam to SGD we perform reverse-decay.
if backward:
global_step_float = tf.cast(self.global_step, tf.float32)
sampling_decay_exponent = global_step_float / 100000.0
sampling_decay = tf.maximum(0.05,
tf.pow(0.5, sampling_decay_exponent))
self.sampling = sampling_rate * 0.05 / sampling_decay
else:
self.sampling = tf.constant(0.0)
#todo Cache variables on cpu if needed.
if num_replicas > 1 or num_gpus > 1:
with tf.device("/cpu:0"):
caching_const = tf.constant(0)
tf.get_variable_scope().set_caching_device(caching_const.op.device)
def gpu_avg(l):
if l[0] is None:
for elem in l:
assert elem is None
return 0.0
if len(l) < 2:
return l[0]
return sum(l) / float(num_gpus)
self.length_tensor = tf.placeholder(tf.int32, name="length")
with tf.device("/cpu:0"):
emb_weights = tf.get_variable(
"embedding", [niclass, vec_size],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
if beam_size > 0:
target_emb_weights = tf.get_variable(
"target_embedding", [noclass, nmaps],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
e0 = tf.scatter_update(emb_weights,
tf.constant(0, dtype=tf.int32, shape=[1]),
tf.zeros([1, vec_size]))
output_w = tf.get_variable("output_w", [nmaps, noclass],
tf.float32)
def conv_rate(layer):
if atrous:
return 2**layer
return 1
# pylint: disable=cell-var-from-loop
def enc_step(step):
"""Encoder step."""
if autoenc_decay < 1.0:
quant_step = autoenc_quantize(step, 16, nmaps,
self.do_training)
if backward:
exp_glob = tf.train.exponential_decay(
1.0, self.global_step - 10000, 1000, autoenc_decay)
dec_factor = 1.0 - exp_glob # * self.do_training
dec_factor = tf.cond(tf.less(self.global_step, 10500),
lambda: tf.constant(0.05),
lambda: dec_factor)
else:
dec_factor = 1.0
cur = tf.cond(tf.less(tf.random_uniform([]), dec_factor),
lambda: quant_step, lambda: step)
else:
cur = step
if dropout > 0.0001:
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
# Do nconvs-many CGRU steps.
if do_jit and tf.get_variable_scope().reuse:
with jit_scope():
for layer in range(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps,
conv_rate(layer), cutoff,
"ecgru_%d" % layer, do_layer_norm)
else:
for layer in range(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps, conv_rate(layer),
cutoff, "ecgru_%d" % layer, do_layer_norm)
return cur
zero_tgt = tf.zeros([batch_size, nmaps, 1])
zero_tgt.set_shape([None, nmaps, 1])
def dec_substep(step, decided):
"""Decoder sub-step."""
cur = step
if dropout > 0.0001:
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
# Do nconvs-many CGRU steps.
if do_jit and tf.get_variable_scope().reuse:
with jit_scope():
for layer in range(nconvs):
cur = conv_gru([decided], cur, kw, kh, nmaps,
conv_rate(layer), cutoff,
"dcgru_%d" % layer, do_layer_norm)
else:
for layer in range(nconvs):
cur = conv_gru([decided], cur, kw, kh, nmaps,
conv_rate(layer), cutoff,
"dcgru_%d" % layer, do_layer_norm)
return cur
# pylint: enable=cell-var-from-loop
def dec_step(step, it, it_int, decided, output_ta, tgts, mloss,
nupd_in, out_idx, beam_cost):
"""Decoder step."""
nupd, mem_loss = 0, 0.0
if mem_size > 0:
it_incr = tf.minimum(it + 1, length - 1)
mem, mem_loss, nupd = memory_run(
step, nmaps, mem_size, batch_size, noclass,
self.global_step, self.do_training, self.update_mem, 10,
num_gpus, target_emb_weights, output_w, gpu_targets_tn,
it_incr)
step = dec_substep(step, decided)
output_l = tf.expand_dims(tf.expand_dims(step[:, it, 0, :], 1), 1)
# Calculate argmax output.
output = tf.reshape(output_l, [-1, nmaps])
# pylint: disable=cell-var-from-loop
output = tf.matmul(output, output_w)
if beam_size > 1:
beam_cost, output, out, reordered = reorder_beam(
beam_size, batch_size, beam_cost, output, it_int == 0,
[output_l, out_idx, step, decided])
[output_l, out_idx, step, decided] = reordered
else:
# Scheduled sampling.
out = tf.multinomial(tf.stop_gradient(output), 1)
out = tf.to_int32(tf.squeeze(out, [1]))
out_write = output_ta.write(it, output_l[:batch_size, :, :, :])
output = tf.gather(target_emb_weights, out)
output = tf.reshape(output, [-1, 1, nmaps])
output = tf.concat(axis=1, values=[output] * height)
tgt = tgts[it, :, :, :]
selected = tf.cond(tf.less(tf.random_uniform([]), self.sampling),
lambda: output, lambda: tgt)
# pylint: enable=cell-var-from-loop
dec_write = place_at14(decided, tf.expand_dims(selected, 1), it)
out_idx = place_at13(
out_idx, tf.reshape(out, [beam_size * batch_size, 1, 1]), it)
if mem_size > 0:
mem = tf.concat(axis=2, values=[mem] * height)
dec_write = place_at14(dec_write, mem, it_incr)
return (step, dec_write, out_write, mloss + mem_loss,
nupd_in + nupd, out_idx, beam_cost)
# Main model construction.
gpu_outputs = []
gpu_losses = []
gpu_grad_norms = []
grads_list = []
gpu_out_idx = []
self.after_enc_step = []
for gpu in range(
num_gpus): # Multi-GPU towers, average gradients later.
length = self.length_tensor
length_float = tf.cast(length, tf.float32)
if gpu > 0:
tf.get_variable_scope().reuse_variables()
gpu_outputs.append([])
gpu_losses.append([])
gpu_grad_norms.append([])
with tf.name_scope("gpu%d" % gpu), tf.device("/gpu:%d" % gpu):
# Main graph creation loop.
data.print_out("Creating model.")
start_time = time.time()
# Embed inputs and calculate mask.
with tf.device("/cpu:0"):
tgt_shape = tf.shape(tf.squeeze(gpu_target[gpu], [1]))
weights = tf.where(
tf.squeeze(gpu_target[gpu], [1]) > 0,
tf.ones(tgt_shape), tf.zeros(tgt_shape))
# Embed inputs and targets.
with tf.control_dependencies([e0]):
start = tf.gather(emb_weights,
gpu_input[gpu]) # b x h x l x nmaps
gpu_targets_tn = gpu_target[gpu] # b x 1 x len
if beam_size > 0:
embedded_targets_tn = tf.gather(
target_emb_weights, gpu_targets_tn)
embedded_targets_tn = tf.transpose(
embedded_targets_tn,
[2, 0, 1, 3]) # len x b x 1 x nmaps
embedded_targets_tn = tf.concat(
axis=2, values=[embedded_targets_tn] * height)
# First image comes from start by applying convolution and adding 0s.
start = tf.transpose(start,
[0, 2, 1, 3]) # Now b x len x h x vec_s
first = conv_linear(start, 1, 1, vec_size, nmaps, 1, True, 0.0,
"input")
first = layer_norm(first, nmaps, "input")
# Computation steps.
keep_prob = dropout * 3.0 / tf.sqrt(length_float)
keep_prob = 1.0 - self.do_training * keep_prob
act_noise_scale = act_noise * self.do_training
# Start with a convolutional gate merging previous step.
step = conv_gru([gpu_prev_step[gpu]], first, kw, kh, nmaps, 1,
cutoff, "first", do_layer_norm)
# This is just for running a baseline RNN seq2seq model.
if do_rnn:
self.after_enc_step.append(
step) # Not meaningful here, but needed.
def lstm_cell():
return tf.contrib.rnn.BasicLSTMCell(height * nmaps)
cell = tf.contrib.rnn.MultiRNNCell(
[lstm_cell() for _ in range(nconvs)])
with tf.variable_scope("encoder"):
encoder_outputs, encoder_state = tf.nn.dynamic_rnn(
cell,
tf.reshape(step,
[batch_size, length, height * nmaps]),
dtype=tf.float32,
time_major=False)
# Attention.
attn = tf.layers.dense(encoder_outputs,
height * nmaps,
name="attn1")
# pylint: disable=cell-var-from-loop
@function.Defun(noinline=True)
def attention_query(query, attn_v):
vecs = tf.tanh(attn + tf.expand_dims(query, 1))
mask = tf.reduce_sum(
vecs * tf.reshape(attn_v, [1, 1, -1]), 2)
mask = tf.nn.softmax(mask)
return tf.reduce_sum(
encoder_outputs * tf.expand_dims(mask, 2), 1)
with tf.variable_scope("decoder"):
def decoder_loop_fn(state_prev_cell_out, _,
cell_inp_cur_tgt):
state, prev_cell_out = state_prev_cell_out
cell_inp, cur_tgt = cell_inp_cur_tgt
"""Decoder loop function."""
attn_q = tf.layers.dense(prev_cell_out,
height * nmaps,
name="attn_query")
attn_res = attention_query(
attn_q,
tf.get_variable(
"attn_v", [height * nmaps],
initializer=tf.random_uniform_initializer(
-0.1, 0.1)))
concatenated = tf.reshape(
tf.concat(axis=1, values=[cell_inp, attn_res]),
[batch_size, 2 * height * nmaps])
cell_inp = tf.layers.dense(concatenated,
height * nmaps,
name="attn_merge")
output, new_state = cell(cell_inp, state)
mem_loss = 0.0
if mem_size > 0:
res, mask, mem_loss = memory_call(
output, cur_tgt, height * nmaps, mem_size,
noclass, num_gpus, self.update_mem)
res = tf.gather(target_emb_weights, res)
res *= tf.expand_dims(mask[:, 0], 1)
output = tf.layers.dense(tf.concat(
axis=1, values=[output, res]),
height * nmaps,
name="rnnmem")
return new_state, output, mem_loss
# pylint: enable=cell-var-from-loop
gpu_targets = tf.squeeze(gpu_target[gpu],
[1]) # b x len
gpu_tgt_trans = tf.transpose(gpu_targets, [1, 0])
dec_zero = tf.zeros([batch_size, 1], dtype=tf.int32)
dec_inp = tf.concat(axis=1,
values=[dec_zero, gpu_targets])
dec_inp = dec_inp[:, :length]
embedded_dec_inp = tf.gather(target_emb_weights,
dec_inp)
embedded_dec_inp_proj = tf.layers.dense(
embedded_dec_inp, height * nmaps, name="dec_proj")
embedded_dec_inp_proj = tf.transpose(
embedded_dec_inp_proj, [1, 0, 2])
init_vals = (encoder_state,
tf.zeros([batch_size,
height * nmaps]), 0.0)
_, dec_outputs, mem_losses = tf.scan(
decoder_loop_fn,
(embedded_dec_inp_proj, gpu_tgt_trans),
initializer=init_vals)
mem_loss = tf.reduce_mean(mem_losses)
outputs = tf.layers.dense(dec_outputs,
nmaps,
name="out_proj")
# Final convolution to get logits, list outputs.
outputs = tf.matmul(tf.reshape(outputs, [-1, nmaps]),
output_w)
outputs = tf.reshape(outputs,
[length, batch_size, noclass])
gpu_out_idx.append(tf.argmax(outputs, 2))
else: # Here we go with the Neural GPU.
# Encoder.
enc_length = length
step = enc_step(step) # First step hard-coded.
# pylint: disable=cell-var-from-loop
i = tf.constant(1)
c = lambda i, _s: tf.less(i, enc_length)
def enc_step_lambda(i, step):
with tf.variable_scope(tf.get_variable_scope(),
reuse=True):
new_step = enc_step(step)
return (i + 1, new_step)
_, step = tf.while_loop(c,
enc_step_lambda, [i, step],
parallel_iterations=1,
swap_memory=True)
# pylint: enable=cell-var-from-loop
self.after_enc_step.append(step)
# Decoder.
if beam_size > 0:
output_ta = tf.TensorArray(dtype=tf.float32,
size=length,
dynamic_size=False,
infer_shape=False,
name="outputs")
out_idx = tf.zeros([beam_size * batch_size, length, 1],
dtype=tf.int32)
decided_t = tf.zeros(
[beam_size * batch_size, length, height, vec_size])
# Prepare for beam search.
tgts = tf.concat(axis=1,
values=[embedded_targets_tn] *
beam_size)
beam_cost = tf.zeros([batch_size, beam_size])
step = tf.concat(axis=0, values=[step] * beam_size)
# First step hard-coded.
step, decided_t, output_ta, mem_loss, nupd, oi, bc = dec_step(
step, 0, 0, decided_t, output_ta, tgts, 0.0, 0,
out_idx, beam_cost)
tf.get_variable_scope().reuse_variables()
# pylint: disable=cell-var-from-loop
def step_lambda(i, step, dec_t, out_ta, ml, nu, oi,
bc):
with tf.variable_scope(tf.get_variable_scope(),
reuse=True):
s, d, t, nml, nu, oi, bc = dec_step(
step, i, 1, dec_t, out_ta, tgts, ml, nu,
oi, bc)
return (i + 1, s, d, t, nml, nu, oi, bc)
i = tf.constant(1)
c = lambda i, _s, _d, _o, _ml, _nu, _oi, _bc: tf.less(
i, length)
_, step, _, output_ta, mem_loss, nupd, out_idx, _ = tf.while_loop(
c,
step_lambda, [
i, step, decided_t, output_ta, mem_loss, nupd,
oi, bc
],
parallel_iterations=1,
swap_memory=True)
# pylint: enable=cell-var-from-loop
gpu_out_idx.append(tf.squeeze(out_idx, [2]))
outputs = output_ta.stack()
outputs = tf.squeeze(outputs,
[2, 3]) # Now l x b x nmaps
else:
# If beam_size is 0 or less, we don't have a decoder.
mem_loss = 0.0
outputs = tf.transpose(step[:, :, 1, :], [1, 0, 2])
gpu_out_idx.append(tf.argmax(outputs, 2))
# Final convolution to get logits, list outputs.
outputs = tf.matmul(tf.reshape(outputs, [-1, nmaps]),
output_w)
outputs = tf.reshape(outputs,
[length, batch_size, noclass])
gpu_outputs[gpu] = tf.nn.softmax(outputs)
# Calculate cross-entropy loss and normalize it.
targets_soft = make_dense(tf.squeeze(gpu_target[gpu], [1]),
noclass, 0.1)
targets_soft = tf.reshape(targets_soft, [-1, noclass])
targets_hard = make_dense(tf.squeeze(gpu_target[gpu], [1]),
noclass, 0.0)
targets_hard = tf.reshape(targets_hard, [-1, noclass])
output = tf.transpose(outputs, [1, 0, 2])
xent_soft = tf.reshape(
tf.nn.softmax_cross_entropy_with_logits(
logits=tf.reshape(output, [-1, noclass]),
labels=targets_soft), [batch_size, length])
xent_hard = tf.reshape(
tf.nn.softmax_cross_entropy_with_logits(
logits=tf.reshape(output, [-1, noclass]),
labels=targets_hard), [batch_size, length])
low, high = 0.1 / float(noclass - 1), 0.9
const = high * tf.log(high) + float(noclass -
1) * low * tf.log(low)
weight_sum = tf.reduce_sum(weights) + 1e-20
true_perp = tf.reduce_sum(xent_hard * weights) / weight_sum
soft_loss = tf.reduce_sum(xent_soft * weights) / weight_sum
perp_loss = soft_loss + const
# Final loss: cross-entropy + shared parameter relaxation part + extra.
mem_loss = 0.5 * tf.reduce_mean(mem_loss) / length_float
total_loss = perp_loss + mem_loss
gpu_losses[gpu].append(true_perp)
# Gradients.
if backward:
data.print_out("Creating backward pass for the model.")
grads = tf.gradients(total_loss,
tf.trainable_variables(),
colocate_gradients_with_ops=True)
for g_i, g in enumerate(grads):
if isinstance(g, tf.IndexedSlices):
grads[g_i] = tf.convert_to_tensor(g)
grads, norm = tf.clip_by_global_norm(grads, max_grad_norm)
gpu_grad_norms[gpu].append(norm)
for g in grads:
if grad_noise_scale > 0.001:
g += tf.truncated_normal(
tf.shape(g)) * self.noise_param
grads_list.append(grads)
else:
gpu_grad_norms[gpu].append(0.0)
data.print_out("Created model for gpu %d in %.2f s." %
(gpu, time.time() - start_time))
self.updates = []
self.after_enc_step = tf.concat(
axis=0, values=self.after_enc_step) # Concat GPUs.
if backward:
tf.get_variable_scope()._reuse = False
tf.get_variable_scope().set_caching_device(None)
grads = [
gpu_avg([grads_list[g][i] for g in range(num_gpus)])
for i in range(len(grads_list[0]))
]
update = adam_update(grads)
self.updates.append(update)
else:
self.updates.append(tf.no_op())
self.losses = [
gpu_avg([gpu_losses[g][i] for g in range(num_gpus)])
for i in range(len(gpu_losses[0]))
]
self.out_idx = tf.concat(axis=0, values=gpu_out_idx)
self.grad_norms = [
gpu_avg([gpu_grad_norms[g][i] for g in range(num_gpus)])
for i in range(len(gpu_grad_norms[0]))
]
self.outputs = [
tf.concat(axis=1, values=[gpu_outputs[g] for g in range(num_gpus)])
]
self.quantize_op = quantize_weights_op(512, 8)
if backward:
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=10)
def __init__(self, nmaps, vec_size, niclass, noclass, dropout, rx_step,
max_grad_norm, cutoff, nconvs, kw, kh, height, mode,
learning_rate, pull, pull_incr, min_length, act_noise=0.0):
# Feeds for parameters and ops to update them.
self.global_step = tf.Variable(0, trainable=False)
self.cur_length = tf.Variable(min_length, trainable=False)
self.cur_length_incr_op = self.cur_length.assign_add(1)
self.lr = tf.Variable(float(learning_rate), trainable=False)
self.lr_decay_op = self.lr.assign(self.lr * 0.98)
self.pull = tf.Variable(float(pull), trainable=False)
self.pull_incr_op = self.pull.assign(self.pull * pull_incr)
self.do_training = tf.placeholder(tf.float32, name="do_training")
self.noise_param = tf.placeholder(tf.float32, name="noise_param")
# Feeds for inputs, targets, outputs, losses, etc.
self.input = []
self.target = []
for l in xrange(data_utils.forward_max + 1):
self.input.append(tf.placeholder(tf.int32, name="inp{0}".format(l)))
self.target.append(tf.placeholder(tf.int32, name="tgt{0}".format(l)))
self.outputs = []
self.losses = []
self.grad_norms = []
self.updates = []
# Computation.
inp0_shape = tf.shape(self.input[0])
batch_size = inp0_shape[0]
with tf.device("/cpu:0"):
emb_weights = tf.get_variable(
"embedding", [niclass, vec_size],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
e0 = tf.scatter_update(emb_weights,
tf.constant(0, dtype=tf.int32, shape=[1]),
tf.zeros([1, vec_size]))
adam = tf.train.AdamOptimizer(self.lr, epsilon=1e-4)
# Main graph creation loop, for every bin in data_utils.
self.steps = []
for length in sorted(list(set(data_utils.bins + [data_utils.forward_max]))):
data_utils.print_out("Creating model for bin of length %d." % length)
start_time = time.time()
if length > data_utils.bins[0]:
tf.get_variable_scope().reuse_variables()
# Embed inputs and calculate mask.
with tf.device("/cpu:0"):
with tf.control_dependencies([e0]):
embedded = [tf.nn.embedding_lookup(emb_weights, self.input[l])
for l in xrange(length)]
# Mask to 0-out padding space in each step.
imask = [check_for_zero(self.input[l]) for l in xrange(length)]
omask = [check_for_zero(self.target[l]) for l in xrange(length)]
mask = [1.0 - (imask[i] * omask[i]) for i in xrange(length)]
mask = [tf.reshape(m, [-1, 1]) for m in mask]
# Use a shifted mask for step scaling and concatenated for weights.
shifted_mask = mask + [tf.zeros_like(mask[0])]
scales = [shifted_mask[i] * (1.0 - shifted_mask[i+1])
for i in xrange(length)]
scales = [tf.reshape(s, [-1, 1, 1, 1]) for s in scales]
mask = tf.concat(1, mask[0:length]) # batch x length
weights = mask
# Add a height dimension to mask to use later for masking.
mask = tf.reshape(mask, [-1, length, 1, 1])
mask = tf.concat(2, [mask for _ in xrange(height)]) + tf.zeros(
tf.pack([batch_size, length, height, nmaps]), dtype=tf.float32)
# Start is a length-list of batch-by-nmaps tensors, reshape and concat.
start = [tf.tanh(embedded[l]) for l in xrange(length)]
start = [tf.reshape(start[l], [-1, 1, nmaps]) for l in xrange(length)]
start = tf.reshape(tf.concat(1, start), [-1, length, 1, nmaps])
# First image comes from start by applying one convolution and adding 0s.
first = conv_linear(start, 1, 1, vec_size, nmaps, True, 0.0, "input")
first = [first] + [tf.zeros(tf.pack([batch_size, length, 1, nmaps]),
dtype=tf.float32) for _ in xrange(height - 1)]
first = tf.concat(2, first)
# Computation steps.
keep_prob = 1.0 - self.do_training * (dropout * 8.0 / float(length))
step = [tf.nn.dropout(first, keep_prob) * mask]
act_noise_scale = act_noise * self.do_training * self.pull
outputs = []
for it in xrange(length):
with tf.variable_scope("RX%d" % (it % rx_step)) as vs:
if it >= rx_step:
vs.reuse_variables()
cur = step[it]
# Do nconvs-many CGRU steps.
for layer in xrange(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps, cutoff, "cgru_%d" % layer)
cur *= mask
outputs.append(tf.slice(cur, [0, 0, 0, 0], [-1, -1, 1, -1]))
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
step.append(cur * mask)
self.steps.append([tf.reshape(s, [-1, length, height * nmaps])
for s in step])
# Output is the n-th step output; n = current length, as in scales.
output = tf.add_n([outputs[i] * scales[i] for i in xrange(length)])
# Final convolution to get logits, list outputs.
output = conv_linear(output, 1, 1, nmaps, noclass, True, 0.0, "output")
output = tf.reshape(output, [-1, length, noclass])
external_output = [tf.reshape(o, [-1, noclass])
for o in list(tf.split(1, length, output))]
external_output = [tf.nn.softmax(o) for o in external_output]
self.outputs.append(external_output)
# Calculate cross-entropy loss and normalize it.
targets = tf.concat(1, [make_dense(self.target[l], noclass)
for l in xrange(length)])
targets = tf.reshape(targets, [-1, noclass])
xent = tf.reshape(tf.nn.softmax_cross_entropy_with_logits(
tf.reshape(output, [-1, noclass]), targets), [-1, length])
perp_loss = tf.reduce_sum(xent * weights)
perp_loss /= tf.cast(batch_size, dtype=tf.float32)
perp_loss /= length
# Final loss: cross-entropy + shared parameter relaxation part.
relax_dist, self.avg_op = relaxed_distance(rx_step)
total_loss = perp_loss + relax_dist * self.pull
self.losses.append(perp_loss)
# Gradients and Adam update operation.
if length == data_utils.bins[0] or (mode == 0 and
length < data_utils.bins[-1] + 1):
data_utils.print_out("Creating backward for bin of length %d." % length)
params = tf.trainable_variables()
grads = tf.gradients(total_loss, params)
grads, norm = tf.clip_by_global_norm(grads, max_grad_norm)
self.grad_norms.append(norm)
for grad in grads:
if isinstance(grad, tf.Tensor):
grad += tf.truncated_normal(tf.shape(grad)) * self.noise_param
update = adam.apply_gradients(zip(grads, params),
global_step=self.global_step)
self.updates.append(update)
data_utils.print_out("Created model for bin of length %d in"
" %.2f s." % (length, time.time() - start_time))
self.saver = tf.train.Saver(tf.all_variables())
def __init__(self, nmaps, vec_size, niclass, noclass, dropout,
max_grad_norm, cutoff, nconvs, kw, kh, height, mem_size,
learning_rate, min_length, num_gpus, num_replicas,
grad_noise_scale, sampling_rate, act_noise=0.0, do_rnn=False,
atrous=False, beam_size=1, backward=True, do_layer_norm=False,
autoenc_decay=1.0):
# Feeds for parameters and ops to update them.
self.nmaps = nmaps
if backward:
self.global_step = tf.Variable(0, trainable=False, name="global_step")
self.cur_length = tf.Variable(min_length, trainable=False)
self.cur_length_incr_op = self.cur_length.assign_add(1)
self.lr = tf.Variable(learning_rate, trainable=False)
self.lr_decay_op = self.lr.assign(self.lr * 0.995)
self.do_training = tf.placeholder(tf.float32, name="do_training")
self.update_mem = tf.placeholder(tf.int32, name="update_mem")
self.noise_param = tf.placeholder(tf.float32, name="noise_param")
# Feeds for inputs, targets, outputs, losses, etc.
self.input = tf.placeholder(tf.int32, name="inp")
self.target = tf.placeholder(tf.int32, name="tgt")
self.prev_step = tf.placeholder(tf.float32, name="prev_step")
gpu_input = tf.split(axis=0, num_or_size_splits=num_gpus, value=self.input)
gpu_target = tf.split(axis=0, num_or_size_splits=num_gpus, value=self.target)
gpu_prev_step = tf.split(axis=0, num_or_size_splits=num_gpus, value=self.prev_step)
batch_size = tf.shape(gpu_input[0])[0]
if backward:
adam_lr = 0.005 * self.lr
adam = tf.train.AdamOptimizer(adam_lr, epsilon=1e-3)
def adam_update(grads):
return adam.apply_gradients(zip(grads, tf.trainable_variables()),
global_step=self.global_step,
name="adam_update")
# When switching from Adam to SGD we perform reverse-decay.
if backward:
global_step_float = tf.cast(self.global_step, tf.float32)
sampling_decay_exponent = global_step_float / 100000.0
sampling_decay = tf.maximum(0.05, tf.pow(0.5, sampling_decay_exponent))
self.sampling = sampling_rate * 0.05 / sampling_decay
else:
self.sampling = tf.constant(0.0)
# Cache variables on cpu if needed.
if num_replicas > 1 or num_gpus > 1:
with tf.device("/cpu:0"):
caching_const = tf.constant(0)
tf.get_variable_scope().set_caching_device(caching_const.op.device)
# partitioner = tf.variable_axis_size_partitioner(1024*256*4)
# tf.get_variable_scope().set_partitioner(partitioner)
def gpu_avg(l):
if l[0] is None:
for elem in l:
assert elem is None
return 0.0
if len(l) < 2:
return l[0]
return sum(l) / float(num_gpus)
self.length_tensor = tf.placeholder(tf.int32, name="length")
with tf.device("/cpu:0"):
emb_weights = tf.get_variable(
"embedding", [niclass, vec_size],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
if beam_size > 0:
target_emb_weights = tf.get_variable(
"target_embedding", [noclass, nmaps],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
e0 = tf.scatter_update(emb_weights,
tf.constant(0, dtype=tf.int32, shape=[1]),
tf.zeros([1, vec_size]))
output_w = tf.get_variable("output_w", [nmaps, noclass], tf.float32)
def conv_rate(layer):
if atrous:
return 2**layer
return 1
# pylint: disable=cell-var-from-loop
def enc_step(step):
"""Encoder step."""
if autoenc_decay < 1.0:
quant_step = autoenc_quantize(step, 16, nmaps, self.do_training)
if backward:
exp_glob = tf.train.exponential_decay(1.0, self.global_step - 10000,
1000, autoenc_decay)
dec_factor = 1.0 - exp_glob # * self.do_training
dec_factor = tf.cond(tf.less(self.global_step, 10500),
lambda: tf.constant(0.05), lambda: dec_factor)
else:
dec_factor = 1.0
cur = tf.cond(tf.less(tf.random_uniform([]), dec_factor),
lambda: quant_step, lambda: step)
else:
cur = step
if dropout > 0.0001:
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
# Do nconvs-many CGRU steps.
if do_jit and tf.get_variable_scope().reuse:
with jit_scope():
for layer in xrange(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps, conv_rate(layer),
cutoff, "ecgru_%d" % layer, do_layer_norm)
else:
for layer in xrange(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps, conv_rate(layer),
cutoff, "ecgru_%d" % layer, do_layer_norm)
return cur
zero_tgt = tf.zeros([batch_size, nmaps, 1])
zero_tgt.set_shape([None, nmaps, 1])
def dec_substep(step, decided):
"""Decoder sub-step."""
cur = step
if dropout > 0.0001:
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
# Do nconvs-many CGRU steps.
if do_jit and tf.get_variable_scope().reuse:
with jit_scope():
for layer in xrange(nconvs):
cur = conv_gru([decided], cur, kw, kh, nmaps, conv_rate(layer),
cutoff, "dcgru_%d" % layer, do_layer_norm)
else:
for layer in xrange(nconvs):
cur = conv_gru([decided], cur, kw, kh, nmaps, conv_rate(layer),
cutoff, "dcgru_%d" % layer, do_layer_norm)
return cur
# pylint: enable=cell-var-from-loop
def dec_step(step, it, it_int, decided, output_ta, tgts,
mloss, nupd_in, out_idx, beam_cost):
"""Decoder step."""
nupd, mem_loss = 0, 0.0
if mem_size > 0:
it_incr = tf.minimum(it+1, length - 1)
mem, mem_loss, nupd = memory_run(
step, nmaps, mem_size, batch_size, noclass, self.global_step,
self.do_training, self.update_mem, 10, num_gpus,
target_emb_weights, output_w, gpu_targets_tn, it_incr)
step = dec_substep(step, decided)
output_l = tf.expand_dims(tf.expand_dims(step[:, it, 0, :], 1), 1)
# Calculate argmax output.
output = tf.reshape(output_l, [-1, nmaps])
# pylint: disable=cell-var-from-loop
output = tf.matmul(output, output_w)
if beam_size > 1:
beam_cost, output, out, reordered = reorder_beam(
beam_size, batch_size, beam_cost, output, it_int == 0,
[output_l, out_idx, step, decided])
[output_l, out_idx, step, decided] = reordered
else:
# Scheduled sampling.
out = tf.multinomial(tf.stop_gradient(output), 1)
out = tf.to_int32(tf.squeeze(out, [1]))
out_write = output_ta.write(it, output_l[:batch_size, :, :, :])
output = tf.gather(target_emb_weights, out)
output = tf.reshape(output, [-1, 1, nmaps])
output = tf.concat(axis=1, values=[output] * height)
tgt = tgts[it, :, :, :]
selected = tf.cond(tf.less(tf.random_uniform([]), self.sampling),
lambda: output, lambda: tgt)
# pylint: enable=cell-var-from-loop
dec_write = place_at14(decided, tf.expand_dims(selected, 1), it)
out_idx = place_at13(
out_idx, tf.reshape(out, [beam_size * batch_size, 1, 1]), it)
if mem_size > 0:
mem = tf.concat(axis=2, values=[mem] * height)
dec_write = place_at14(dec_write, mem, it_incr)
return (step, dec_write, out_write, mloss + mem_loss, nupd_in + nupd,
out_idx, beam_cost)
# Main model construction.
gpu_outputs = []
gpu_losses = []
gpu_grad_norms = []
grads_list = []
gpu_out_idx = []
self.after_enc_step = []
for gpu in xrange(num_gpus): # Multi-GPU towers, average gradients later.
length = self.length_tensor
length_float = tf.cast(length, tf.float32)
if gpu > 0:
tf.get_variable_scope().reuse_variables()
gpu_outputs.append([])
gpu_losses.append([])
gpu_grad_norms.append([])
with tf.name_scope("gpu%d" % gpu), tf.device("/gpu:%d" % gpu):
# Main graph creation loop.
data.print_out("Creating model.")
start_time = time.time()
# Embed inputs and calculate mask.
with tf.device("/cpu:0"):
tgt_shape = tf.shape(tf.squeeze(gpu_target[gpu], [1]))
weights = tf.where(tf.squeeze(gpu_target[gpu], [1]) > 0,
tf.ones(tgt_shape), tf.zeros(tgt_shape))
# Embed inputs and targets.
with tf.control_dependencies([e0]):
start = tf.gather(emb_weights, gpu_input[gpu]) # b x h x l x nmaps
gpu_targets_tn = gpu_target[gpu] # b x 1 x len
if beam_size > 0:
embedded_targets_tn = tf.gather(target_emb_weights,
gpu_targets_tn)
embedded_targets_tn = tf.transpose(
embedded_targets_tn, [2, 0, 1, 3]) # len x b x 1 x nmaps
embedded_targets_tn = tf.concat(axis=2, values=[embedded_targets_tn] * height)
# First image comes from start by applying convolution and adding 0s.
start = tf.transpose(start, [0, 2, 1, 3]) # Now b x len x h x vec_s
first = conv_linear(start, 1, 1, vec_size, nmaps, 1, True, 0.0, "input")
first = layer_norm(first, nmaps, "input")
# Computation steps.
keep_prob = dropout * 3.0 / tf.sqrt(length_float)
keep_prob = 1.0 - self.do_training * keep_prob
act_noise_scale = act_noise * self.do_training
# Start with a convolutional gate merging previous step.
step = conv_gru([gpu_prev_step[gpu]], first,
kw, kh, nmaps, 1, cutoff, "first", do_layer_norm)
# This is just for running a baseline RNN seq2seq model.
if do_rnn:
self.after_enc_step.append(step) # Not meaningful here, but needed.
lstm_cell = tf.contrib.rnn.BasicLSTMCell(height * nmaps)
cell = tf.contrib.rnn.MultiRNNCell([lstm_cell] * nconvs)
with tf.variable_scope("encoder"):
encoder_outputs, encoder_state = tf.nn.dynamic_rnn(
cell, tf.reshape(step, [batch_size, length, height * nmaps]),
dtype=tf.float32, time_major=False)
# Attention.
attn = tf.layers.dense(
encoder_outputs, height * nmaps, name="attn1")
# pylint: disable=cell-var-from-loop
@function.Defun(noinline=True)
def attention_query(query, attn_v):
vecs = tf.tanh(attn + tf.expand_dims(query, 1))
mask = tf.reduce_sum(vecs * tf.reshape(attn_v, [1, 1, -1]), 2)
mask = tf.nn.softmax(mask)
return tf.reduce_sum(encoder_outputs * tf.expand_dims(mask, 2), 1)
with tf.variable_scope("decoder"):
def decoder_loop_fn((state, prev_cell_out, _), (cell_inp, cur_tgt)):
"""Decoder loop function."""
attn_q = tf.layers.dense(prev_cell_out, height * nmaps,
name="attn_query")
attn_res = attention_query(attn_q, tf.get_variable(
"attn_v", [height * nmaps],
initializer=tf.random_uniform_initializer(-0.1, 0.1)))
concatenated = tf.reshape(tf.concat(axis=1, values=[cell_inp, attn_res]),
[batch_size, 2 * height * nmaps])
cell_inp = tf.layers.dense(
concatenated, height * nmaps, name="attn_merge")
output, new_state = cell(cell_inp, state)
mem_loss = 0.0
if mem_size > 0:
res, mask, mem_loss = memory_call(
output, cur_tgt, height * nmaps, mem_size, noclass,
num_gpus, self.update_mem)
res = tf.gather(target_emb_weights, res)
res *= tf.expand_dims(mask[:, 0], 1)
output = tf.layers.dense(
tf.concat(axis=1, values=[output, res]), height * nmaps, name="rnnmem")
return new_state, output, mem_loss
# pylint: enable=cell-var-from-loop
gpu_targets = tf.squeeze(gpu_target[gpu], [1]) # b x len
gpu_tgt_trans = tf.transpose(gpu_targets, [1, 0])
dec_zero = tf.zeros([batch_size, 1], dtype=tf.int32)
dec_inp = tf.concat(axis=1, values=[dec_zero, gpu_targets])
dec_inp = dec_inp[:, :length]
embedded_dec_inp = tf.gather(target_emb_weights, dec_inp)
embedded_dec_inp_proj = tf.layers.dense(
embedded_dec_inp, height * nmaps, name="dec_proj")
embedded_dec_inp_proj = tf.transpose(embedded_dec_inp_proj,
[1, 0, 2])
init_vals = (encoder_state,
tf.zeros([batch_size, height * nmaps]), 0.0)
_, dec_outputs, mem_losses = tf.scan(
decoder_loop_fn, (embedded_dec_inp_proj, gpu_tgt_trans),
initializer=init_vals)
mem_loss = tf.reduce_mean(mem_losses)
outputs = tf.layers.dense(dec_outputs, nmaps, name="out_proj")
# Final convolution to get logits, list outputs.
outputs = tf.matmul(tf.reshape(outputs, [-1, nmaps]), output_w)
outputs = tf.reshape(outputs, [length, batch_size, noclass])
gpu_out_idx.append(tf.argmax(outputs, 2))
else: # Here we go with the Neural GPU.
# Encoder.
enc_length = length
step = enc_step(step) # First step hard-coded.
# pylint: disable=cell-var-from-loop
i = tf.constant(1)
c = lambda i, _s: tf.less(i, enc_length)
def enc_step_lambda(i, step):
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
new_step = enc_step(step)
return (i + 1, new_step)
_, step = tf.while_loop(
c, enc_step_lambda, [i, step],
parallel_iterations=1, swap_memory=True)
# pylint: enable=cell-var-from-loop
self.after_enc_step.append(step)
# Decoder.
if beam_size > 0:
output_ta = tf.TensorArray(
dtype=tf.float32, size=length, dynamic_size=False,
infer_shape=False, name="outputs")
out_idx = tf.zeros([beam_size * batch_size, length, 1],
dtype=tf.int32)
decided_t = tf.zeros([beam_size * batch_size, length,
height, vec_size])
# Prepare for beam search.
tgts = tf.concat(axis=1, values=[embedded_targets_tn] * beam_size)
beam_cost = tf.zeros([batch_size, beam_size])
step = tf.concat(axis=0, values=[step] * beam_size)
# First step hard-coded.
step, decided_t, output_ta, mem_loss, nupd, oi, bc = dec_step(
step, 0, 0, decided_t, output_ta, tgts, 0.0, 0, out_idx,
beam_cost)
tf.get_variable_scope().reuse_variables()
# pylint: disable=cell-var-from-loop
def step_lambda(i, step, dec_t, out_ta, ml, nu, oi, bc):
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
s, d, t, nml, nu, oi, bc = dec_step(
step, i, 1, dec_t, out_ta, tgts, ml, nu, oi, bc)
return (i + 1, s, d, t, nml, nu, oi, bc)
i = tf.constant(1)
c = lambda i, _s, _d, _o, _ml, _nu, _oi, _bc: tf.less(i, length)
_, step, _, output_ta, mem_loss, nupd, out_idx, _ = tf.while_loop(
c, step_lambda,
[i, step, decided_t, output_ta, mem_loss, nupd, oi, bc],
parallel_iterations=1, swap_memory=True)
# pylint: enable=cell-var-from-loop
gpu_out_idx.append(tf.squeeze(out_idx, [2]))
outputs = output_ta.stack()
outputs = tf.squeeze(outputs, [2, 3]) # Now l x b x nmaps
else:
# If beam_size is 0 or less, we don't have a decoder.
mem_loss = 0.0
outputs = tf.transpose(step[:, :, 1, :], [1, 0, 2])
gpu_out_idx.append(tf.argmax(outputs, 2))
# Final convolution to get logits, list outputs.
outputs = tf.matmul(tf.reshape(outputs, [-1, nmaps]), output_w)
outputs = tf.reshape(outputs, [length, batch_size, noclass])
gpu_outputs[gpu] = tf.nn.softmax(outputs)
# Calculate cross-entropy loss and normalize it.
targets_soft = make_dense(tf.squeeze(gpu_target[gpu], [1]),
noclass, 0.1)
targets_soft = tf.reshape(targets_soft, [-1, noclass])
targets_hard = make_dense(tf.squeeze(gpu_target[gpu], [1]),
noclass, 0.0)
targets_hard = tf.reshape(targets_hard, [-1, noclass])
output = tf.transpose(outputs, [1, 0, 2])
xent_soft = tf.reshape(tf.nn.softmax_cross_entropy_with_logits(
logits=tf.reshape(output, [-1, noclass]), labels=targets_soft),
[batch_size, length])
xent_hard = tf.reshape(tf.nn.softmax_cross_entropy_with_logits(
logits=tf.reshape(output, [-1, noclass]), labels=targets_hard),
[batch_size, length])
low, high = 0.1 / float(noclass - 1), 0.9
const = high * tf.log(high) + float(noclass - 1) * low * tf.log(low)
weight_sum = tf.reduce_sum(weights) + 1e-20
true_perp = tf.reduce_sum(xent_hard * weights) / weight_sum
soft_loss = tf.reduce_sum(xent_soft * weights) / weight_sum
perp_loss = soft_loss + const
# Final loss: cross-entropy + shared parameter relaxation part + extra.
mem_loss = 0.5 * tf.reduce_mean(mem_loss) / length_float
total_loss = perp_loss + mem_loss
gpu_losses[gpu].append(true_perp)
# Gradients.
if backward:
data.print_out("Creating backward pass for the model.")
grads = tf.gradients(
total_loss, tf.trainable_variables(),
colocate_gradients_with_ops=True)
for g_i, g in enumerate(grads):
if isinstance(g, tf.IndexedSlices):
grads[g_i] = tf.convert_to_tensor(g)
grads, norm = tf.clip_by_global_norm(grads, max_grad_norm)
gpu_grad_norms[gpu].append(norm)
for g in grads:
if grad_noise_scale > 0.001:
g += tf.truncated_normal(tf.shape(g)) * self.noise_param
grads_list.append(grads)
else:
gpu_grad_norms[gpu].append(0.0)
data.print_out("Created model for gpu %d in %.2f s."
% (gpu, time.time() - start_time))
def __init__(self, nmaps, vec_size, niclass, noclass, dropout, rx_step,
max_grad_norm, cutoff, nconvs, kw, kh, height, mode,
learning_rate, pull, pull_incr, min_length, act_noise=0.0):
# Feeds for parameters and ops to update them.
self.global_step = tf.Variable(0, trainable=False)
self.cur_length = tf.Variable(min_length, trainable=False)
self.cur_length_incr_op = self.cur_length.assign_add(1)
self.lr = tf.Variable(float(learning_rate), trainable=False)
self.lr_decay_op = self.lr.assign(self.lr * 0.98)
self.pull = tf.Variable(float(pull), trainable=False)
self.pull_incr_op = self.pull.assign(self.pull * pull_incr)
self.do_training = tf.placeholder(tf.float32, name="do_training")
self.noise_param = tf.placeholder(tf.float32, name="noise_param")
# Feeds for inputs, targets, outputs, losses, etc.
self.input = []
self.target = []
for l in xrange(data_utils.forward_max + 1):
self.input.append(tf.placeholder(tf.int32, name="inp{0}".format(l)))
self.target.append(tf.placeholder(tf.int32, name="tgt{0}".format(l)))
self.outputs = []
self.losses = []
self.grad_norms = []
self.updates = []
# Computation.
inp0_shape = tf.shape(self.input[0])
batch_size = inp0_shape[0]
with tf.device("/cpu:0"):
emb_weights = tf.get_variable(
"embedding", [niclass, vec_size],
initializer=tf.random_uniform_initializer(-1.7, 1.7))
e0 = tf.scatter_update(emb_weights,
tf.constant(0, dtype=tf.int32, shape=[1]),
tf.zeros([1, vec_size]))
adam = tf.train.AdamOptimizer(self.lr, epsilon=1e-4)
# Main graph creation loop, for every bin in data_utils.
self.steps = []
for length in sorted(list(set(data_utils.bins + [data_utils.forward_max]))):
data_utils.print_out("Creating model for bin of length %d." % length)
start_time = time.time()
if length > data_utils.bins[0]:
tf.get_variable_scope().reuse_variables()
# Embed inputs and calculate mask.
with tf.device("/cpu:0"):
with tf.control_dependencies([e0]):
embedded = [tf.nn.embedding_lookup(emb_weights, self.input[l])
for l in xrange(length)]
# Mask to 0-out padding space in each step.
imask = [check_for_zero(self.input[l]) for l in xrange(length)]
omask = [check_for_zero(self.target[l]) for l in xrange(length)]
mask = [1.0 - (imask[i] * omask[i]) for i in xrange(length)]
mask = [tf.reshape(m, [-1, 1]) for m in mask]
# Use a shifted mask for step scaling and concatenated for weights.
shifted_mask = mask + [tf.zeros_like(mask[0])]
scales = [shifted_mask[i] * (1.0 - shifted_mask[i+1])
for i in xrange(length)]
scales = [tf.reshape(s, [-1, 1, 1, 1]) for s in scales]
mask = tf.concat(1, mask[0:length]) # batch x length
weights = mask
# Add a height dimension to mask to use later for masking.
mask = tf.reshape(mask, [-1, length, 1, 1])
mask = tf.concat(2, [mask for _ in xrange(height)]) + tf.zeros(
tf.pack([batch_size, length, height, nmaps]), dtype=tf.float32)
# Start is a length-list of batch-by-nmaps tensors, reshape and concat.
start = [tf.tanh(embedded[l]) for l in xrange(length)]
start = [tf.reshape(start[l], [-1, 1, nmaps]) for l in xrange(length)]
start = tf.reshape(tf.concat(1, start), [-1, length, 1, nmaps])
# First image comes from start by applying one convolution and adding 0s.
first = conv_linear(start, 1, 1, vec_size, nmaps, True, 0.0, "input")
first = [first] + [tf.zeros(tf.pack([batch_size, length, 1, nmaps]),
dtype=tf.float32) for _ in xrange(height - 1)]
first = tf.concat(2, first)
# Computation steps.
keep_prob = 1.0 - self.do_training * (dropout * 8.0 / float(length))
step = [tf.nn.dropout(first, keep_prob) * mask]
act_noise_scale = act_noise * self.do_training * self.pull
outputs = []
for it in xrange(length):
with tf.variable_scope("RX%d" % (it % rx_step)) as vs:
if it >= rx_step:
vs.reuse_variables()
cur = step[it]
# Do nconvs-many CGRU steps.
for layer in xrange(nconvs):
cur = conv_gru([], cur, kw, kh, nmaps, cutoff, "cgru_%d" % layer)
cur *= mask
outputs.append(tf.slice(cur, [0, 0, 0, 0], [-1, -1, 1, -1]))
cur = tf.nn.dropout(cur, keep_prob)
if act_noise > 0.00001:
cur += tf.truncated_normal(tf.shape(cur)) * act_noise_scale
step.append(cur * mask)
self.steps.append([tf.reshape(s, [-1, length, height * nmaps])
for s in step])
# Output is the n-th step output; n = current length, as in scales.
output = tf.add_n([outputs[i] * scales[i] for i in xrange(length)])
# Final convolution to get logits, list outputs.
output = conv_linear(output, 1, 1, nmaps, noclass, True, 0.0, "output")
output = tf.reshape(output, [-1, length, noclass])
external_output = [tf.reshape(o, [-1, noclass])
for o in list(tf.split(1, length, output))]
external_output = [tf.nn.softmax(o) for o in external_output]
self.outputs.append(external_output)
# Calculate cross-entropy loss and normalize it.
targets = tf.concat(1, [make_dense(self.target[l], noclass)
for l in xrange(length)])
targets = tf.reshape(targets, [-1, noclass])
xent = tf.reshape(tf.nn.softmax_cross_entropy_with_logits(
tf.reshape(output, [-1, noclass]), targets), [-1, length])
perp_loss = tf.reduce_sum(xent * weights)
perp_loss /= tf.cast(batch_size, dtype=tf.float32)
perp_loss /= length
# Final loss: cross-entropy + shared parameter relaxation part.
relax_dist, self.avg_op = relaxed_distance(rx_step)
total_loss = perp_loss + relax_dist * self.pull
self.losses.append(perp_loss)
# Gradients and Adam update operation.
if length == data_utils.bins[0] or (mode == 0 and
length < data_utils.bins[-1] + 1):
data_utils.print_out("Creating backward for bin of length %d." % length)
params = tf.trainable_variables()
grads = tf.gradients(total_loss, params)
grads, norm = tf.clip_by_global_norm(grads, max_grad_norm)
self.grad_norms.append(norm)
for grad in grads:
if isinstance(grad, tf.Tensor):
grad += tf.truncated_normal(tf.shape(grad)) * self.noise_param
update = adam.apply_gradients(zip(grads, params),
global_step=self.global_step)
self.updates.append(update)
data_utils.print_out("Created model for bin of length %d in"
" %.2f s." % (length, time.time() - start_time))
self.saver = tf.train.Saver(tf.all_variables())
def train():
"""Train the model."""
batch_size = FLAGS.batch_size
tasks = FLAGS.task.split("-")
with tf.Session() as sess:
(model, min_length, max_length, checkpoint_dir, curriculum,
_) = initialize(sess)
quant_op = neural_gpu.quantize_weights_op(512, 8)
max_cur_length = min(min_length + 3, max_length)
prev_acc_perp = [1000000 for _ in xrange(3)]
prev_seq_err = 1.0
# Main traning loop.
while True:
global_step, pull, max_cur_length, learning_rate = sess.run(
[model.global_step, model.pull, model.cur_length, model.lr])
acc_loss, acc_total, acc_errors, acc_seq_err = 0.0, 0, 0, 0
acc_grad_norm, step_count, step_time = 0.0, 0, 0.0
for _ in xrange(FLAGS.steps_per_checkpoint):
global_step += 1
task = random.choice(tasks)
# Select the length for curriculum learning.
l = np.random.randint(max_cur_length - min_length +
1) + min_length
# Prefer longer stuff 60% of time.
if np.random.randint(100) < 60:
l1 = np.random.randint(max_cur_length - min_length +
1) + min_length
l = max(l, l1)
# Mixed curriculum learning: in 25% of cases go to any larger length.
if np.random.randint(100) < 25:
l1 = np.random.randint(max_length - min_length +
1) + min_length
l = max(l, l1)
# Run a step and time it.
start_time = time.time()
inp, target = data.get_batch(l, batch_size, True, task)
noise_param = math.sqrt(
math.pow(global_step, -0.55) *
prev_seq_err) * FLAGS.grad_noise_scale
loss, res, gnorm, _ = model.step(sess, inp, target, True,
noise_param)
step_time += time.time() - start_time
acc_grad_norm += float(gnorm)
# Accumulate statistics only if we did not exceed curriculum length.
if l < max_cur_length + 1:
step_count += 1
acc_loss += loss
errors, total, seq_err = data.accuracy(
inp, res, target, batch_size, 0)
acc_total += total
acc_errors += errors
acc_seq_err += seq_err
# Normalize and print out accumulated statistics.
acc_loss /= step_count
step_time /= FLAGS.steps_per_checkpoint
acc_seq_err = float(acc_seq_err) / (step_count * batch_size)
prev_seq_err = max(0.0,
acc_seq_err - 0.02) # No noise at error < 2%.
acc_errors = float(
acc_errors) / acc_total if acc_total > 0 else 1.0
msg1 = "step %d step-time %.2f" % (global_step, step_time)
msg2 = "lr %.8f pull %.3f" % (learning_rate, pull)
msg3 = ("%s %s grad-norm %.8f" %
(msg1, msg2, acc_grad_norm / FLAGS.steps_per_checkpoint))
data.print_out(
"%s len %d ppx %.8f errors %.2f sequence-errors %.2f" %
(msg3, max_cur_length, data.safe_exp(acc_loss),
100 * acc_errors, 100 * acc_seq_err))
# If errors are below the curriculum threshold, move curriculum forward.
if curriculum > acc_seq_err:
if FLAGS.quantize:
# Quantize weights.
data.print_out(" Quantizing parameters.")
sess.run([quant_op])
# Increase current length (until the next with training data).
do_incr = True
while do_incr and max_cur_length < max_length:
sess.run(model.cur_length_incr_op)
for t in tasks:
if data.train_set[t]: do_incr = False
# Forget last perplexities if we're not yet at the end.
if max_cur_length < max_length:
prev_acc_perp.append(1000000)
# Either increase pull or, if it's large, average parameters.
if pull < 0.1:
sess.run(model.pull_incr_op)
else:
data.print_out(" Averaging parameters.")
sess.run(model.avg_op)
if acc_seq_err < (curriculum / 3.0):
sess.run(model.lr_decay_op)
# Lower learning rate if we're worse than the last 3 checkpoints.
acc_perp = data.safe_exp(acc_loss)
if acc_perp > max(prev_acc_perp[-3:]):
sess.run(model.lr_decay_op)
prev_acc_perp.append(acc_perp)
# Save checkpoint.
checkpoint_path = os.path.join(checkpoint_dir, "neural_gpu.ckpt")
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step)
# Run evaluation.
bound = data.bins[-1] + 1
for t in tasks:
l = min_length
while l < max_length + EXTRA_EVAL and l < bound:
_, seq_err, _ = single_test(l, model, sess, t,
FLAGS.nprint, batch_size)
l += 1
while l < bound + 1 and not data.test_set[t][l]:
l += 1
if seq_err < 0.05: # Run larger test if we're good enough.
_, seq_err = multi_test(data.forward_max, model, sess, t,
FLAGS.nprint, batch_size * 4)
if seq_err < 0.01: # Super-large test on 1-task large-forward models.
if data.forward_max > 4000 and len(tasks) == 1:
multi_test(data.forward_max, model, sess, tasks[0],
FLAGS.nprint, batch_size * 16, 0)
def train():
"""Train the model."""
batch_size = FLAGS.batch_size
tasks = FLAGS.task.split("-")
with tf.Session() as sess:
(model, min_length, max_length, checkpoint_dir,
curriculum, _) = initialize(sess)
quant_op = neural_gpu.quantize_weights_op(512, 8)
max_cur_length = min(min_length + 3, max_length)
prev_acc_perp = [1000000 for _ in xrange(3)]
prev_seq_err = 1.0
# Main traning loop.
while True:
global_step, pull, max_cur_length, learning_rate = sess.run(
[model.global_step, model.pull, model.cur_length, model.lr])
acc_loss, acc_total, acc_errors, acc_seq_err = 0.0, 0, 0, 0
acc_grad_norm, step_count, step_time = 0.0, 0, 0.0
for _ in xrange(FLAGS.steps_per_checkpoint):
global_step += 1
task = random.choice(tasks)
# Select the length for curriculum learning.
l = np.random.randint(max_cur_length - min_length + 1) + min_length
# Prefer longer stuff 60% of time.
if np.random.randint(100) < 60:
l1 = np.random.randint(max_cur_length - min_length+1) + min_length
l = max(l, l1)
# Mixed curriculum learning: in 25% of cases go to any larger length.
if np.random.randint(100) < 25:
l1 = np.random.randint(max_length - min_length + 1) + min_length
l = max(l, l1)
# Run a step and time it.
start_time = time.time()
inp, target = data.get_batch(l, batch_size, True, task)
noise_param = math.sqrt(math.pow(global_step, -0.55) *
prev_seq_err) * FLAGS.grad_noise_scale
loss, res, gnorm, _ = model.step(sess, inp, target, True, noise_param)
step_time += time.time() - start_time
acc_grad_norm += float(gnorm)
# Accumulate statistics only if we did not exceed curriculum length.
if l < max_cur_length + 1:
step_count += 1
acc_loss += loss
errors, total, seq_err = data.accuracy(inp, res, target,
batch_size, 0)
acc_total += total
acc_errors += errors
acc_seq_err += seq_err
# Normalize and print out accumulated statistics.
acc_loss /= step_count
step_time /= FLAGS.steps_per_checkpoint
acc_seq_err = float(acc_seq_err) / (step_count * batch_size)
prev_seq_err = max(0.0, acc_seq_err - 0.02) # No noise at error < 2%.
acc_errors = float(acc_errors) / acc_total if acc_total > 0 else 1.0
msg1 = "step %d step-time %.2f" % (global_step, step_time)
msg2 = "lr %.8f pull %.3f" % (learning_rate, pull)
msg3 = ("%s %s grad-norm %.8f"
% (msg1, msg2, acc_grad_norm / FLAGS.steps_per_checkpoint))
data.print_out("%s len %d ppx %.8f errors %.2f sequence-errors %.2f" %
(msg3, max_cur_length, data.safe_exp(acc_loss),
100*acc_errors, 100*acc_seq_err))
# If errors are below the curriculum threshold, move curriculum forward.
if curriculum > acc_seq_err:
if FLAGS.quantize:
# Quantize weights.
data.print_out(" Quantizing parameters.")
sess.run([quant_op])
# Increase current length (until the next with training data).
do_incr = True
while do_incr and max_cur_length < max_length:
sess.run(model.cur_length_incr_op)
for t in tasks:
if data.train_set[t]: do_incr = False
# Forget last perplexities if we're not yet at the end.
if max_cur_length < max_length:
prev_acc_perp.append(1000000)
# Either increase pull or, if it's large, average parameters.
if pull < 0.1:
sess.run(model.pull_incr_op)
else:
data.print_out(" Averaging parameters.")
sess.run(model.avg_op)
if acc_seq_err < (curriculum / 3.0):
sess.run(model.lr_decay_op)
# Lower learning rate if we're worse than the last 3 checkpoints.
acc_perp = data.safe_exp(acc_loss)
if acc_perp > max(prev_acc_perp[-3:]):
sess.run(model.lr_decay_op)
prev_acc_perp.append(acc_perp)
# Save checkpoint.
checkpoint_path = os.path.join(checkpoint_dir, "neural_gpu.ckpt")
model.saver.save(sess, checkpoint_path,
global_step=model.global_step)
# Run evaluation.
bound = data.bins[-1] + 1
for t in tasks:
l = min_length
while l < max_length + EXTRA_EVAL and l < bound:
_, seq_err, _ = single_test(l, model, sess, t,
FLAGS.nprint, batch_size)
l += 1
while l < bound + 1 and not data.test_set[t][l]:
l += 1
if seq_err < 0.05: # Run larger test if we're good enough.
_, seq_err = multi_test(data.forward_max, model, sess, t,
FLAGS.nprint, batch_size * 4)
if seq_err < 0.01: # Super-large test on 1-task large-forward models.
if data.forward_max > 4000 and len(tasks) == 1:
multi_test(data.forward_max, model, sess, tasks[0], FLAGS.nprint,
batch_size * 16, 0)
def initialize(sess):
"""Initialize data and model."""
if FLAGS.jobid >= 0:
data.log_filename = os.path.join(FLAGS.train_dir, "log%d" % FLAGS.jobid)
data.print_out("NN ", newline=False)
# Set random seed.
seed = FLAGS.random_seed + max(0, FLAGS.jobid)
tf.set_random_seed(seed)
random.seed(seed)
np.random.seed(seed)
# Check data sizes.
assert data.bins
min_length = 3
max_length = min(FLAGS.max_length, data.bins[-1])
assert max_length + 1 > min_length
while len(data.bins) > 1 and data.bins[-2] > max_length + EXTRA_EVAL:
data.bins = data.bins[:-1]
assert data.bins[0] > FLAGS.rx_step
data.forward_max = max(FLAGS.forward_max, data.bins[-1])
nclass = min(FLAGS.niclass, FLAGS.noclass)
data_size = FLAGS.train_data_size if FLAGS.mode == 0 else 1000
# Initialize data for each task.
tasks = FLAGS.task.split("-")
for t in tasks:
for l in xrange(max_length + EXTRA_EVAL - 1):
data.init_data(t, l, data_size, nclass)
data.init_data(t, data.bins[-2], data_size, nclass)
data.init_data(t, data.bins[-1], data_size, nclass)
end_size = 4 * 1024 if FLAGS.mode > 0 else 1024
data.init_data(t, data.forward_max, end_size, nclass)
# Print out parameters.
curriculum = FLAGS.curriculum_bound
msg1 = ("layers %d kw %d h %d kh %d relax %d batch %d noise %.2f task %s"
% (FLAGS.nconvs, FLAGS.kw, FLAGS.height, FLAGS.kh, FLAGS.rx_step,
FLAGS.batch_size, FLAGS.grad_noise_scale, FLAGS.task))
msg2 = "data %d %s" % (FLAGS.train_data_size, msg1)
msg3 = ("cut %.2f pull %.3f lr %.2f iw %.2f cr %.2f nm %d d%.4f gn %.2f %s" %
(FLAGS.cutoff, FLAGS.pull_incr, FLAGS.lr, FLAGS.init_weight,
curriculum, FLAGS.nmaps, FLAGS.dropout, FLAGS.max_grad_norm, msg2))
data.print_out(msg3)
# Create checkpoint directory if it does not exist.
checkpoint_dir = os.path.join(FLAGS.train_dir, "neural_gpu%s"
% ("" if FLAGS.jobid < 0 else str(FLAGS.jobid)))
if not gfile.IsDirectory(checkpoint_dir):
data.print_out("Creating checkpoint directory %s." % checkpoint_dir)
gfile.MkDir(checkpoint_dir)
# Create model and initialize it.
tf.get_variable_scope().set_initializer(
tf.uniform_unit_scaling_initializer(factor=1.8 * FLAGS.init_weight))
model = neural_gpu.NeuralGPU(
FLAGS.nmaps, FLAGS.nmaps, FLAGS.niclass, FLAGS.noclass, FLAGS.dropout,
FLAGS.rx_step, FLAGS.max_grad_norm, FLAGS.cutoff, FLAGS.nconvs,
FLAGS.kw, FLAGS.kh, FLAGS.height, FLAGS.mode, FLAGS.lr,
FLAGS.pull, FLAGS.pull_incr, min_length + 3)
data.print_out("Created model.")
sess.run(tf.initialize_all_variables())
data.print_out("Initialized variables.")
# Load model from parameters if a checkpoint exists.
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
data.print_out("Reading model parameters from %s"
% ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
# Check if there are ensemble models and get their checkpoints.
ensemble = []
ensemble_dir_list = [d for d in FLAGS.ensemble.split(",") if d]
for ensemble_dir in ensemble_dir_list:
ckpt = tf.train.get_checkpoint_state(ensemble_dir)
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
data.print_out("Found ensemble model %s" % ckpt.model_checkpoint_path)
ensemble.append(ckpt.model_checkpoint_path)
# Return the model and needed variables.
return (model, min_length, max_length, checkpoint_dir, curriculum, ensemble)
def initialize(sess=None):
"""Initialize data and model."""
global MAXLEN_F
# Create training directory if it does not exist.
if not tf.gfile.IsDirectory(FLAGS.train_dir):
data.print_out("Creating training directory %s." % FLAGS.train_dir)
tf.gfile.MkDir(FLAGS.train_dir)
decode_suffix = "beam%dln%d" % (FLAGS.beam_size,
int(100 * FLAGS.length_norm))
if FLAGS.mode == 0:
decode_suffix = ""
if FLAGS.task >= 0:
data.log_filename = os.path.join(FLAGS.train_dir,
"log%d%s" % (FLAGS.task, decode_suffix))
else:
data.log_filename = os.path.join(FLAGS.train_dir, "neural_gpu/log")
# Set random seed.
if FLAGS.random_seed > 0:
seed = FLAGS.random_seed + max(0, FLAGS.task)
tf.set_random_seed(seed)
random.seed(seed)
np.random.seed(seed)
# Check data sizes.
assert data.bins
max_length = min(FLAGS.max_length, data.bins[-1])
while len(data.bins) > 1 and data.bins[-2] >= max_length + EXTRA_EVAL:
data.bins = data.bins[:-1]
if sess is None and FLAGS.task == 0 and FLAGS.num_replicas > 1:
if max_length > 60:
max_length = max_length * 1 / 2 # Save memory on chief.
min_length = min(14, max_length - 3) if FLAGS.problem == "wmt" else 3
for p in FLAGS.problem.split("-"):
if p in ["progeval", "progsynth"]:
min_length = max(26, min_length)
assert max_length + 1 > min_length
while len(data.bins) > 1 and data.bins[-2] >= max_length + EXTRA_EVAL:
data.bins = data.bins[:-1]
# Create checkpoint directory if it does not exist.
if FLAGS.mode == 0 or FLAGS.task < 0:
checkpoint_dir = os.path.join(FLAGS.train_dir, "neural_gpu%s"
% ("" if FLAGS.task < 0 else str(FLAGS.task)))
else:
checkpoint_dir = FLAGS.train_dir
if not tf.gfile.IsDirectory(checkpoint_dir):
data.print_out("Creating checkpoint directory %s." % checkpoint_dir)
tf.gfile.MkDir(checkpoint_dir)
# Prepare data.
if FLAGS.problem == "wmt":
# Prepare WMT data.
data.print_out("Preparing WMT data in %s" % FLAGS.data_dir)
if FLAGS.simple_tokenizer:
MAXLEN_F = 3.5
(en_train, fr_train, en_dev, fr_dev,
en_path, fr_path) = wmt.prepare_wmt_data(
FLAGS.data_dir, FLAGS.vocab_size,
tokenizer=wmt.space_tokenizer,
normalize_digits=FLAGS.normalize_digits)
else:
(en_train, fr_train, en_dev, fr_dev,
en_path, fr_path) = wmt.prepare_wmt_data(
FLAGS.data_dir, FLAGS.vocab_size)
# Read data into buckets and compute their sizes.
fr_vocab, rev_fr_vocab = wmt.initialize_vocabulary(fr_path)
data.vocab = fr_vocab
data.rev_vocab = rev_fr_vocab
data.print_out("Reading development and training data (limit: %d)."
% FLAGS.max_train_data_size)
dev_set = {}
dev_set["wmt"] = read_data(en_dev, fr_dev, data.bins)
def data_read(size, print_out):
read_data_into_global(en_train, fr_train, data.bins, size, print_out)
data_read(50000, False)
read_thread_small = threading.Thread(
name="reading-data-small", target=lambda: data_read(900000, False))
read_thread_small.start()
read_thread_full = threading.Thread(
name="reading-data-full",
target=lambda: data_read(FLAGS.max_train_data_size, True))
read_thread_full.start()
data.print_out("Data reading set up.")
else:
# Prepare algorithmic data.
en_path, fr_path = None, None
tasks = FLAGS.problem.split("-")
data_size = FLAGS.train_data_size
for t in tasks:
data.print_out("Generating data for %s." % t)
if t in ["progeval", "progsynth"]:
data.init_data(t, data.bins[-1], 20 * data_size, FLAGS.vocab_size)
if len(program_utils.prog_vocab) > FLAGS.vocab_size - 2:
raise ValueError("Increase vocab_size to %d for prog-tasks."
% (len(program_utils.prog_vocab) + 2))
data.rev_vocab = program_utils.prog_vocab
data.vocab = program_utils.prog_rev_vocab
else:
for l in xrange(max_length + EXTRA_EVAL - 1):
data.init_data(t, l, data_size, FLAGS.vocab_size)
data.init_data(t, data.bins[-2], data_size, FLAGS.vocab_size)
data.init_data(t, data.bins[-1], data_size, FLAGS.vocab_size)
if t not in global_train_set:
global_train_set[t] = []
global_train_set[t].append(data.train_set[t])
calculate_buckets_scale(data.train_set[t], data.bins, t)
dev_set = data.test_set
# Grid-search parameters.
lr = FLAGS.lr
init_weight = FLAGS.init_weight
max_grad_norm = FLAGS.max_grad_norm
if sess is not None and FLAGS.task > -1:
def job_id_factor(step):
"""If jobid / step mod 3 is 0, 1, 2: say 0, 1, -1."""
return ((((FLAGS.task / step) % 3) + 1) % 3) - 1
lr *= math.pow(2, job_id_factor(1))
init_weight *= math.pow(1.5, job_id_factor(3))
max_grad_norm *= math.pow(2, job_id_factor(9))
# Print out parameters.
curriculum = FLAGS.curriculum_seq
msg1 = ("layers %d kw %d h %d kh %d batch %d noise %.2f"
% (FLAGS.nconvs, FLAGS.kw, FLAGS.height, FLAGS.kh,
FLAGS.batch_size, FLAGS.grad_noise_scale))
msg2 = ("cut %.2f lr %.3f iw %.2f cr %.2f nm %d d%.4f gn %.2f %s"
% (FLAGS.cutoff, lr, init_weight, curriculum, FLAGS.nmaps,
FLAGS.dropout, max_grad_norm, msg1))
data.print_out(msg2)
# Create model and initialize it.
tf.get_variable_scope().set_initializer(
tf.orthogonal_initializer(gain=1.8 * init_weight))
max_sampling_rate = FLAGS.max_sampling_rate if FLAGS.mode == 0 else 0.0
o = FLAGS.vocab_size if FLAGS.max_target_vocab < 1 else FLAGS.max_target_vocab
ngpu.CHOOSE_K = FLAGS.soft_mem_size
do_beam_model = FLAGS.train_beam_freq > 0.0001 and FLAGS.beam_size > 1
beam_size = FLAGS.beam_size if FLAGS.mode > 0 and not do_beam_model else 1
beam_size = min(beam_size, FLAGS.beam_size)
beam_model = None
def make_ngpu(cur_beam_size, back):
return ngpu.NeuralGPU(
FLAGS.nmaps, FLAGS.vec_size, FLAGS.vocab_size, o,
FLAGS.dropout, max_grad_norm, FLAGS.cutoff, FLAGS.nconvs,
FLAGS.kw, FLAGS.kh, FLAGS.height, FLAGS.mem_size,
lr / math.sqrt(FLAGS.num_replicas), min_length + 3, FLAGS.num_gpus,
FLAGS.num_replicas, FLAGS.grad_noise_scale, max_sampling_rate,
atrous=FLAGS.atrous, do_rnn=FLAGS.rnn_baseline,
do_layer_norm=FLAGS.layer_norm, beam_size=cur_beam_size, backward=back)
if sess is None:
with tf.device(tf.train.replica_device_setter(FLAGS.ps_tasks)):
model = make_ngpu(beam_size, True)
if do_beam_model:
tf.get_variable_scope().reuse_variables()
beam_model = make_ngpu(FLAGS.beam_size, False)
else:
model = make_ngpu(beam_size, True)
if do_beam_model:
tf.get_variable_scope().reuse_variables()
beam_model = make_ngpu(FLAGS.beam_size, False)
sv = None
if sess is None:
# The supervisor configuration has a few overriden options.
sv = tf.train.Supervisor(logdir=checkpoint_dir,
is_chief=(FLAGS.task < 1),
saver=model.saver,
summary_op=None,
save_summaries_secs=60,
save_model_secs=15 * 60,
global_step=model.global_step)
config = tf.ConfigProto(allow_soft_placement=True)
sess = sv.PrepareSession(FLAGS.master, config=config)
data.print_out("Created model. Checkpoint dir %s" % checkpoint_dir)
# Load model from parameters if a checkpoint exists.
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and tf.gfile.Exists(ckpt.model_checkpoint_path + ".index"):
data.print_out("Reading model parameters from %s"
% ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
elif sv is None:
sess.run(tf.global_variables_initializer())
data.print_out("Initialized variables (no supervisor mode).")
elif FLAGS.task < 1 and FLAGS.mem_size > 0:
# sess.run(model.mem_norm_op)
data.print_out("Created new model and normalized mem (on chief).")
# Return the model and needed variables.
return (model, beam_model, min_length, max_length, checkpoint_dir,
(global_train_set, dev_set, en_path, fr_path), sv, sess)
def initialize(sess):
"""Initialize data and model."""
if FLAGS.jobid >= 0:
data.log_filename = os.path.join(FLAGS.train_dir,
"log%d" % FLAGS.jobid)
data.print_out("NN ", newline=False)
# Set random seed.
seed = FLAGS.random_seed + max(0, FLAGS.jobid)
tf.set_random_seed(seed)
random.seed(seed)
np.random.seed(seed)
# Check data sizes.
assert data.bins
min_length = 3
max_length = min(FLAGS.max_length, data.bins[-1])
assert max_length + 1 > min_length
while len(data.bins) > 1 and data.bins[-2] > max_length + EXTRA_EVAL:
data.bins = data.bins[:-1]
assert data.bins[0] > FLAGS.rx_step
data.forward_max = max(FLAGS.forward_max, data.bins[-1])
nclass = min(FLAGS.niclass, FLAGS.noclass)
data_size = FLAGS.train_data_size if FLAGS.mode == 0 else 1000
# Initialize data for each task.
tasks = FLAGS.task.split("-")
for t in tasks:
for l in xrange(max_length + EXTRA_EVAL - 1):
data.init_data(t, l, data_size, nclass)
data.init_data(t, data.bins[-2], data_size, nclass)
data.init_data(t, data.bins[-1], data_size, nclass)
end_size = 4 * 1024 if FLAGS.mode > 0 else 1024
data.init_data(t, data.forward_max, end_size, nclass)
# Print out parameters.
curriculum = FLAGS.curriculum_bound
msg1 = ("layers %d kw %d h %d kh %d relax %d batch %d noise %.2f task %s" %
(FLAGS.nconvs, FLAGS.kw, FLAGS.height, FLAGS.kh, FLAGS.rx_step,
FLAGS.batch_size, FLAGS.grad_noise_scale, FLAGS.task))
msg2 = "data %d %s" % (FLAGS.train_data_size, msg1)
msg3 = (
"cut %.2f pull %.3f lr %.2f iw %.2f cr %.2f nm %d d%.4f gn %.2f %s" %
(FLAGS.cutoff, FLAGS.pull_incr, FLAGS.lr, FLAGS.init_weight,
curriculum, FLAGS.nmaps, FLAGS.dropout, FLAGS.max_grad_norm, msg2))
data.print_out(msg3)
# Create checkpoint directory if it does not exist.
checkpoint_dir = os.path.join(
FLAGS.train_dir,
"neural_gpu%s" % ("" if FLAGS.jobid < 0 else str(FLAGS.jobid)))
if not gfile.IsDirectory(checkpoint_dir):
data.print_out("Creating checkpoint directory %s." % checkpoint_dir)
gfile.MkDir(checkpoint_dir)
# Create model and initialize it.
tf.get_variable_scope().set_initializer(
tf.uniform_unit_scaling_initializer(factor=1.8 * FLAGS.init_weight))
model = neural_gpu.NeuralGPU(FLAGS.nmaps, FLAGS.nmaps, FLAGS.niclass,
FLAGS.noclass, FLAGS.dropout, FLAGS.rx_step,
FLAGS.max_grad_norm, FLAGS.cutoff,
FLAGS.nconvs, FLAGS.kw, FLAGS.kh,
FLAGS.height, FLAGS.mode, FLAGS.lr,
FLAGS.pull, FLAGS.pull_incr, min_length + 3)
data.print_out("Created model.")
sess.run(tf.initialize_all_variables())
data.print_out("Initialized variables.")
# Load model from parameters if a checkpoint exists.
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
data.print_out("Reading model parameters from %s" %
ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
# Check if there are ensemble models and get their checkpoints.
ensemble = []
ensemble_dir_list = [d for d in FLAGS.ensemble.split(",") if d]
for ensemble_dir in ensemble_dir_list:
ckpt = tf.train.get_checkpoint_state(ensemble_dir)
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
data.print_out("Found ensemble model %s" %
ckpt.model_checkpoint_path)
ensemble.append(ckpt.model_checkpoint_path)
# Return the model and needed variables.
return (model, min_length, max_length, checkpoint_dir, curriculum,
ensemble)
def evaluate():
"""Evaluate an existing model."""
batch_size = FLAGS.batch_size * FLAGS.num_gpus
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
(model, beam_model, _, _, _,
(_, dev_set, en_vocab_path, fr_vocab_path), _, sess) = initialize(sess)
for p in FLAGS.problem.split("-"):
for bin_id in xrange(len(data.bins)):
if (FLAGS.task >= 0 and bin_id > 4) or (FLAGS.nprint == 0 and
bin_id > 8 and p == "wmt"):
break
single_test(bin_id, model, sess, FLAGS.nprint, batch_size, dev_set, p,
beam_model=beam_model)
path = FLAGS.test_file_prefix
xid = "" if FLAGS.task < 0 else ("%.4d" % (FLAGS.task+FLAGS.decode_offset))
en_path, fr_path = path + ".en" + xid, path + ".fr" + xid
# Evaluate the test file if they exist.
if path and tf.gfile.Exists(en_path) and tf.gfile.Exists(fr_path):
data.print_out("Translating test set %s" % en_path)
# Read lines.
en_lines, fr_lines = [], []
with tf.gfile.GFile(en_path, mode="r") as f:
for line in f:
en_lines.append(line.strip())
with tf.gfile.GFile(fr_path, mode="r") as f:
for line in f:
fr_lines.append(line.strip())
# Tokenize and convert to ids.
en_vocab, _ = wmt.initialize_vocabulary(en_vocab_path)
_, rev_fr_vocab = wmt.initialize_vocabulary(fr_vocab_path)
if FLAGS.simple_tokenizer:
en_ids = [wmt.sentence_to_token_ids(
l, en_vocab, tokenizer=wmt.space_tokenizer,
normalize_digits=FLAGS.normalize_digits)
for l in en_lines]
else:
en_ids = [wmt.sentence_to_token_ids(l, en_vocab) for l in en_lines]
# Translate.
results = []
for idx, token_ids in enumerate(en_ids):
if idx % 5 == 0:
data.print_out("Translating example %d of %d." % (idx, len(en_ids)))
# Which bucket does it belong to?
buckets = [b for b in xrange(len(data.bins))
if data.bins[b] >= len(token_ids)]
if buckets:
result, result_cost = [], 100000000.0
for bucket_id in buckets:
if data.bins[bucket_id] > MAXLEN_F * len(token_ids) + EVAL_LEN_INCR:
break
# Get a 1-element batch to feed the sentence to the model.
used_batch_size = 1 # batch_size
inp, target = data.get_batch(
bucket_id, used_batch_size, None, FLAGS.height,
preset=([token_ids], [[]]))
loss, output_logits, _, _ = model.step(
sess, inp, target, None, beam_size=FLAGS.beam_size)
outputs = [int(o[0]) for o in output_logits]
loss = loss[0] - (data.bins[bucket_id] * FLAGS.length_norm)
if FLAGS.simple_tokenizer:
cur_out = outputs
if wmt.EOS_ID in cur_out:
cur_out = cur_out[:cur_out.index(wmt.EOS_ID)]
res_tags = [rev_fr_vocab[o] for o in cur_out]
bad_words, bad_brack = wmt.parse_constraints(token_ids, res_tags)
loss += 1000.0 * bad_words + 100.0 * bad_brack
# print (bucket_id, loss)
if loss < result_cost:
result = outputs
result_cost = loss
final = linearize(result, rev_fr_vocab)
results.append("%s\t%s\n" % (final, fr_lines[idx]))
# print result_cost
sys.stderr.write(results[-1])
sys.stderr.flush()
else:
sys.stderr.write("TOOO_LONG\t%s\n" % fr_lines[idx])
sys.stderr.flush()
if xid:
decode_suffix = "beam%dln%dn" % (FLAGS.beam_size,
int(100 * FLAGS.length_norm))
with tf.gfile.GFile(path + ".res" + decode_suffix + xid, mode="w") as f:
for line in results:
f.write(line)
