def _make_basic_master_spec():
"""Constructs a simple spec.
Modified version of nlp/saft/opensource/dragnn/tools/parser_trainer.py
Returns:
spec_pb2.MasterSpec instance.
"""
# Construct the "lookahead" ComponentSpec. This is a simple right-to-left RNN
# sequence model, which encodes the context to the right of each token. It has
# no loss except for the downstream components.
lookahead = spec_builder.ComponentSpecBuilder('lookahead')
lookahead.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256')
lookahead.set_transition_system(name='shift-only', left_to_right='true')
lookahead.add_fixed_feature(name='words',
fml='input.word',
embedding_dim=64)
lookahead.add_rnn_link(embedding_dim=-1)
# Construct the ComponentSpec for parsing.
parser = spec_builder.ComponentSpecBuilder('parser')
parser.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256')
parser.set_transition_system(name='arc-standard')
parser.add_token_link(source=lookahead,
fml='input.focus',
embedding_dim=32)
master_spec = spec_pb2.MasterSpec()
master_spec.component.extend([lookahead.spec, parser.spec])
return master_spec
def testFillsTaggerTransitions(self):
lexicon_dir = tempfile.mkdtemp()
def write_lines(filename, lines):
with open(os.path.join(lexicon_dir, filename), 'w') as f:
f.write(''.join('{}\n'.format(line) for line in lines))
# Label map is required, even though it isn't used
write_lines('label-map', ['0'])
write_lines('word-map', ['2', 'miranda 1', 'rights 1'])
write_lines('tag-map', ['2', 'NN 1', 'NNP 1'])
write_lines('tag-to-category', ['NN\tNOUN', 'NNP\tNOUN'])
tagger = spec_builder.ComponentSpecBuilder('tagger')
tagger.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256')
tagger.set_transition_system(name='tagger')
tagger.add_fixed_feature(name='words',
fml='input.word',
embedding_dim=64)
tagger.add_rnn_link(embedding_dim=-1)
tagger.fill_from_resources(lexicon_dir)
fixed_feature, = tagger.spec.fixed_feature
linked_feature, = tagger.spec.linked_feature
self.assertEqual(fixed_feature.vocabulary_size, 5)
self.assertEqual(fixed_feature.size, 1)
self.assertEqual(fixed_feature.size, 1)
self.assertEqual(linked_feature.size, 1)
self.assertEqual(tagger.spec.num_actions, 2)
def testComponentSpecBuilderEmpty(self):
builder = spec_builder.ComponentSpecBuilder('test')
self.assertSpecEqual(r"""
name: "test"
backend { registered_name: "SyntaxNetComponent" }
component_builder { registered_name: "DynamicComponentBuilder" }
""", builder.spec)
def testComponentSpecBuilderLinkedFeature(self):
builder1 = spec_builder.ComponentSpecBuilder('test1')
builder1.set_transition_system('shift-only')
builder1.add_fixed_feature(name='words', fml='input.word', embedding_dim=16)
builder2 = spec_builder.ComponentSpecBuilder('test2')
builder2.set_network_unit('IdentityNetwork')
builder2.set_transition_system('tagger')
builder2.add_token_link(
source=builder1,
source_layer='words',
fml='input.focus',
embedding_dim=-1)
self.assertSpecEqual(r"""
name: "test2"
linked_feature { name: "test1" source_component: "test1" source_layer: "words"
source_translator: "identity" fml: "input.focus"
embedding_dim: -1 }
backend { registered_name: "SyntaxNetComponent" }
component_builder { registered_name: "DynamicComponentBuilder" }
network_unit { registered_name: "IdentityNetwork" }
transition_system { registered_name: "tagger" }
""", builder2.spec)
def testComponentSpecBuilderFixedFeature(self):
builder = spec_builder.ComponentSpecBuilder('test')
builder.set_network_unit('FeedForwardNetwork', hidden_layer_sizes='64,64')
builder.set_transition_system('shift-only')
builder.add_fixed_feature(name='words', fml='input.word', embedding_dim=16)
self.assertSpecEqual(r"""
name: "test"
fixed_feature { name: "words" fml: "input.word" embedding_dim: 16 }
backend { registered_name: "SyntaxNetComponent" }
component_builder { registered_name: "DynamicComponentBuilder" }
network_unit { registered_name: "FeedForwardNetwork"
parameters { key: "hidden_layer_sizes" value: "64,64" } }
transition_system { registered_name: "shift-only" }
""", builder.spec)
def main(argv):
del argv # unused
# Constructs lexical resources for SyntaxNet in the given resource path, from
# the training data.
lexicon.build_lexicon(lexicon_dir,
training_sentence,
training_corpus_format='sentence-prototext')
# Construct the ComponentSpec for tagging. This is a simple left-to-right RNN
# sequence tagger.
tagger = spec_builder.ComponentSpecBuilder('tagger')
tagger.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256')
tagger.set_transition_system(name='tagger')
tagger.add_fixed_feature(name='words', fml='input.word', embedding_dim=64)
tagger.add_rnn_link(embedding_dim=-1)
tagger.fill_from_resources(lexicon_dir)
master_spec = spec_pb2.MasterSpec()
master_spec.component.extend([tagger.spec])
hyperparam_config = spec_pb2.GridPoint()
# Build the TensorFlow graph.
graph = tf.Graph()
with graph.as_default():
builder = graph_builder.MasterBuilder(master_spec, hyperparam_config)
target = spec_pb2.TrainTarget()
target.name = 'all'
target.unroll_using_oracle.extend([True])
dry_run = builder.add_training_from_config(target, trace_only=True)
# Read in serialized protos from training data.
sentence = sentence_pb2.Sentence()
text_format.Merge(open(training_sentence).read(), sentence)
training_set = [sentence.SerializeToString()]
with tf.Session(graph=graph) as sess:
# Make sure to re-initialize all underlying state.
sess.run(tf.initialize_all_variables())
traces = sess.run(dry_run['traces'],
feed_dict={dry_run['input_batch']: training_set})
with open('dragnn_tutorial_1.html', 'w') as f:
f.write(
visualization.trace_html(traces[0],
height='300px').encode('utf-8'))
def main(unused_argv):
logging.set_verbosity(logging.INFO)
if not gfile.IsDirectory(FLAGS.resource_path):
gfile.MakeDirs(FLAGS.resource_path)
# Constructs lexical resources for SyntaxNet in the given resource path, from
# the training data.
if FLAGS.compute_lexicon:
logging.info('Computing lexicon...')
lexicon.build_lexicon(FLAGS.resource_path, FLAGS.training_corpus_path)
# Construct the "lookahead" ComponentSpec. This is a simple right-to-left RNN
# sequence model, which encodes the context to the right of each token. It has
# no loss except for the downstream components.
lookahead = spec_builder.ComponentSpecBuilder('lookahead')
lookahead.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
lookahead.set_transition_system(name='shift-only', left_to_right='false')
lookahead.add_fixed_feature(name='char',
fml='input(-1).char input.char input(1).char',
embedding_dim=32)
lookahead.add_fixed_feature(name='char-bigram',
fml='input.char-bigram',
embedding_dim=32)
lookahead.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
# Construct the ComponentSpec for segmentation.
segmenter = spec_builder.ComponentSpecBuilder('segmenter')
segmenter.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='128')
segmenter.set_transition_system(name='binary-segment-transitions')
segmenter.add_token_link(source=lookahead,
fml='input.focus stack.focus',
embedding_dim=64)
segmenter.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
# Build and write master_spec.
master_spec = spec_pb2.MasterSpec()
master_spec.component.extend([lookahead.spec, segmenter.spec])
logging.info('Constructed master spec: %s', str(master_spec))
with gfile.GFile(FLAGS.resource_path + '/master_spec', 'w') as f:
f.write(str(master_spec).encode('utf-8'))
hyperparam_config = spec_pb2.GridPoint()
try:
text_format.Parse(FLAGS.hyperparams, hyperparam_config)
except text_format.ParseError:
text_format.Parse(base64.b64decode(FLAGS.hyperparams),
hyperparam_config)
# Build the TensorFlow graph.
graph = tf.Graph()
with graph.as_default():
builder = graph_builder.MasterBuilder(master_spec, hyperparam_config)
component_targets = spec_builder.default_targets_from_spec(master_spec)
trainers = [
builder.add_training_from_config(target)
for target in component_targets
]
assert len(trainers) == 1
annotator = builder.add_annotation()
builder.add_saver()
# Read in serialized protos from training data.
training_set = ConllSentenceReader(FLAGS.training_corpus_path,
projectivize=False).corpus()
dev_set = ConllSentenceReader(FLAGS.dev_corpus_path,
projectivize=False).corpus()
# Convert word-based docs to char-based documents for segmentation training
# and evaluation.
with tf.Session(graph=tf.Graph()) as tmp_session:
char_training_set_op = gen_parser_ops.segmenter_training_data_constructor(
training_set)
char_dev_set_op = gen_parser_ops.char_token_generator(dev_set)
char_training_set = tmp_session.run(char_training_set_op)
char_dev_set = tmp_session.run(char_dev_set_op)
# Ready to train!
logging.info('Training on %d sentences.', len(training_set))
logging.info('Tuning on %d sentences.', len(dev_set))
pretrain_steps = [0]
train_steps = [FLAGS.num_epochs * len(training_set)]
tf.logging.info('Creating TensorFlow checkpoint dir...')
gfile.MakeDirs(os.path.dirname(FLAGS.checkpoint_filename))
summary_writer = trainer_lib.get_summary_writer(FLAGS.tensorboard_dir)
with tf.Session(FLAGS.tf_master, graph=graph) as sess:
# Make sure to re-initialize all underlying state.
sess.run(tf.global_variables_initializer())
trainer_lib.run_training(
sess, trainers, annotator, evaluation.segmentation_summaries,
pretrain_steps, train_steps, char_training_set, char_dev_set,
dev_set, FLAGS.batch_size, summary_writer, FLAGS.report_every,
builder.saver, FLAGS.checkpoint_filename)
def build_master_spec():
'''
# Left-to-right, character-based LSTM.
char2word = spec_builder.ComponentSpecBuilder('char_lstm')
char2word.set_network_unit(
name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
char2word.set_transition_system(name='char-shift-only', left_to_right='true')
char2word.add_fixed_feature(name='chars', fml='char-input.text-char',
embedding_dim=16)
# Lookahead LSTM reads right-to-left to represent the rightmost context of the
# words. It gets word embeddings from the char model.
lookahead = spec_builder.ComponentSpecBuilder('lookahead')
lookahead.set_network_unit(
name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
lookahead.set_transition_system(name='shift-only', left_to_right='false')
lookahead.add_link(source=char2word, fml='input.last-char-focus',
embedding_dim=64)
'''
# Construct the 'lookahead' ComponentSpec. This is a simple right-to-left RNN
# sequence model, which encodes the context to the right of each token. It has
# no loss except for the downstream components.
lookahead = spec_builder.ComponentSpecBuilder('lookahead')
lookahead.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256')
lookahead.set_transition_system(name='shift-only', left_to_right='true')
lookahead.add_fixed_feature(name='words',
fml='input.word',
embedding_dim=64)
lookahead.add_rnn_link(embedding_dim=-1)
# Construct the tagger. This is a simple left-to-right LSTM sequence tagger.
tagger = spec_builder.ComponentSpecBuilder('tagger')
tagger.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
tagger.set_transition_system(name='tagger')
tagger.add_token_link(source=lookahead,
fml='input.focus',
embedding_dim=64)
# Construct the parser.
parser = spec_builder.ComponentSpecBuilder('parser')
parser.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256',
layer_norm_hidden='true')
parser.set_transition_system(name='arc-standard')
parser.add_token_link(source=lookahead,
fml='input.focus',
embedding_dim=64)
parser.add_token_link(source=tagger,
fml='input.focus stack.focus stack(1).focus',
embedding_dim=64)
# Add discrete features of the predicted parse tree so far, like in Parsey
# McParseface.
parser.add_fixed_feature(
name='labels',
embedding_dim=16,
fml=' '.join([
'stack.child(1).label', 'stack.child(1).sibling(-1).label',
'stack.child(-1).label', 'stack.child(-1).sibling(1).label',
'stack(1).child(1).label', 'stack(1).child(1).sibling(-1).label',
'stack(1).child(-1).label', 'stack(1).child(-1).sibling(1).label',
'stack.child(2).label', 'stack.child(-2).label',
'stack(1).child(2).label', 'stack(1).child(-2).label'
]))
# Recurrent connection for the arc-standard parser. For both tokens on the
# stack, we connect to the last time step to either SHIFT or REDUCE that
# token. This allows the parser to build up compositional representations of
# phrases.
parser.add_link(
source=parser, # recurrent connection
name='rnn-stack', # unique identifier
fml='stack.focus stack(1).focus', # look for both stack tokens
source_translator='shift-reduce-step', # maps token indices -> step
embedding_dim=64) # project down to 64 dims
master_spec = spec_pb2.MasterSpec()
'''
master_spec.component.extend(
[char2word.spec, lookahead.spec, tagger.spec, parser.spec])
'''
master_spec.component.extend([lookahead.spec, tagger.spec, parser.spec])
return master_spec
def main(unused_argv):
logging.set_verbosity(logging.INFO)
if not gfile.IsDirectory(FLAGS.resource_path):
gfile.MakeDirs(FLAGS.resource_path)
# Constructs lexical resources for SyntaxNet in the given resource path, from
# the training data.
if FLAGS.compute_lexicon:
logging.info('Computing lexicon...')
lexicon.build_lexicon(FLAGS.resource_path, FLAGS.training_corpus_path)
# Construct the "lookahead" ComponentSpec. This is a simple right-to-left RNN
# sequence model, which encodes the context to the right of each token. It has
# no loss except for the downstream components.
char2word = spec_builder.ComponentSpecBuilder('char_lstm')
char2word.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
char2word.set_transition_system(name='char-shift-only',
left_to_right='true')
char2word.add_fixed_feature(name='chars',
fml='char-input.text-char',
embedding_dim=16)
char2word.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
lookahead = spec_builder.ComponentSpecBuilder('lookahead')
lookahead.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
lookahead.set_transition_system(name='shift-only', left_to_right='false')
lookahead.add_link(source=char2word,
fml='input.last-char-focus',
embedding_dim=32)
lookahead.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
# Construct the ComponentSpec for tagging. This is a simple left-to-right RNN
# sequence tagger.
tagger = spec_builder.ComponentSpecBuilder('tagger')
tagger.set_network_unit(name='wrapped_units.LayerNormBasicLSTMNetwork',
hidden_layer_sizes='256')
tagger.set_transition_system(name='tagger')
tagger.add_token_link(source=lookahead,
fml='input.focus',
embedding_dim=32)
tagger.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
# Construct the ComponentSpec for parsing.
parser = spec_builder.ComponentSpecBuilder('parser')
parser.set_network_unit(name='FeedForwardNetwork',
hidden_layer_sizes='256',
layer_norm_hidden='True')
parser.set_transition_system(name='arc-standard')
parser.add_token_link(source=lookahead,
fml='input.focus',
embedding_dim=32)
parser.add_token_link(source=tagger,
fml='input.focus stack.focus stack(1).focus',
embedding_dim=32)
# Recurrent connection for the arc-standard parser. For both tokens on the
# stack, we connect to the last time step to either SHIFT or REDUCE that
# token. This allows the parser to build up compositional representations of
# phrases.
parser.add_link(
source=parser, # recurrent connection
name='rnn-stack', # unique identifier
fml='stack.focus stack(1).focus', # look for both stack tokens
source_translator='shift-reduce-step', # maps token indices -> step
embedding_dim=32) # project down to 32 dims
parser.fill_from_resources(FLAGS.resource_path, FLAGS.tf_master)
master_spec = spec_pb2.MasterSpec()
master_spec.component.extend(
[char2word.spec, lookahead.spec, tagger.spec, parser.spec])
logging.info('Constructed master spec: %s', str(master_spec))
hyperparam_config = spec_pb2.GridPoint()
hyperparam_config.decay_steps = 128000
hyperparam_config.learning_rate = 0.001
hyperparam_config.learning_method = 'adam'
hyperparam_config.adam_beta1 = 0.9
hyperparam_config.adam_beta2 = 0.9
hyperparam_config.adam_eps = 0.0001
hyperparam_config.gradient_clip_norm = 1
hyperparam_config.self_norm_alpha = 1.0
hyperparam_config.use_moving_average = True
hyperparam_config.dropout_rate = 0.7
hyperparam_config.seed = 1
# Build the TensorFlow graph.
graph = tf.Graph()
with graph.as_default():
builder = graph_builder.MasterBuilder(master_spec, hyperparam_config)
component_targets = spec_builder.default_targets_from_spec(master_spec)
trainers = [
builder.add_training_from_config(target)
for target in component_targets
]
assert len(trainers) == 2
annotator = builder.add_annotation()
builder.add_saver()
# Read in serialized protos from training data.
training_set = sentence_io.ConllSentenceReader(
FLAGS.training_corpus_path,
projectivize=FLAGS.projectivize_training_set).corpus()
dev_set = sentence_io.ConllSentenceReader(FLAGS.dev_corpus_path,
projectivize=False).corpus()
# Ready to train!
logging.info('Training on %d sentences.', len(training_set))
logging.info('Tuning on %d sentences.', len(dev_set))
pretrain_steps = [100, 0]
tagger_steps = 1000
train_steps = [tagger_steps, 8 * tagger_steps]
tf.logging.info('Creating TensorFlow checkpoint dir...')
gfile.MakeDirs(os.path.dirname(FLAGS.checkpoint_filename))
summary_writer = trainer_lib.get_summary_writer(FLAGS.tensorboard_dir)
with tf.Session(FLAGS.tf_master, graph=graph) as sess:
# Make sure to re-initialize all underlying state.
sess.run(tf.global_variables_initializer())
trainer_lib.run_training(sess, trainers, annotator,
evaluation.parser_summaries, pretrain_steps,
train_steps, training_set, dev_set, dev_set,
FLAGS.batch_size, summary_writer,
FLAGS.report_every, builder.saver,
FLAGS.checkpoint_filename)
def main(argv):
del argv # unused
# Constructs lexical resources for SyntaxNet in the given resource path, from
# the training data.
lexicon.build_lexicon(
lexicon_dir,
training_sentence,
training_corpus_format='sentence-prototext')
# Construct the ComponentSpec for tagging. This is a simple left-to-right RNN
# sequence tagger.
tagger = spec_builder.ComponentSpecBuilder('tagger')
tagger.set_network_unit(name='FeedForwardNetwork', hidden_layer_sizes='256')
tagger.set_transition_system(name='tagger')
tagger.add_fixed_feature(name='words', fml='input.word', embedding_dim=64)
tagger.add_rnn_link(embedding_dim=-1)
tagger.fill_from_resources(lexicon_dir)
# Construct the ComponentSpec for parsing.
parser = spec_builder.ComponentSpecBuilder('parser')
parser.set_network_unit(
name='FeedForwardNetwork',
hidden_layer_sizes='256',
layer_norm_hidden='True')
parser.set_transition_system(name='arc-standard')
parser.add_token_link(
source=tagger,
fml='input.focus stack.focus stack(1).focus',
embedding_dim=32,
source_layer='logits')
# Recurrent connection for the arc-standard parser. For both tokens on the
# stack, we connect to the last time step to either SHIFT or REDUCE that
# token. This allows the parser to build up compositional representations of
# phrases.
parser.add_link(
source=parser, # recurrent connection
name='rnn-stack', # unique identifier
fml='stack.focus stack(1).focus', # look for both stack tokens
source_translator='shift-reduce-step', # maps token indices -> step
embedding_dim=32) # project down to 32 dims
parser.fill_from_resources(lexicon_dir)
master_spec = spec_pb2.MasterSpec()
master_spec.component.extend([tagger.spec, parser.spec])
hyperparam_config = spec_pb2.GridPoint()
# Build the TensorFlow graph.
graph = tf.Graph()
with graph.as_default():
builder = graph_builder.MasterBuilder(master_spec, hyperparam_config)
target = spec_pb2.TrainTarget()
target.name = 'all'
target.unroll_using_oracle.extend([True, True])
dry_run = builder.add_training_from_config(target, trace_only=True)
# Read in serialized protos from training data.
sentence = sentence_pb2.Sentence()
text_format.Merge(open(training_sentence).read(), sentence)
training_set = [sentence.SerializeToString()]
with tf.Session(graph=graph) as sess:
# Make sure to re-initialize all underlying state.
sess.run(tf.initialize_all_variables())
traces = sess.run(
dry_run['traces'], feed_dict={dry_run['input_batch']: training_set})
with open('dragnn_tutorial_2.html', 'w') as f:
f.write(
visualization.trace_html(
traces[0], height='400px', master_spec=master_spec).encode('utf-8'))
