def sample(args):
numpy.random.seed(args.random_seed)
if args.debug:
theano.config.compute_test_value = 'warn'
else:
theano.config.compute_test_value = 'off'
with h5py.File(args.model_path, 'r') as state:
print("Reading vocabulary from network state.")
sys.stdout.flush()
vocabulary = Vocabulary.from_state(state)
print("Number of words in vocabulary:", vocabulary.num_words())
print("Number of word classes:", vocabulary.num_classes())
print("Building neural network.")
sys.stdout.flush()
architecture = Architecture.from_state(state)
network = Network(vocabulary,
architecture,
mode=Network.Mode(minibatch=False))
print("Restoring neural network state.")
network.set_state(state)
print("Building text sampler.")
sys.stdout.flush()
sampler = TextSampler(network)
sequences = sampler.generate(30, args.num_sentences)
for sequence in sequences:
try:
eos_pos = sequence.index('</s>')
sequence = sequence[:eos_pos + 1]
except:
pass
args.output_file.write(' '.join(sequence) + '\n')
def restoreModel(path):
with h5py.File(path, 'r') as state:
print("Reading vocabulary from network state.")
sys.stdout.flush()
vocabulary = Vocabulary.from_state(state)
print("Number of words in vocabulary:", vocabulary.num_words())
print("Number of words in shortlist:", vocabulary.num_shortlist_words())
print("Number of word classes:", vocabulary.num_classes())
print("Building neural network.")
sys.stdout.flush()
architecture = Architecture.from_state(state)
network = Network(architecture, vocabulary, mode=Network.Mode(minibatch=False))
print("Restoring neural network state.")
network.set_state(state)
return network
def sample(args):
"""A function that performs the "theanolm sample" command.
:type args: argparse.Namespace
:param args: a collection of command line arguments
"""
numpy.random.seed(args.random_seed)
if args.debug:
theano.config.compute_test_value = 'warn'
else:
theano.config.compute_test_value = 'off'
with h5py.File(args.model_path, 'r') as state:
logging.info("Reading vocabulary from network state.")
vocabulary = Vocabulary.from_state(state)
logging.info("Number of words in vocabulary: %d",
vocabulary.num_words())
logging.info("Number of words in shortlist: %d",
vocabulary.num_shortlist_words())
logging.info("Number of word classes: %d", vocabulary.num_classes())
logging.info("Building neural network.")
architecture = Architecture.from_state(state)
default_device = get_default_device(args.default_device)
network = Network(architecture,
vocabulary,
mode=Network.Mode(minibatch=False),
default_device=default_device)
logging.info("Restoring neural network state.")
network.set_state(state)
logging.info("Building text sampler.")
sampler = TextSampler(network)
sequences = sampler.generate(args.sentence_length,
args.num_sentences,
seed_sequence=args.seed_sequence)
for sequence in sequences:
try:
eos_pos = sequence.index('</s>')
sequence = sequence[:eos_pos + 1]
except ValueError:
pass
args.output_file.write(' '.join(sequence) + '\n')
def __init__(self, model_path):
self.model_path = model_path
numpy.random.seed()
theano.config.compute_test_value = 'off'
with h5py.File(model_path, 'r') as self.state:
print("Reading vocabulary from network state.")
#sys.stdout.flush()
self.vocabulary = Vocabulary.from_state(self.state)
print("Number of words in vocabulary:", self.vocabulary.num_words())
print("Number of words in shortlist:", self.vocabulary.num_shortlist_words())
print("Number of word classes:", self.vocabulary.num_classes())
print("Building neural network.")
#sys.stdout.flush()
self.architecture = Architecture.from_state(self.state)
self.network = Network(self.architecture, self.vocabulary, mode=Network.Mode(minibatch=False))
print("Restoring neural network state.")
self.network.set_state(self.state)
print("Building text sampler.")
#sys.stdout.flush()
self.sampler = TextSampler(self.network)
def decode(args):
"""A function that performs the "theanolm decode" command.
:type args: argparse.Namespace
:param args: a collection of command line arguments
"""
log_file = args.log_file
log_level = getattr(logging, args.log_level.upper(), None)
if not isinstance(log_level, int):
print("Invalid logging level requested:", args.log_level)
sys.exit(1)
log_format = '%(asctime)s %(funcName)s: %(message)s'
if args.log_file == '-':
logging.basicConfig(stream=sys.stdout,
format=log_format,
level=log_level)
else:
logging.basicConfig(filename=log_file,
format=log_format,
level=log_level)
if args.debug:
theano.config.compute_test_value = 'warn'
else:
theano.config.compute_test_value = 'off'
theano.config.profile = args.profile
theano.config.profile_memory = args.profile
network = Network.from_file(args.model_path,
mode=Network.Mode(minibatch=False))
log_scale = 1.0 if args.log_base is None else numpy.log(args.log_base)
if args.wi_penalty is None:
wi_penalty = None
else:
wi_penalty = args.wi_penalty * log_scale
if args.unk_penalty is None:
ignore_unk = False
unk_penalty = None
elif args.unk_penalty == 0:
ignore_unk = True
unk_penalty = None
else:
ignore_unk = False
unk_penalty = args.unk_penalty
decoding_options = {
'nnlm_weight': args.nnlm_weight,
'lm_scale': args.lm_scale,
'wi_penalty': wi_penalty,
'ignore_unk': ignore_unk,
'unk_penalty': unk_penalty,
'linear_interpolation': args.linear_interpolation,
'max_tokens_per_node': args.max_tokens_per_node,
'beam': args.beam,
'recombination_order': args.recombination_order
}
logging.debug("DECODING OPTIONS")
for option_name, option_value in decoding_options.items():
logging.debug("%s: %s", option_name, str(option_value))
print("Building word lattice decoder.")
sys.stdout.flush()
decoder = LatticeDecoder(network, decoding_options)
# Combine paths from command line and lattice list.
lattices = args.lattices
if args.lattice_list is not None:
lattices.extend(args.lattice_list.readlines())
lattices = [path.strip() for path in lattices]
# Ignore empty lines in the lattice list.
lattices = [x for x in lattices if x]
# Pick every Ith lattice, if --num-jobs is specified and > 1.
if args.num_jobs < 1:
print("Invalid number of jobs specified:", args.num_jobs)
sys.exit(1)
if (args.job < 0) or (args.job > args.num_jobs - 1):
print("Invalid job specified:", args.job)
sys.exit(1)
lattices = lattices[args.job::args.num_jobs]
file_type = TextFileType('r')
for index, path in enumerate(lattices):
logging.info("Reading word lattice: %s", path)
lattice_file = file_type(path)
lattice = SLFLattice(lattice_file)
if lattice.utterance_id is not None:
utterance_id = lattice.utterance_id
else:
utterance_id = os.path.basename(lattice_file.name)
logging.info("Utterance `%s' -- %d/%d of job %d", utterance_id,
index + 1, len(lattices), args.job)
tokens = decoder.decode(lattice)
for index in range(min(args.n_best, len(tokens))):
line = format_token(tokens[index], utterance_id,
network.vocabulary, log_scale, args.output)
args.output_file.write(line + "\n")
def decode(args):
"""A function that performs the "theanolm decode" command.
:type args: argparse.Namespace
:param args: a collection of command line arguments
"""
log_file = args.log_file
log_level = getattr(logging, args.log_level.upper(), None)
if not isinstance(log_level, int):
print("Invalid logging level requested:", args.log_level,
file=sys.stderr)
sys.exit(1)
log_format = '%(asctime)s %(funcName)s: %(message)s'
if args.log_file == '-':
logging.basicConfig(stream=sys.stdout, format=log_format, level=log_level)
else:
logging.basicConfig(filename=log_file, format=log_format, level=log_level)
if args.debug:
theano.config.compute_test_value = 'warn'
else:
theano.config.compute_test_value = 'off'
theano.config.profile = args.profile
theano.config.profile_memory = args.profile
if (args.lattice_format == 'kaldi') or (args.output == 'kaldi'):
if args.kaldi_vocabulary is None:
print("Kaldi lattice vocabulary is not given.", file=sys.stderr)
sys.exit(1)
default_device = get_default_device(args.default_device)
network = Network.from_file(args.model_path,
mode=Network.Mode(minibatch=False),
default_device=default_device)
log_scale = 1.0 if args.log_base is None else numpy.log(args.log_base)
if (args.log_base is not None) and (args.lattice_format == 'kaldi'):
logging.info("Warning: Kaldi lattice reader doesn't support logarithm "
"base conversion.")
if args.wi_penalty is None:
wi_penalty = None
else:
wi_penalty = args.wi_penalty * log_scale
decoding_options = {
'nnlm_weight': args.nnlm_weight,
'lm_scale': args.lm_scale,
'wi_penalty': wi_penalty,
'unk_penalty': args.unk_penalty,
'use_shortlist': args.shortlist,
'unk_from_lattice': args.unk_from_lattice,
'linear_interpolation': args.linear_interpolation,
'max_tokens_per_node': args.max_tokens_per_node,
'beam': args.beam,
'recombination_order': args.recombination_order,
'prune_relative': args.prune_relative,
'abs_min_max_tokens': args.abs_min_max_tokens,
'abs_min_beam': args.abs_min_beam
}
logging.debug("DECODING OPTIONS")
for option_name, option_value in decoding_options.items():
logging.debug("%s: %s", option_name, str(option_value))
logging.info("Building word lattice decoder.")
decoder = LatticeDecoder(network, decoding_options)
batch = LatticeBatch(args.lattices, args.lattice_list, args.lattice_format,
args.kaldi_vocabulary, args.num_jobs, args.job)
for lattice_number, lattice in enumerate(batch):
if lattice.utterance_id is None:
lattice.utterance_id = str(lattice_number)
logging.info("Utterance `%s´ -- %d of job %d",
lattice.utterance_id,
lattice_number + 1,
args.job)
log_free_mem()
final_tokens, recomb_tokens = decoder.decode(lattice)
if (args.output == "slf") or (args.output == "kaldi"):
rescored_lattice = RescoredLattice(lattice,
final_tokens,
recomb_tokens,
network.vocabulary)
rescored_lattice.lm_scale = args.lm_scale
rescored_lattice.wi_penalty = args.wi_penalty
if args.output == "slf":
rescored_lattice.write_slf(args.output_file)
else:
assert args.output == "kaldi"
rescored_lattice.write_kaldi(args.output_file,
batch.kaldi_word_to_id)
else:
for token in final_tokens[:min(args.n_best, len(final_tokens))]:
line = format_token(token,
lattice.utterance_id,
network.vocabulary,
log_scale,
args.output)
args.output_file.write(line + "\n")
gc.collect()
