def __init__(self, nmt_specs, decoder_args):
"""Set up the NMT model used by the decoder.
Args:
nmt_specs (list): List of tuples of which the first element
is the path to the NMT model (.npz) file
and the second is the NMT configuration
decoder_args (object): Decoder configuration passed through
from configuration API.
"""
super(BlocksNMTEnsembleVanillaDecoder, self).__init__(decoder_args)
self.n_networks = len(nmt_specs)
if self.n_networks < 2:
logging.fatal("The NMT ensemble vanilla decoder needs at "
"least two NMT systems")
global_config = nmt_specs[0][1]
self.src_vocab_size = global_config['src_vocab_size']
self.beam_size = global_config['beam_size']
self.src_sparse_feat_map = global_config['src_sparse_feat_map'] \
if global_config['src_sparse_feat_map'] else FlatSparseFeatMap()
if global_config['trg_sparse_feat_map']:
logging.fatal("Using sparse feature maps on the target size is "
"currently not supported by the ensemble vanilla "
"decoder")
self.trg_sparse_feat_map = FlatSparseFeatMap()
self.set_up_decoder(nmt_specs)
self.src_eos = self.src_sparse_feat_map.word2dense(utils.EOS_ID)
def set_up_predictor(self, nmt_model_path):
"""Initializes the predictor with the given NMT model. Code
following ``blocks.machine_translation.main``.
"""
self.src_vocab_size = self.config['src_vocab_size']
self.trgt_vocab_size = self.config['trg_vocab_size']
self.nmt_model = NMTModel(self.config)
self.nmt_model.set_up()
loader = LoadNMTUtils(nmt_model_path,
self.config['saveto'],
self.nmt_model.search_model)
loader.load_weights()
self.best_models = []
self.val_bleu_curve = []
self.src_sparse_feat_map = self.config['src_sparse_feat_map'] \
if self.config['src_sparse_feat_map'] else FlatSparseFeatMap()
if self.config['trg_sparse_feat_map']:
self.trg_sparse_feat_map = self.config['trg_sparse_feat_map']
self.search_algorithm = MyopticSparseSearch(
samples=self.nmt_model.samples,
trg_sparse_feat_map=self.trg_sparse_feat_map)
else:
self.trg_sparse_feat_map = FlatSparseFeatMap()
self.search_algorithm = MyopticSearch(samples=self.nmt_model.samples)
self.search_algorithm.compile()
def __init__(self,
src_file,
trgt_file,
src_vocab_size,
trgt_vocab_size,
preprocess=None,
src_sparse_feat_map=None,
trg_sparse_feat_map=None):
"""Constructor like for ``TextFile``"""
global _current_src_sparse_feat_map, _current_trg_sparse_feat_map
super(ParallelTextFile, self).__init__()
self.src_file = src_file
self.trgt_file = trgt_file
self.src_vocab_size = src_vocab_size
self.trgt_vocab_size = trgt_vocab_size
self.preprocess = preprocess
with open(self.src_file) as f:
self.src_sentences = f.readlines()
with open(self.trgt_file) as f:
self.trgt_sentences = f.readlines()
self.num_examples = len(self.src_sentences)
if self.num_examples != len(self.trgt_sentences):
raise ValueError
self.src_sparse_feat_map = src_sparse_feat_map if src_sparse_feat_map \
else FlatSparseFeatMap()
self.trg_sparse_feat_map = trg_sparse_feat_map if trg_sparse_feat_map \
else FlatSparseFeatMap()
# Pickle hacks
_current_src_sparse_feat_map = self.src_sparse_feat_map
_current_trg_sparse_feat_map = self.trg_sparse_feat_map
def __init__(self,
parallel_sources,
src_vocab_size=30000,
trg_vocab_size=30000,
src_sparse_feat_map=None,
trg_sparse_feat_map=None):
"""Creates a new ``Dataset`` which allows switching between
multiple parallel sources. The first source is activated per
default
Args:
parallel_sources (list): List of ``ParallelSource`` objects
src_vocab_size (int): Size of source vocabulary
trg_vocab_size (int): Size of target vocabulary
src_sparse_feat_map (SparseFeatMap): Map between words and
their features if you
use sparse vectors.
trg_sparse_feat_map (SparseFeatMap): Map between words and
their features if you
use sparse vectors.
Raises:
IndexError. If the list is empty
"""
super(ParallelSourceSwitchDataset, self).__init__()
global _current_parallel_sources
global _current_src_sparse_feat_map, _current_trg_sparse_feat_map
if not parallel_sources:
raise ValueError
self.parallel_sources = parallel_sources
self.src_vocab_size = src_vocab_size
self.trg_vocab_size = trg_vocab_size
self.active_idx = 0
self.src_sparse_feat_map = src_sparse_feat_map if src_sparse_feat_map \
else FlatSparseFeatMap()
self.trg_sparse_feat_map = trg_sparse_feat_map if trg_sparse_feat_map \
else FlatSparseFeatMap()
# pickle hacks
_current_parallel_sources = parallel_sources
_current_src_sparse_feat_map = self.src_sparse_feat_map
_current_trg_sparse_feat_map = self.trg_sparse_feat_map
def set_up_decoder(self, nmt_model_path):
"""This method uses the NMT configuration in ``self.config`` to
initialize the NMT model. This method basically corresponds to
``blocks.machine_translation.main``.
Args:
nmt_model_path (string): Path to the NMT model file (.npz)
"""
self.nmt_model = NMTModel(self.config)
self.nmt_model.set_up()
loader = LoadNMTUtils(nmt_model_path, self.config['saveto'],
self.nmt_model.search_model)
loader.load_weights()
self.src_sparse_feat_map = self.config['src_sparse_feat_map'] \
if self.config['src_sparse_feat_map'] else FlatSparseFeatMap()
if self.config['trg_sparse_feat_map']:
self.trg_sparse_feat_map = self.config['trg_sparse_feat_map']
self.beam_search = SparseBeamSearch(
samples=self.nmt_model.samples,
trg_sparse_feat_map=self.trg_sparse_feat_map)
else:
self.trg_sparse_feat_map = FlatSparseFeatMap()
self.beam_search = BeamSearch(samples=self.nmt_model.samples)
class BlocksUnboundedNMTPredictor(BlocksNMTPredictor,
UnboundedVocabularyPredictor):
"""This is a version of the NMT predictor which assumes an
unbounded vocabulary. Therefore, this predictor can only be used
when other predictors (like fst) define the words to score. Using
this predictor is mandatory when a target sparse feature map is
provided.
"""
def __init__(self, nmt_model_path, gnmt_beta, config):
"""Creates a new NMT predictor with unbounded vocabulary.
Args:
nmt_model_path (string): Path to the NMT model file (.npz)
config (dict): NMT configuration,
"""
super(BlocksUnboundedNMTPredictor, self).__init__(nmt_model_path,
gnmt_beta,
False,
config)
def set_up_predictor(self, nmt_model_path):
"""Initializes the predictor with the given NMT model. Code
following ``blocks.machine_translation.main``.
"""
self.src_vocab_size = self.config['src_vocab_size']
self.trgt_vocab_size = self.config['trg_vocab_size']
self.nmt_model = NMTModel(self.config)
self.nmt_model.set_up()
loader = LoadNMTUtils(nmt_model_path,
self.config['saveto'],
self.nmt_model.search_model)
loader.load_weights()
self.best_models = []
self.val_bleu_curve = []
self.src_sparse_feat_map = self.config['src_sparse_feat_map'] \
if self.config['src_sparse_feat_map'] else FlatSparseFeatMap()
if self.config['trg_sparse_feat_map']:
self.trg_sparse_feat_map = self.config['trg_sparse_feat_map']
self.search_algorithm = MyopticSparseSearch(
samples=self.nmt_model.samples,
trg_sparse_feat_map=self.trg_sparse_feat_map)
else:
self.trg_sparse_feat_map = FlatSparseFeatMap()
self.search_algorithm = MyopticSearch(samples=self.nmt_model.samples)
self.search_algorithm.compile()
def predict_next(self, words):
"""Uses cache or runs the decoder network to get the
distribution over the next target words.
Returns:
np array. Full distribution over the entire NMT vocabulary
for the next target token.
"""
logprobs = self.search_algorithm.compute_logprobs(self.contexts,
self.states)
if self.trg_sparse_feat_map.dim > 1:
return {w: -sparse.dense_euclidean2(
logprobs[0],
self.trg_sparse_feat_map.word2dense(w))
for w in words}
else:
# logprobs are negative log probs, i.e. greater than 0
posterior = np.multiply(logprobs[0], -1.0)
return {w: posterior[w] for w in words}
def get_unk_probability(self, posterior):
"""Returns negative inf as this is a unbounded predictor. """
return NEG_INF
def consume(self, word):
"""Feeds back ``word`` to the decoder network. This includes
embedding of ``word``, running the attention network and update
the recurrent decoder layer.
"""
if word >= self.trgt_vocab_size:
word = utils.UNK_ID
self.consumed.append(word)
self.states.update(self.search_algorithm.compute_next_states(
self.contexts,
self.states,
[self.trg_sparse_feat_map.word2dense(word)]))
def is_equal(self, state1, state2):
"""Returns true if the history is the same """
_,consumed1 = state1
_,consumed2 = state2
return consumed1 == consumed2
def __init__(self,
source_sentence,
samples,
model,
data_stream,
config,
n_best=1,
track_n_models=1,
normalize=True,
store_full_main_loop=False,
**kwargs):
"""Creates a new extension which adds model selection based on
the accuracy score to the training main loop.
Args:
source_sentence (Variable): Input variable to the sampling
computation graph
samples (Variable): Samples variable of the CG
model (NMTModel): See the model module
data_stream (DataStream): Data stream to the development
set
config (dict): NMT configuration
n_best (int): beam size
track_n_models (int): Number of n-best models for which to
create checkpoints.
normalize (boolean): Enables length normalization
store_full_main_loop (boolean): Stores the iteration state
in the old style of
Blocks 0.1. Not recommended
"""
super(AccValidator, self).__init__(**kwargs)
self.store_full_main_loop = store_full_main_loop
self.source_sentence = source_sentence
self.samples = samples
self.model = model
self.data_stream = data_stream
self.config = config
self.n_best = n_best
self.track_n_models = track_n_models
self.normalize = normalize
self.best_models = []
self.val_bleu_curve = []
self.src_sparse_feat_map = config['src_sparse_feat_map'] if config['src_sparse_feat_map'] \
else FlatSparseFeatMap()
if config['trg_sparse_feat_map']:
self.trg_sparse_feat_map = config['trg_sparse_feat_map']
self.beam_search = SparseBeamSearch(
samples=samples,
trg_sparse_feat_map=self.trg_sparse_feat_map)
else:
self.trg_sparse_feat_map = FlatSparseFeatMap()
self.beam_search = BeamSearch(samples=samples)
# Create saving directory if it does not exist
if not os.path.exists(self.config['saveto']):
os.makedirs(self.config['saveto'])
if self.config['reload']:
try:
bleu_score = numpy.load(os.path.join(self.config['saveto'],
'val_bleu_scores.npz'))
self.val_bleu_curve = bleu_score['bleu_scores'].tolist()
# Track n best previous bleu scores
for i, bleu in enumerate(
sorted(self.val_bleu_curve, reverse=True)):
if i < self.track_n_models:
self.best_models.append(ModelInfo(bleu))
logging.info("BleuScores Reloaded")
except:
logging.info("BleuScores not Found")
self.verbose = self.config.get('val_set_out', None)
utils.load_trg_wmap(self.config['trg_wmap'])
self.trg_wmap = utils.trg_wmap
