def __init__(
self,
input_path=None,
rating_scale=5,
verbose=False,
):
self.uid_map = load_dict(os.path.join(input_path, "uid_map"), sep=",")
self.iid_map = load_dict(os.path.join(input_path, "iid_map"), sep=",")
self.aspect_id_map = load_dict(os.path.join(input_path,
"aspect_id_map"),
sep=",")
self.opinion_id_map = load_dict(os.path.join(input_path,
"opinion_id_map"),
sep=",")
self.U = np.load(os.path.join(input_path, "U.npy"))
self.I = np.load(os.path.join(input_path, "I.npy"))
self.A = np.load(os.path.join(input_path, "A.npy"))
self.O = np.load(os.path.join(input_path, "O.npy"))
self.G1 = np.load(os.path.join(input_path, "G1.npy"))
self.G2 = np.load(os.path.join(input_path, "G2.npy"))
self.G3 = np.load(os.path.join(input_path, "G3.npy"))
self.rating_scale = rating_scale
self.id2aspect = {v: k for k, v in self.aspect_id_map.items()}
self.verbose = verbose
if self.verbose:
print("Load MTER from %s" % input_path)
def __init__(
self,
input_path=None,
alpha=0.85,
num_most_cared_aspects=15,
rating_scale=5,
verbose=False,
):
self.uid_map = load_dict(os.path.join(input_path, "uid_map"), sep=",")
self.iid_map = load_dict(os.path.join(input_path, "iid_map"), sep=",")
self.aspect_id_map = load_dict(os.path.join(input_path,
"aspect_id_map"),
sep=",")
self.U1 = np.load(os.path.join(input_path, "U1.npy"))
self.U2 = np.load(os.path.join(input_path, "U2.npy"))
self.V = np.load(os.path.join(input_path, "V.npy"))
self.H1 = np.load(os.path.join(input_path, "H1.npy"))
self.H2 = np.load(os.path.join(input_path, "H2.npy"))
self.alpha = alpha
self.n_cared_aspects = num_most_cared_aspects
self.rating_scale = rating_scale
self.id2aspect = {v: k for k, v in self.aspect_id_map.items()}
self.verbose = verbose
if self.verbose:
print("Load EFM from %s" % input_path)
def __init__(self, source, target,
source_dict, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True):
if shuffle_each_epoch:
shuffle.main([source, target])
self.source = fopen(source+'.shuf', 'r')
self.target = fopen(target+'.shuf', 'r')
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dict = load_dict(source_dict)
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * 20
self.end_of_data = False
def __init__(self,
source,
target,
source_dict,
target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True):
if shuffle_each_epoch:
shuffle.main([source, target])
self.source = fopen(source + '.shuf', 'r')
self.target = fopen(target + '.shuf', 'r')
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dict = load_dict(source_dict)
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * 20
self.end_of_data = False
def main(models,
source_file,
nbest_file,
saveto,
b=80,
normalize=False,
verbose=False,
alignweights=False):
# load model model_options
options = []
for model in args.models:
try:
with open('%s.json' % model, 'rb') as f:
options.append(json.load(f))
except:
with open('%s.pkl' % model, 'rb') as f:
options.append(pkl.load(f))
#hacks for using old models with missing options
if not 'dropout_embedding' in options[-1]:
options[-1]['dropout_embedding'] = 0
if not 'dropout_hidden' in options[-1]:
options[-1]['dropout_hidden'] = 0
if not 'dropout_source' in options[-1]:
options[-1]['dropout_source'] = 0
if not 'dropout_target' in options[-1]:
options[-1]['dropout_target'] = 0
dictionaries = options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
dictionary_target = dictionaries[-1]
# load source dictionary and invert
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
if options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= options[0]['n_words_src']:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
# load target dictionary and invert
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
rescore_model(source_file, nbest_file, saveto, models, options, b,
normalize, verbose, alignweights)
def __init__(self, source, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
indomain_source='', indomain_target='',
interpolation_rate=0.1,
maxibatch_size=20):
if shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main([self.source_orig, self.target_orig], temporary=True)
self.indomain_source_orig = indomain_source
self.indomain_target_orig = indomain_target
self.indomain_source, self.indomain_target = shuffle.main([self.indomain_source_orig, self.indomain_target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.indomain_source = fopen(indomain_source, 'r')
self.indomain_target = fopen(indomain_target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
self.interpolation_rate = interpolation_rate
self.cur_interpolation_rate = self.interpolation_rate
self.indomain_k = int(math.ceil(self.cur_interpolation_rate * self.k))
self.outdomain_k = self.k - self.indomain_k
def __init__(self, source, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
use_factor=False,
maxibatch_size=20,
keep_data_in_memory=False):
if keep_data_in_memory:
self.source, self.target = FileWrapper(source), FileWrapper(target)
if shuffle_each_epoch:
r = numpy.random.permutation(len(self.source))
self.source.shuffle_lines(r)
self.target.shuffle_lines(r)
elif shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main([self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.keep_data_in_memory = keep_data_in_memory
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.use_factor = use_factor
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def __init__(self, source, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True,
indomain_source='', indomain_target='',
interpolation_rate=0.1,
maxibatch_size=20):
if shuffle_each_epoch:
shuffle.main([source, target])
shuffle.main([indomain_source, indomain_target])
self.source = fopen(source+'.shuf', 'r')
self.target = fopen(target+'.shuf', 'r')
self.indomain_source = fopen(indomain_source+'.shuf', 'r')
self.indomain_target = fopen(indomain_target+'.shuf', 'r')
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.indomain_source = fopen(indomain_source, 'r')
self.indomain_target = fopen(indomain_target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
self.interpolation_rate = interpolation_rate
self.indomain_k = int(math.ceil(self.interpolation_rate * self.k))
self.outdomain_k = self.k - self.indomain_k
def __init__(self,
source,
target,
source_dicts,
target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
use_factor=False,
maxibatch_size=20,
token_batch_size=0):
if shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main(
[self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.use_factor = use_factor
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.token_batch_size = token_batch_size
self.end_of_data = False
def __init__(
self,
source,
target,
source_dicts,
target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20,
):
global epoch_num
if shuffle_each_epoch:
shuffle.main([source, target], epoch_num)
self.source = fopen(source + '.shuf', 'r')
self.target = fopen(target + '.shuf', 'r')
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
self.embeddings = embeddings
def __init__(self,
source,
target,
source_dict,
target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
if shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main(
[self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
#for line in self.source.readlines():
#print line
#aline = self.target.readline()
#print aline
self.source_dict = load_dict(source_dict)
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0: # if source number is specified
for key, idx in self.source_dict.items():
if idx >= self.n_words_source:
del self.source_dict[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = [] # source instance in memory
self.target_buffer = [] # target instance in memory
self.k = batch_size * maxibatch_size # number of instance in memory in total
self.end_of_data = False
def main(models, source_file, nbest_file, saveto, b=80,
normalize=False, verbose=False, alignweights=False):
# load model model_options
options = []
for model in args.models:
try:
with open('%s.json' % model, 'rb') as f:
options.append(json.load(f))
except:
with open('%s.pkl' % model, 'rb') as f:
options.append(pkl.load(f))
#hacks for using old models with missing options
if not 'dropout_embedding' in options[-1]:
options[-1]['dropout_embedding'] = 0
if not 'dropout_hidden' in options[-1]:
options[-1]['dropout_hidden'] = 0
if not 'dropout_source' in options[-1]:
options[-1]['dropout_source'] = 0
if not 'dropout_target' in options[-1]:
options[-1]['dropout_target'] = 0
dictionaries = options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
dictionary_target = dictionaries[-1]
# load source dictionary and invert
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
if options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= options[0]['n_words_src']:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
# load target dictionary and invert
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
rescore_model(source_file, nbest_file, saveto, models, options, b, normalize, verbose, alignweights)
def read_model(self, params):
user_file = os.path.join(params['config_path'],
params['user_filename'])
item_file = os.path.join(params['config_path'],
params['item_filename'])
vocab_file = os.path.join(params['config_path'],
params['vocab_filename'])
aspect_file = os.path.join(params['config_path'],
params['aspect_filename'])
opinion_file = os.path.join(params['config_path'],
params['opinion_filename'])
aspect_opinions_file = os.path.join(params['config_path'],
params['aspect_opinions_filename'])
model_file = os.path.join(params['config_path'],
params['model_filename'])
context_word_units = int(params['unit'])
lstm_hidden_units = IN_TO_OUT_UNITS_RATIO * context_word_units
target_word_units = IN_TO_OUT_UNITS_RATIO * context_word_units
user2index = load_dict(user_file)
item2index = load_dict(item_file)
word2index = load_dict(vocab_file)
aspect2index = load_dict(aspect_file)
opinion2index = load_dict(opinion_file)
aspect_opinions = load_json(aspect_opinions_file)
n_user = max(user2index.values()) + 1
n_item = max(item2index.values()) + 1
n_vocab = max(word2index.values()) + 1
n_aspect = max(aspect2index.values()) + 1
n_encode = n_aspect
# dummy word counts - not used for eval
cs = [1 for _ in range(n_vocab)]
# dummy loss func - not used for eval
loss_func = L.NegativeSampling(target_word_units, cs,
NEGATIVE_SAMPLING_NUM)
if params['model_type'] == 'c2v':
model = Context2Vec(self.gpu, n_vocab, context_word_units,
lstm_hidden_units, target_word_units,
loss_func, self.resume)
elif params['model_type'] in ['asc2v', 'asc2v-mter']:
model = AspectSentiContext2Vec(self.gpu, n_vocab, n_encode,
context_word_units,
lstm_hidden_units,
target_word_units, loss_func,
self.resume)
S.load_npz(model_file, model)
w = model.loss_func.W.data
return user2index, item2index, w, word2index, aspect2index, opinion2index, aspect_opinions, model
def testLoad(self,cfg):
cfg = self.cfg
entity2id = load_dict(cfg['data_folder'] + cfg['entity2id'])
word2id = load_dict(cfg['data_folder'] + cfg['word2id'])
relation2id = load_dict(cfg['data_folder'] + cfg['relation2id'])
train_documents = load_documents(cfg['data_folder'] + cfg['train_documents'])
train_document_entity_indices, train_document_texts = index_document_entities(train_documents, word2id,
entity2id,
cfg['max_document_word'])
train_data = DataLoader(cfg['data_folder'] + cfg['train_data'], train_documents, train_document_entity_indices,
train_document_texts, word2id, relation2id, entity2id, cfg['max_query_word'],
cfg['max_document_word'], cfg['use_kb'], cfg['use_doc'], cfg['use_inverse_relation'])
def predict():
word_to_ix = load_dict(word_dict_file)
tag_to_ix = load_dict(tag_dict_file)
ix_to_tag = {v: k for k, v in tag_to_ix.items()}
model_file = model_path + 'params.pkl'
if os.path.exists(model_file):
model.load_state_dict(torch.load(model_file))
for wordss, tagss, lengths in pred_helper.gen_batch():
sentence_in = prepare_sequence(wordss, word_to_ix)
predict_scores, predict_ix_seqs = model(sentence_in, lengths)
for word, ix in zip(wordss[0], predict_ix_seqs[0]):
print(word, ix_to_tag[ix])
print()
def __init__(self,
source,
target,
source_dict,
target_dict,
batch_size=128,
maxlen=None,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
if shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main(
[self.source_orig, self.target_orig], temporary=True)
else:
self.source = data_utils.fopen(source, 'r')
self.target = data_utils.fopen(target, 'r')
self.source_dict = load_dict(source_dict)
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for key, idx in self.source_dict.items():
if idx >= self.n_words_source:
del self.source_dict[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def __init__(self, source, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
if shuffle_each_epoch:
shuffle.main([source, target])
self.source = fopen(source+'.shuf', 'r')
self.target = fopen(target+'.shuf', 'r')
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
print "K=", self.k
self.end_of_data = False
def __init__(self, options, **kwargs):
self.src_dict = load_dict(options['dictionaries'][0])
self.trg_dict = load_dict(options['dictionaries'][1])
self.tmp_dir = kwargs['tmp_dir']
self.translate_script = kwargs['translate_script']
self.bleu_script = kwargs['bleu_script']
self.valid_src = kwargs['bleuvalid_src']
self.valid_trg = kwargs['bleuvalid_trg']
self.n_words_src = options['n_words_src']
self.batch_size = 16
self.batches = self.prepare_data()
os.system('mkdir -p %s' % self.tmp_dir)
self.check_script() # check bleu script
self.trg_idict = dict()
for k, v in self.trg_dict.iteritems():
self.trg_idict[v] = k
def test():
entity2id = load_dict(entity2id_file)
test_data = TypedataLoader(test_file, entity2id)
my_model = get_model(entity2id)
test_acc = inference(my_model, test_data, entity2id, log_info=True)
return test_acc
def __init__(self, sources, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=[-1],
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
if shuffle_each_epoch:
shuffle.main(sources + [target])
self.sources = [fopen(source+'.shuf', 'r') for source in sources]
self.target = fopen(target+'.shuf', 'r')
else:
self.sources = [fopen(source, 'r') for source in sources]
self.target = fopen(target, 'r')
self.source_dicts = []
for factor_dicts in source_dicts:
self.source_dicts.append([load_dict(source_dict) for source_dict in factor_dicts])
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
for i, n_words in enumerate(self.n_words_source):
if n_words > 0:
for d in self.source_dicts[i]:
for key, idx in d.items():
if idx >= n_words:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffers = [list() for _ in range(len(self.sources))]
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def __init__(self, config, documents, mode='train'):
self.mode = mode
self.use_doc = config['use_doc']
self.use_inverse_relation = config['use_inverse_relation']
self.max_query_word = config['max_query_word']
self.max_document_word = config['max_document_word']
self.max_char = config['max_char']
self.documents = documents
self.data_file = config['data_folder'] + config['{}_data'.format(mode)]
self.batch_size = config['batch_size'] if mode == 'train' else config[
'batch_size']
self.max_rel_words = config['max_rel_words']
self.type_rels = config['type_rels']
self.fact_drop = config['fact_drop']
# read all data
self.data = []
with open(self.data_file) as f:
for line in tqdm(list(f)):
self.data.append(json.loads(line))
# word and kb vocab
self.word2id = load_dict(config['data_folder'] + config['word2id'])
self.relation2id = load_dict(config['data_folder'] +
config['relation2id'])
self.entity2id = load_dict(config['data_folder'] + config['entity2id'])
self.id2entity = {i: entity for entity, i in self.entity2id.items()}
self.rel_word_idx = np.load(config['data_folder'] + 'rel_word_idx.npy')
# for batching
self.max_local_entity = 0 # max num of candidates
self.max_relevant_docs = 0 # max num of retired documents
self.max_kb_neighbors = config[
'max_num_neighbors'] # max num of neighbors for entity
self.max_kb_neighbors_ = config[
'max_num_neighbors'] # kb relations are directed
self.max_linked_entities = 0 # max num of linked entities for each doc
self.max_linked_documents = 50 # max num of linked documents for each entity
self.num_kb_relation = 2 * len(
self.relation2id) if self.use_inverse_relation else len(
self.relation2id)
# get the batching parameters
self.get_stats()
def train():
print("training ...")
#prepare data
entity2id = load_dict(entity2id_file)
train_data = TypedataLoader(train_file, entity2id)
dev_data = TypedataLoader(dev_file, entity2id)
test_data = TypedataLoader(test_file, entity2id)
my_model = get_model(entity2id)
trainable_parameters = [
p for p in my_model.parameters() if p.requires_grad
]
optimizer = torch.optim.Adam(trainable_parameters, lr=learning_rate)
best_dev_acc = 0.0
for i in range(epoch):
try:
print('epoch', i)
my_model.train()
train_loss, train_acc = [], []
for iteration in tqdm(range(train_data.num_data // batch_size)):
batch = train_data.get_batch(iteration, batch_size)
loss, pred, _ = my_model(batch)
pred = pred.data.cpu().numpy()
acc = cal_type_acc(pred, batch[-1])
train_loss.append(loss.data[0])
train_acc.append(acc)
# back propogate
my_model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(my_model.parameters(),
gradient_clip)
optimizer.step()
print('avg_training_loss', sum(train_loss) / len(train_loss))
print('avg_training_acc', sum(train_acc) / len(train_acc))
print("validating ...")
eval_acc = inference(my_model, dev_data, entity2id)
if eval_acc > best_dev_acc and save_model:
print("saving model to", save_model)
torch.save(my_model.state_dict(), save_model)
best_dev_acc = eval_acc
except KeyboardInterrupt:
break
# Test set evaluation
print("evaluating on test")
print('loading model from ...', test_file)
my_model.load_state_dict(torch.load(save_model))
test_acc = inference(my_model, test_data, entity2id, log_info=True)
print("test_acc:", test_acc)
return test_acc
def test(cfg):
entity2id = load_dict(cfg['data_folder'] + cfg['entity2id'])
word2id = load_dict(cfg['data_folder'] + cfg['word2id'])
relation2id = load_dict(cfg['data_folder'] + cfg['relation2id'])
test_documents = load_documents(cfg['data_folder'] + cfg['test_documents'])
test_document_entity_indices, test_document_texts = index_document_entities(
test_documents, word2id, entity2id, cfg['max_document_word'])
test_data = DataLoader(cfg['data_folder'] + cfg['test_data'],
test_documents, test_document_entity_indices,
test_document_texts, word2id, relation2id,
entity2id, cfg['max_query_word'],
cfg['max_document_word'], cfg['use_kb'],
cfg['use_doc'], cfg['use_inverse_relation'])
my_model = get_model(cfg, test_data.num_kb_relation, len(entity2id),
len(word2id)).to(device)
test_acc = inference(my_model, test_data, entity2id, cfg, log_info=True)
return test_acc
def _build_dictionaries(self, source_dic, target_dic):
"""
Builds and inverts source and target dictionaries, taken
from the first model since all of them must have the same
vocabulary.
"""
if source_dic == None or target_dic == None:
dictionaries = self._options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
dictionary_target = dictionaries[-1]
else:
dictionaries_source = [source_dic]
dictionary_target = target_dic
# load and invert source dictionaries
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
if self._options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= self._options[0]['n_words_src']:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
self._word_dicts = word_dicts
self._word_idicts = word_idicts
# load and invert target dictionary
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
self._word_idict_trg = word_idict_trg
def __init__(self,
source,
target,
source_dict,
target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
# 每次epoch都，打乱文件顺序
if shuffle_each_epoch:
shuffle.main([source, target])
self.source = fopen(source + '.shuf')
self.target = fopen(target + '.shuf')
else:
self.source = fopen(source)
self.target = fopen(target)
self.source_dict = load_dict(source_dict)
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for key, idx in self.source_dict.items():
if idx >= self.n_words_source:
del self.source_dict[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def _build_dictionaries(self):
"""
Builds and inverts source and target dictionaries, taken
from the first model since all of them must have the same
vocabulary.
"""
dictionaries = self._options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
dictionary_target = dictionaries[-1]
# load and invert source dictionaries
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
# n_words is a list containing the max len of each dictionary
if self._options[0]['n_words'][0]:
for key, idx in word_dict.items():
if idx >= self._options[0]['n_words'][0]:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
self._word_dicts = word_dicts
self._word_idicts = word_idicts
# load and invert target dictionary
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
self._word_idict_trg = word_idict_trg
def _determine_vocab_size_from_file(path, plus_one):
""" plus_one give place for unk
"""
try:
d = load_dict(path)
except IOError as x:
logging.error('failed to determine vocabulary size from file: '
'{}: {}'.format(path, str(x)))
sys.exit(1)
except:
logging.error('failed to determine vocabulary size from file: '
'{}'.format(path))
sys.exit(1)
return max(d.values()) + 1 if plus_one else max(d.values())
def __init__(self, dict_file=DICT_FILE, schema_file=SCHEMA_FILE):
"""
init
"""
#self.logger.info("hook")
word_dict = util.load_dict(dict_file)
schema_pos, schema_output = util.get_parse_shitu_conf(schema_file)
self.word_dict = word_dict
self.schema_pos = schema_pos
self.schema_output = schema_output
dict_size = len(word_dict)
schema_pos_size = len(schema_pos)
schema_output_size = len(schema_output)
def __init__(self,
datasets,
dicts,
n_words_dicts=None,
batch_size=128,
maxlen=100,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
factors=1,
outputs=1,
maxibatch_size=20):
if shuffle_each_epoch:
self.datasets_orig = datasets
self.datasets = shuffle.main(datasets, temporary=True)
else:
self.datasets = [fopen(fp, 'r') for fp in datasets]
self.dicts = []
for dict_ in dicts:
self.dicts.append(load_dict(dict_))
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.factors = factors
self.outputs = outputs
assert len(
datasets) == 1 + outputs, 'Datasets and dictionaries mismatch'
self.n_words_dicts = n_words_dicts
if self.n_words_dicts:
for d, max_ in zip(self.dicts, self.n_words_dicts):
for key, idx in d.items():
if idx >= max_:
del d[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.buffers = [[] for _ in range(len(datasets))]
self.k = batch_size * maxibatch_size
self.end_of_data = False
def __init__(self,
source,
source_dicts,
batch_size=128,
maxlen=100,
n_words_source=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
maxibatch_size=20):
if shuffle_each_epoch:
self.source_orig = source
self.source = shuffle.main([self.source_orig], temporary=True)
self.source = self.source[0] # ???
print('this had better be a file:', type(self.source))
else:
self.source = fopen(source, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.n_words_source = n_words_source
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def load_dict_from_model_config(models):
import re
re.sub(' +', ' ', models)
model = models.split(" ")[0]
options = []
try:
with open('%s.json' % model, 'rb') as f:
options.append(json.load(f))
except:
with open('%s.pkl' % model, 'rb') as f:
options.append(pkl.load(f))
dictionaries = options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
#dictionary_target = dictionaries[-1]
word_dict = load_dict(dictionaries_source[0])
if options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= options[0]['n_words_src']:
del word_dict[key]
del word_dict['<EOS>']
del word_dict['<UNK>']
return word_dict
def init_prefs(prefs_path):
'''
Initializes the default preferences for the stoichiometry program, and
loads any preferences that have been saved to the specified file.
Returns a dictionary of preferences.
'''
# Define default preferences.
dprefs = {
'wdir': os.path.abspath('.'),
'autosave_prefs': True,
'delimiter': '\t'
}
# Check for preferences file; load it if it exists.
lprefs = {}
if os.path.isfile(prefs_path):
lprefs = util.load_dict(prefs_path)
# Supply default preferences that are missing.
for key in dprefs.keys():
if key not in lprefs:
lprefs[key] = dprefs[key]
# Return preferences.
return lprefs
def parse_args():
parser = argparse.ArgumentParser()
data = parser.add_argument_group('data sets; model loading and saving')
data.add_argument('--source_dataset',
type=str,
metavar='PATH',
help="parallel training corpus (source)")
data.add_argument('--target_dataset',
type=str,
metavar='PATH',
help="parallel training corpus (target)")
# parallel training corpus (source and target). Hidden option for backward compatibility
data.add_argument('--datasets',
type=str,
metavar='PATH',
nargs=2,
help=argparse.SUPPRESS)
data.add_argument(
'--dictionaries',
type=str,
required=True,
metavar='PATH',
nargs="+",
help=
"network vocabularies (one per source factor, plus target vocabulary)")
data.add_argument('--saveFreq',
type=int,
default=30000,
metavar='INT',
help="save frequency (default: %(default)s)")
data.add_argument('--model',
'--saveto',
type=str,
default='model',
metavar='PATH',
dest='saveto',
help="model file name (default: %(default)s)")
data.add_argument(
'--reload',
type=str,
default=None,
metavar='PATH',
help=
"load existing model from this path. Set to \"latest_checkpoint\" to reload the latest checkpoint in the same directory of --saveto"
)
data.add_argument(
'--no_reload_training_progress',
action='store_false',
dest='reload_training_progress',
help="don't reload training progress (only used if --reload is enabled)"
)
data.add_argument(
'--summary_dir',
type=str,
required=False,
metavar='PATH',
help=
"directory for saving summaries (default: same directory as the --saveto file)"
)
data.add_argument(
'--summaryFreq',
type=int,
default=0,
metavar='INT',
help=
"Save summaries after INT updates, if 0 do not save summaries (default: %(default)s)"
)
network = parser.add_argument_group('network parameters')
network.add_argument('--embedding_size',
'--dim_word',
type=int,
default=512,
metavar='INT',
help="embedding layer size (default: %(default)s)")
network.add_argument('--state_size',
'--dim',
type=int,
default=1000,
metavar='INT',
help="hidden state size (default: %(default)s)")
network.add_argument(
'--source_vocab_sizes',
'--n_words_src',
type=int,
default=None,
nargs='+',
metavar='INT',
help=
"source vocabulary sizes (one per input factor) (default: %(default)s)"
)
network.add_argument('--target_vocab_size',
'--n_words',
type=int,
default=-1,
metavar='INT',
help="target vocabulary size (default: %(default)s)")
network.add_argument('--factors',
type=int,
default=1,
metavar='INT',
help="number of input factors (default: %(default)s)")
network.add_argument(
'--dim_per_factor',
type=int,
default=None,
nargs='+',
metavar='INT',
help=
"list of word vector dimensionalities (one per factor): '--dim_per_factor 250 200 50' for total dimensionality of 500 (default: %(default)s)"
)
network.add_argument(
'--enc_depth',
type=int,
default=1,
metavar='INT',
help="number of encoder layers (default: %(default)s)")
network.add_argument(
'--enc_recurrence_transition_depth',
type=int,
default=1,
metavar='INT',
help=
"number of GRU transition operations applied in the encoder. Minimum is 1. (Only applies to gru). (default: %(default)s)"
)
network.add_argument(
'--dec_depth',
type=int,
default=1,
metavar='INT',
help="number of decoder layers (default: %(default)s)")
network.add_argument(
'--dec_base_recurrence_transition_depth',
type=int,
default=2,
metavar='INT',
help=
"number of GRU transition operations applied in the first layer of the decoder. Minimum is 2. (Only applies to gru_cond). (default: %(default)s)"
)
network.add_argument(
'--dec_high_recurrence_transition_depth',
type=int,
default=1,
metavar='INT',
help=
"number of GRU transition operations applied in the higher layers of the decoder. Minimum is 1. (Only applies to gru). (default: %(default)s)"
)
network.add_argument(
'--dec_deep_context',
action='store_true',
help="pass context vector (from first layer) to deep decoder layers")
network.add_argument('--use_dropout',
action="store_true",
help="use dropout layer (default: %(default)s)")
network.add_argument(
'--dropout_embedding',
type=float,
default=0.2,
metavar="FLOAT",
help=
"dropout for input embeddings (0: no dropout) (default: %(default)s)")
network.add_argument(
'--dropout_hidden',
type=float,
default=0.2,
metavar="FLOAT",
help="dropout for hidden layer (0: no dropout) (default: %(default)s)")
network.add_argument(
'--dropout_source',
type=float,
default=0.0,
metavar="FLOAT",
help="dropout source words (0: no dropout) (default: %(default)s)")
network.add_argument(
'--dropout_target',
type=float,
default=0.0,
metavar="FLOAT",
help="dropout target words (0: no dropout) (default: %(default)s)")
network.add_argument(
'--use_layer_norm',
'--layer_normalisation',
action="store_true",
dest="use_layer_norm",
help="Set to use layer normalization in encoder and decoder")
network.add_argument(
'--tie_encoder_decoder_embeddings',
action="store_true",
dest="tie_encoder_decoder_embeddings",
help=
"tie the input embeddings of the encoder and the decoder (first factor only). Source and target vocabulary size must be the same"
)
network.add_argument(
'--tie_decoder_embeddings',
action="store_true",
dest="tie_decoder_embeddings",
help=
"tie the input embeddings of the decoder with the softmax output embeddings"
)
network.add_argument(
'--output_hidden_activation',
type=str,
default='tanh',
choices=['tanh', 'relu', 'prelu', 'linear'],
help=
'activation function in hidden layer of the output network (default: %(default)s)'
)
network.add_argument(
'--softmax_mixture_size',
type=int,
default=1,
metavar="INT",
help="number of softmax components to use (default: %(default)s)")
training = parser.add_argument_group('training parameters')
training.add_argument(
'--maxlen',
type=int,
default=100,
metavar='INT',
help=
"maximum sequence length for training and validation (default: %(default)s)"
)
training.add_argument('--batch_size',
type=int,
default=80,
metavar='INT',
help="minibatch size (default: %(default)s)")
training.add_argument(
'--token_batch_size',
type=int,
default=0,
metavar='INT',
help=
"minibatch size (expressed in number of source or target tokens). Sentence-level minibatch size will be dynamic. If this is enabled, batch_size only affects sorting by length. (default: %(default)s)"
)
training.add_argument(
'--max_epochs',
type=int,
default=5000,
metavar='INT',
help="maximum number of epochs (default: %(default)s)")
training.add_argument(
'--finish_after',
type=int,
default=10000000,
metavar='INT',
help="maximum number of updates (minibatches) (default: %(default)s)")
training.add_argument(
'--decay_c',
type=float,
default=0.0,
metavar='FLOAT',
help="L2 regularization penalty (default: %(default)s)")
training.add_argument(
'--map_decay_c',
type=float,
default=0.0,
metavar='FLOAT',
help=
"MAP-L2 regularization penalty towards original weights (default: %(default)s)"
)
training.add_argument(
'--prior_model',
type=str,
metavar='PATH',
help=
"Prior model for MAP-L2 regularization. Unless using \"--reload\", this will also be used for initialization."
)
training.add_argument(
'--clip_c',
type=float,
default=1.0,
metavar='FLOAT',
help="gradient clipping threshold (default: %(default)s)")
training.add_argument('--learning_rate',
'--lrate',
type=float,
default=0.0001,
metavar='FLOAT',
help="learning rate (default: %(default)s)")
training.add_argument('--label_smoothing',
type=float,
default=0.0,
metavar='FLOAT',
help="label smoothing (default: %(default)s)")
training.add_argument(
'--no_shuffle',
action="store_false",
dest="shuffle_each_epoch",
help="disable shuffling of training data (for each epoch)")
training.add_argument(
'--keep_train_set_in_memory',
action="store_true",
help="Keep training dataset lines stores in RAM during training")
training.add_argument('--no_sort_by_length',
action="store_false",
dest="sort_by_length",
help='do not sort sentences in maxibatch by length')
training.add_argument(
'--maxibatch_size',
type=int,
default=20,
metavar='INT',
help=
'size of maxibatch (number of minibatches that are sorted by length) (default: %(default)s)'
)
training.add_argument('--optimizer',
type=str,
default="adam",
choices=['adam'],
help="optimizer (default: %(default)s)")
validation = parser.add_argument_group('validation parameters')
validation.add_argument(
'--valid_source_dataset',
type=str,
default=None,
metavar='PATH',
help="source validation corpus (default: %(default)s)")
validation.add_argument(
'--valid_target_dataset',
type=str,
default=None,
metavar='PATH',
help="target validation corpus (default: %(default)s)")
# parallel validation corpus (source and target). Hidden option for backward compatibility
validation.add_argument('--valid_datasets',
type=str,
default=None,
metavar='PATH',
nargs=2,
help=argparse.SUPPRESS)
validation.add_argument(
'--valid_batch_size',
type=int,
default=80,
metavar='INT',
help="validation minibatch size (default: %(default)s)")
training.add_argument(
'--valid_token_batch_size',
type=int,
default=0,
metavar='INT',
help=
"validation minibatch size (expressed in number of source or target tokens). Sentence-level minibatch size will be dynamic. If this is enabled, valid_batch_size only affects sorting by length. (default: %(default)s)"
)
validation.add_argument('--validFreq',
type=int,
default=10000,
metavar='INT',
help="validation frequency (default: %(default)s)")
validation.add_argument(
'--valid_script',
type=str,
default=None,
metavar='PATH',
help=
"path to script for external validation (default: %(default)s). The script will be passed an argument specifying the path of a file that contains translations of the source validation corpus. It must write a single score to standard output."
)
validation.add_argument(
'--patience',
type=int,
default=10,
metavar='INT',
help="early stopping patience (default: %(default)s)")
validation.add_argument(
'--run_validation',
action='store_true',
help="Compute validation score on validation dataset")
display = parser.add_argument_group('display parameters')
display.add_argument(
'--dispFreq',
type=int,
default=1000,
metavar='INT',
help="display loss after INT updates (default: %(default)s)")
display.add_argument(
'--sampleFreq',
type=int,
default=10000,
metavar='INT',
help="display some samples after INT updates (default: %(default)s)")
display.add_argument(
'--beamFreq',
type=int,
default=10000,
metavar='INT',
help=
"display some beam_search samples after INT updates (default: %(default)s)"
)
display.add_argument('--beam_size',
type=int,
default=12,
metavar='INT',
help="size of the beam (default: %(default)s)")
translate = parser.add_argument_group('translate parameters')
translate.add_argument('--translate_valid',
action='store_true',
dest='translate_valid',
help='Translate source dataset instead of training')
translate.add_argument(
'--no_normalize',
action='store_false',
dest='normalize',
help="Cost of sentences will not be normalized by length")
translate.add_argument('--n_best',
action='store_true',
dest='n_best',
help="Print full beam")
translate.add_argument(
'--n_threads',
type=int,
default=5,
metavar='INT',
help="Number of threads to use for beam search (default: %(default)s)")
translate.add_argument(
'--translation_maxlen',
type=int,
default=200,
metavar='INT',
help=
"Maximum length of translation output sentence (default: %(default)s)")
config = parser.parse_args()
# allow "--datasets" for backward compatibility
if config.datasets:
if config.source_dataset or config.target_dataset:
logging.error(
'argument clash: --datasets is mutually exclusive with --source_dataset and --target_dataset'
)
sys.exit(1)
else:
config.source_dataset = config.datasets[0]
config.target_dataset = config.datasets[1]
elif not config.source_dataset:
logging.error('--source_dataset is required')
sys.exit(1)
elif not config.target_dataset:
logging.error('--target_dataset is required')
sys.exit(1)
# allow "--valid_datasets" for backward compatibility
if config.valid_datasets:
if config.valid_source_dataset or config.valid_target_dataset:
logging.error(
'argument clash: --valid_datasets is mutually exclusive with --valid_source_dataset and --valid_target_dataset'
)
sys.exit(1)
else:
config.valid_source_dataset = config.valid_datasets[0]
config.valid_target_dataset = config.valid_datasets[1]
# check factor-related options are consistent
if config.dim_per_factor == None:
if config.factors == 1:
config.dim_per_factor = [config.embedding_size]
else:
logging.error(
'if using factored input, you must specify \'dim_per_factor\'\n'
)
sys.exit(1)
if len(config.dim_per_factor) != config.factors:
logging.error(
'mismatch between \'--factors\' ({0}) and \'--dim_per_factor\' ({1} entries)\n'
.format(config.factors, len(config.dim_per_factor)))
sys.exit(1)
if sum(config.dim_per_factor) != config.embedding_size:
logging.error(
'mismatch between \'--embedding_size\' ({0}) and \'--dim_per_factor\' (sums to {1})\n'
.format(config.embedding_size, sum(config.dim_per_factor)))
sys.exit(1)
if len(config.dictionaries) != config.factors + 1:
logging.error(
'\'--dictionaries\' must specify one dictionary per source factor and one target dictionary\n'
)
sys.exit(1)
# determine target_embedding_size
if config.tie_encoder_decoder_embeddings:
config.target_embedding_size = config.dim_per_factor[0]
else:
config.target_embedding_size = config.embedding_size
# set vocabulary sizes
vocab_sizes = []
if config.source_vocab_sizes == None:
vocab_sizes = [-1] * config.factors
elif len(config.source_vocab_sizes) == config.factors:
vocab_sizes = config.source_vocab_sizes
elif len(config.source_vocab_sizes) < config.factors:
num_missing = config.factors - len(config.source_vocab_sizes)
vocab_sizes += config.source_vocab_sizes + [-1] * num_missing
else:
logging.error(
'too many values supplied to \'--source_vocab_sizes\' option (expected one per factor = {0})'
.format(config.factors))
sys.exit(1)
if config.target_vocab_size == -1:
vocab_sizes.append(-1)
else:
vocab_sizes.append(config.target_vocab_size)
# for unspecified vocabulary sizes, determine sizes from vocabulary dictionaries
for i, vocab_size in enumerate(vocab_sizes):
if vocab_size >= 0:
continue
try:
d = util.load_dict(config.dictionaries[i])
except:
logging.error(
'failed to determine vocabulary size from file: {0}'.format(
config.dictionaries[i]))
vocab_sizes[i] = max(d.values()) + 1
config.source_dicts = config.dictionaries[:-1]
config.source_vocab_sizes = vocab_sizes[:-1]
config.target_dict = config.dictionaries[-1]
config.target_vocab_size = vocab_sizes[-1]
# set the model version
config.model_version = 0.2
config.theano_compat = False
return config
def main(models, source_files, saveto, save_alignment, k=5,
normalize=False, n_process=5, chr_level=False, verbose=False,
nbest=False, suppress_unk=False, a_json=False,
print_word_probabilities=False):
# load model model_options
options = []
for model in args.models:
try:
with open('%s.json' % model, 'rb') as f:
options.append(json.load(f))
except:
with open('%s.pkl' % model, 'rb') as f:
options.append(pkl.load(f))
#hacks for using old models with missing options
if not 'dropout_embedding' in options[-1]:
options[-1]['dropout_embedding'] = 0
if 'dropout_hidden' not in options[-1]:
options[-1]['dropout_hidden'] = 0
if 'dropout_source' not in options[-1]:
options[-1]['dropout_source'] = 0
if 'dropout_target' not in options[-1]:
options[-1]['dropout_target'] = 0
if 'factors' not in options[-1]:
options[-1]['factors'] = 1
if 'dim_per_factor' not in options[-1]:
options[-1]['dim_per_factor'] = [options[-1]['dim_word']]
dictionaries = options[0]['dictionaries']
dictionaries_sources = dictionaries[:-1]
print >> sys.stderr, "SRC DICT:", dictionaries_sources
dictionary_target = dictionaries[-1]
print >> sys.stderr, "TRG DICT:", dictionary_target
encoders_word_dicts = []
encoders_word_idicts = []
for dictionaries_source in dictionaries_sources:
# load source dictionary and invert
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
if options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= options[0]['n_words_src']:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
encoders_word_dicts.append(word_dicts)
encoders_word_idicts.append(word_idicts)
# load target dictionary and invert
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
# create input and output queues for processes
queue = Queue()
rqueue = Queue()
processes = [None] * n_process
for midx in xrange(n_process):
processes[midx] = Process(
target=translate_model,
args=(queue, rqueue, midx, models, options, k, normalize, verbose,
nbest, save_alignment is not None, suppress_unk))
processes[midx].start()
# utility function
def _seqs2words(cc):
ww = []
for w in cc:
if w == 0:
break
ww.append(word_idict_trg[w])
return ' '.join(ww)
def _send_jobs(files):
source_sentences = []
for idx, lines in enumerate(zip(*files)):
# print lines
enc_idx = 0;
xs = []
enc_words = []
for src_idx, line in enumerate(lines):
if chr_level:
words = list(line.decode('utf-8').strip())
else:
words = line.strip().split()
x = []
for w in words:
w = [encoders_word_dicts[src_idx][i][f] if f in word_dicts[i] else 1
for (i, f) in enumerate(w.split('|'))]
if len(w) != options[0]['factors'][enc_idx]:
sys.stderr.write('Error: expected {0} factors, but input word has {1}\n'.format(options[0]['factors'][enc_idx], len(w)))
for midx in xrange(n_process):
processes[midx].terminate()
sys.exit(1)
x.append(w)
x += [ [0] * options[0]['factors'][enc_idx] ]
# print "X:", x
xs.append(x)
enc_words.append(words)
# print "XS:", xs
# new_xs = [ [xs[j][i] for j in range(len(xs)) ] for i in range(len(xs[0])) ]
# print "NEW XS:", new_xs
queue.put((idx, xs))
source_sentences.append(enc_words)
enc_idx += 1
return idx+1, source_sentences
def _finish_processes():
for midx in xrange(n_process):
queue.put(None)
def _retrieve_jobs(n_samples):
trans = [None] * n_samples
out_idx = 0
for idx in xrange(n_samples):
resp = rqueue.get()
trans[resp[0]] = resp[1]
if verbose and numpy.mod(idx, 10) == 0:
sys.stderr.write('Sample {0} / {1} Done\n'.format((idx+1), n_samples))
while out_idx < n_samples and trans[out_idx] != None:
yield trans[out_idx]
out_idx += 1
sys.stderr.write('Translating {0} ...\n'.format(':'.join([f.name for f in source_files])))
n_samples, source_sentences = _send_jobs(source_files)
_finish_processes()
for i, trans in enumerate(_retrieve_jobs(n_samples)):
if nbest:
samples, scores, word_probs, alignment = trans
order = numpy.argsort(scores)
for j in order:
saveto.write('{0} ||| {1} ||| {2}\n'.format(i, _seqs2words(samples[j]), scores[j]))
# print alignment matrix for each hypothesis
# header: sentence id ||| translation ||| score ||| source ||| source_token_count+eos translation_token_count+eos
if save_alignment is not None:
if a_json:
print_matrix_json(alignment[j], source_sentences[i], _seqs2words(samples[j]).split(), i, i+j,save_alignment)
else:
save_alignment.write('{0} ||| {1} ||| {2} ||| {3} ||| {4} {5}\n'.format(
i, _seqs2words(samples[j]), scores[j], ' '.join(source_sentences[i]) , len(source_sentences[i])+1, len(samples[j])))
print_matrix(alignment[j], save_alignment)
else:
samples, scores, word_probs, alignment = trans
saveto.write(_seqs2words(samples) + "\n")
if print_word_probabilities:
for prob in word_probs:
saveto.write("{} ".format(prob))
saveto.write('\n')
if save_alignment is not None:
if a_json:
print_matrix_json(trans[1], source_sentences[i], _seqs2words(trans[0]).split(), i, i,save_alignment)
else:
save_alignment.write('{0} ||| {1} ||| {2} ||| {3} ||| {4} {5}\n'.format(
i, _seqs2words(trans[0]), 0, ' '.join(source_sentences[i]) , len(source_sentences[i])+1, len(trans[0])))
print_matrix(trans[3], save_alignment)
sys.stderr.write('Done\n')
def main(models, source_file, saveto, k=5,
normalize=False, n_process=5, chr_level=False, verbose=False, nbest=False, suppress_unk=False):
# load model model_options
options = []
for model in args.models:
try:
with open('%s.json' % model, 'rb') as f:
options.append(json.load(f))
except:
with open('%s.pkl' % model, 'rb') as f:
options.append(pkl.load(f))
#hacks for using old models with missing options
if not 'dropout_embedding' in options[-1]:
options[-1]['dropout_embedding'] = 0
if not 'dropout_hidden' in options[-1]:
options[-1]['dropout_hidden'] = 0
if not 'dropout_source' in options[-1]:
options[-1]['dropout_source'] = 0
if not 'dropout_target' in options[-1]:
options[-1]['dropout_target'] = 0
dictionary, dictionary_target = options[0]['dictionaries']
# load source dictionary and invert
word_dict = load_dict(dictionary)
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
# load target dictionary and invert
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
# create input and output queues for processes
queue = Queue()
rqueue = Queue()
processes = [None] * n_process
for midx in xrange(n_process):
processes[midx] = Process(
target=translate_model,
args=(queue, rqueue, midx, models, options, k, normalize, verbose, nbest, suppress_unk))
processes[midx].start()
# utility function
def _seqs2words(cc):
ww = []
for w in cc:
if w == 0:
break
ww.append(word_idict_trg[w])
return ' '.join(ww)
def _send_jobs(f):
for idx, line in enumerate(f):
if chr_level:
words = list(line.decode('utf-8').strip())
else:
words = line.strip().split()
x = map(lambda w: word_dict[w] if w in word_dict else 1, words)
x = map(lambda ii: ii if ii < options[0]['n_words_src'] else 1, x)
x += [0]
queue.put((idx, x))
return idx+1
def _finish_processes():
for midx in xrange(n_process):
queue.put(None)
def _retrieve_jobs(n_samples):
trans = [None] * n_samples
out_idx = 0
for idx in xrange(n_samples):
resp = rqueue.get()
trans[resp[0]] = resp[1]
if verbose and numpy.mod(idx, 10) == 0:
sys.stderr.write('Sample {0} / {1} Done\n'.format((idx+1), n_samples))
while out_idx < n_samples and trans[out_idx] != None:
yield trans[out_idx]
out_idx += 1
sys.stderr.write('Translating {0} ...\n'.format(source_file.name))
n_samples = _send_jobs(source_file)
_finish_processes()
for i, trans in enumerate(_retrieve_jobs(n_samples)):
if nbest:
samples, scores = trans
order = numpy.argsort(scores)
for j in order:
saveto.write('{0} ||| {1} ||| {2}\n'.format(i, _seqs2words(samples[j]), scores[j]))
else:
saveto.write(_seqs2words(trans) + '\n')
sys.stderr.write('Done\n')
def __init__(self, source, target,
source_dicts, target_dict,
model_type,
batch_size=128,
maxlen=100,
source_vocab_sizes=None,
target_vocab_size=None,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
use_factor=False,
maxibatch_size=20,
token_batch_size=0,
keep_data_in_memory=False):
if keep_data_in_memory:
self.source, self.target = FileWrapper(source), FileWrapper(target)
if shuffle_each_epoch:
r = numpy.random.permutation(len(self.source))
self.source.shuffle_lines(r)
self.target.shuffle_lines(r)
elif shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main([self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict, model_type))
self.target_dict = load_dict(target_dict, model_type)
# Determine the UNK value for each dictionary (the value depends on
# which version of build_dictionary.py was used).
def determine_unk_val(d):
if '<UNK>' in d and d['<UNK>'] == 2:
return 2
return 1
self.source_unk_vals = [determine_unk_val(d)
for d in self.source_dicts]
self.target_unk_val = determine_unk_val(self.target_dict)
self.keep_data_in_memory = keep_data_in_memory
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.use_factor = use_factor
self.source_vocab_sizes = source_vocab_sizes
self.target_vocab_size = target_vocab_size
self.token_batch_size = token_batch_size
if self.source_vocab_sizes != None:
assert len(self.source_vocab_sizes) == len(self.source_dicts)
for d, vocab_size in zip(self.source_dicts, self.source_vocab_sizes):
if vocab_size != None and vocab_size > 0:
for key, idx in list(d.items()):
if idx >= vocab_size:
del d[key]
if self.target_vocab_size != None and self.target_vocab_size > 0:
for key, idx in list(self.target_dict.items()):
if idx >= self.target_vocab_size:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def main(models, source_file, saveto, save_alignment=None, k=5,
normalize=False, n_process=5, chr_level=False, verbose=False, nbest=False, suppress_unk=False, a_json=False, print_word_probabilities=False, return_hyp_graph=False):
# load model model_options
options = []
for model in models:
options.append(load_config(model))
fill_options(options[-1])
dictionaries = options[0]['dictionaries']
dictionaries_source = dictionaries[:-1]
dictionary_target = dictionaries[-1]
# load source dictionary and invert
word_dicts = []
word_idicts = []
for dictionary in dictionaries_source:
word_dict = load_dict(dictionary)
if options[0]['n_words_src']:
for key, idx in word_dict.items():
if idx >= options[0]['n_words_src']:
del word_dict[key]
word_idict = dict()
for kk, vv in word_dict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
word_dicts.append(word_dict)
word_idicts.append(word_idict)
# load target dictionary and invert
word_dict_trg = load_dict(dictionary_target)
word_idict_trg = dict()
for kk, vv in word_dict_trg.iteritems():
word_idict_trg[vv] = kk
word_idict_trg[0] = '<eos>'
word_idict_trg[1] = 'UNK'
# create input and output queues for processes
queue = Queue()
rqueue = Queue()
processes = [None] * n_process
for midx in xrange(n_process):
processes[midx] = Process(
target=translate_model,
args=(queue, rqueue, midx, models, options, k, normalize, verbose, nbest, save_alignment is not None, suppress_unk, return_hyp_graph))
processes[midx].start()
# utility function
def _seqs2words(cc):
ww = []
for w in cc:
if w == 0:
break
ww.append(word_idict_trg[w])
return ' '.join(ww)
def _send_jobs(f):
source_sentences = []
for idx, line in enumerate(f):
if chr_level:
words = list(line.decode('utf-8').strip())
else:
words = line.strip().split()
x = []
for w in words:
w = [word_dicts[i][f] if f in word_dicts[i] else 1 for (i,f) in enumerate(w.split('|'))]
if len(w) != options[0]['factors']:
sys.stderr.write('Error: expected {0} factors, but input word has {1}\n'.format(options[0]['factors'], len(w)))
for midx in xrange(n_process):
processes[midx].terminate()
sys.exit(1)
x.append(w)
x += [[0]*options[0]['factors']]
queue.put((idx, x))
source_sentences.append(words)
return idx+1, source_sentences
def _finish_processes():
for midx in xrange(n_process):
queue.put(None)
def _retrieve_jobs(n_samples):
trans = [None] * n_samples
out_idx = 0
for idx in xrange(n_samples):
resp = rqueue.get()
trans[resp[0]] = resp[1]
if verbose and numpy.mod(idx, 10) == 0:
sys.stderr.write('Sample {0} / {1} Done\n'.format((idx+1), n_samples))
while out_idx < n_samples and trans[out_idx] != None:
yield trans[out_idx]
out_idx += 1
sys.stderr.write('Translating {0} ...\n'.format(source_file.name))
n_samples, source_sentences = _send_jobs(source_file)
_finish_processes()
for i, trans in enumerate(_retrieve_jobs(n_samples)):
if nbest:
samples, scores, word_probs, alignment, hyp_graph = trans
if return_hyp_graph:
renderer = HypGraphRenderer(hyp_graph)
renderer.wordify(word_idict_trg)
renderer.save_png(return_hyp_graph, detailed=True, highlight_best=True)
order = numpy.argsort(scores)
for j in order:
if print_word_probabilities:
probs = " ||| " + " ".join("{0}".format(prob) for prob in word_probs[j])
else:
probs = ""
saveto.write('{0} ||| {1} ||| {2}{3}\n'.format(i, _seqs2words(samples[j]), scores[j], probs))
# print alignment matrix for each hypothesis
# header: sentence id ||| translation ||| score ||| source ||| source_token_count+eos translation_token_count+eos
if save_alignment is not None:
if a_json:
print_matrix_json(alignment[j], source_sentences[i], _seqs2words(samples[j]).split(), i, i+j,save_alignment)
else:
save_alignment.write('{0} ||| {1} ||| {2} ||| {3} ||| {4} {5}\n'.format(
i, _seqs2words(samples[j]), scores[j], ' '.join(source_sentences[i]) , len(source_sentences[i])+1, len(samples[j])))
print_matrix(alignment[j], save_alignment)
else:
samples, scores, word_probs, alignment, hyp_graph = trans
if return_hyp_graph:
renderer = HypGraphRenderer(hyp_graph)
renderer.wordify(word_idict_trg)
renderer.save_png(return_hyp_graph, detailed=True, highlight_best=True)
saveto.write(_seqs2words(samples) + "\n")
if print_word_probabilities:
for prob in word_probs:
saveto.write("{} ".format(prob))
saveto.write('\n')
if save_alignment is not None:
if a_json:
print_matrix_json(alignment, source_sentences[i], _seqs2words(trans[0]).split(), i, i,save_alignment)
else:
save_alignment.write('{0} ||| {1} ||| {2} ||| {3} ||| {4} {5}\n'.format(
i, _seqs2words(trans[0]), 0, ' '.join(source_sentences[i]) , len(source_sentences[i])+1, len(trans[0])))
print_matrix(alignment, save_alignment)
sys.stderr.write('Done\n')
def __init__(self, source, target,
source_dicts, target_dict,
model_type,
batch_size=128,
maxlen=100,
source_vocab_sizes=None,
target_vocab_size=None,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
use_factor=False,
maxibatch_size=20,
token_batch_size=0,
keep_data_in_memory=False):
if keep_data_in_memory:
self.source, self.target = FileWrapper(source), FileWrapper(target)
if shuffle_each_epoch:
r = numpy.random.permutation(len(self.source))
self.source.shuffle_lines(r)
self.target.shuffle_lines(r)
elif shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main([self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict, model_type))
self.target_dict = load_dict(target_dict, model_type)
# Determine the UNK value for each dictionary (the value depends on
# which version of build_dictionary.py was used).
def determine_unk_val(d):
if '<UNK>' in d and d['<UNK>'] == 2:
return 2
return 1
self.source_unk_vals = [determine_unk_val(d)
for d in self.source_dicts]
self.target_unk_val = determine_unk_val(self.target_dict)
self.keep_data_in_memory = keep_data_in_memory
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.use_factor = use_factor
self.source_vocab_sizes = source_vocab_sizes
self.target_vocab_size = target_vocab_size
self.token_batch_size = token_batch_size
if self.source_vocab_sizes != None:
assert len(self.source_vocab_sizes) == len(self.source_dicts)
for d, vocab_size in zip(self.source_dicts, self.source_vocab_sizes):
if vocab_size != None and vocab_size > 0:
for key, idx in list(d.items()):
if idx >= vocab_size:
del d[key]
if self.target_vocab_size != None and self.target_vocab_size > 0:
for key, idx in list(self.target_dict.items()):
if idx >= self.target_vocab_size:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def __init__(self, source, target,
source_dicts, target_dict,
batch_size=128,
maxlen=100,
n_words_source=-1,
n_words_target=-1,
skip_empty=False,
shuffle_each_epoch=False,
sort_by_length=True,
use_factor=False,
maxibatch_size=20):
if shuffle_each_epoch:
self.source_orig = source
self.target_orig = target
self.source, self.target = shuffle.main([self.source_orig, self.target_orig], temporary=True)
else:
self.source = fopen(source, 'r')
self.target = fopen(target, 'r')
print 'scan the dataset.'
for si, _ in enumerate(self.source):
pass
for ti, _ in enumerate(self.target):
pass
self.source.seek(0)
self.target.seek(0)
assert si == ti, 'the number of the source and target document must the same'
print 'scanned {} lines'.format(si)
self.source_dicts = []
for source_dict in source_dicts:
self.source_dicts.append(load_dict(source_dict))
self.target_dict = load_dict(target_dict)
self.batch_size = batch_size
self.maxlen = maxlen
self.skip_empty = skip_empty
self.use_factor = use_factor
self.n_words_source = n_words_source
self.n_words_target = n_words_target
if self.n_words_source > 0:
for d in self.source_dicts:
for key, idx in d.items():
if idx >= self.n_words_source:
del d[key]
if self.n_words_target > 0:
for key, idx in self.target_dict.items():
if idx >= self.n_words_target:
del self.target_dict[key]
self.shuffle = shuffle_each_epoch
self.sort_by_length = sort_by_length
self.source_buffer = []
self.target_buffer = []
self.k = batch_size * maxibatch_size
self.end_of_data = False
def load_alias(self):
''' load an alias dict '''
self.alias_st = util.load_dict(self.filename_alias_st)
self.alias_bigram = util.load_dict(self.filename_alias_bigram)