def __init__(self, *args, **kwargs):
""""""
super(Network, self).__init__(*args, **kwargs)
# hacky!
#hacky_train_files = op.join(self.save_dir, op.basename(self.get("train_files")))
#self._config.set('Configurable', 'train_files', hacky_train_files)
# TODO make this more flexible, maybe specify it in config?
temp_nlp_model = self.nlp_model.from_configurable(self)
if temp_nlp_model.input_vocabs == ['tags']:
word_vocab = WordVocab.from_configurable(self)
word_multivocab = Multivocab.from_configurable(self, [word_vocab], name=word_vocab.name)
tag_vocab = TagVocab.from_configurable(self, initialize_zero=False)
else:
word_vocab = WordVocab.from_configurable(self)
pretrained_vocab = PretrainedVocab.from_vocab(word_vocab)
subtoken_vocab = self.subtoken_vocab.from_vocab(word_vocab)
word_multivocab = Multivocab.from_configurable(self, [word_vocab, pretrained_vocab, subtoken_vocab], name=word_vocab.name)
#word_multivocab = Multivocab.from_configurable(self, [word_vocab, pretrained_vocab], name=word_vocab.name)
tag_vocab = TagVocab.from_configurable(self)
dep_vocab = DepVocab.from_configurable(self)
lemma_vocab = LemmaVocab.from_configurable(self)
xtag_vocab = XTagVocab.from_configurable(self)
head_vocab = HeadVocab.from_configurable(self)
rel_vocab = RelVocab.from_configurable(self)
self._vocabs = [dep_vocab, word_multivocab, lemma_vocab, tag_vocab, xtag_vocab, head_vocab, rel_vocab]
self._global_step = tf.Variable(0., trainable=False, name='global_step')
self._global_epoch = tf.Variable(0., trainable=False, name='global_epoch')
self._optimizer = RadamOptimizer.from_configurable(self, global_step=self.global_step)
return
def __call__(self):
""""""
#création d'un optimizer RAdam
radam_optimizer = RadamOptimizer.from_configurable(self,
learning_rate=1e-1,
decay_steps=500)
x = tf.placeholder(tf.float32, shape=(None, 1), name='x')
y = tf.placeholder(tf.float32, shape=(None, 1), name='y')
def affine(a, b):
return a * tf.log(x) + b
a = tf.get_variable('a',
shape=self.n_zipfs,
dtype=tf.float32,
initializer=tf.random_normal_initializer())
b = tf.get_variable('b',
shape=self.n_zipfs,
dtype=tf.float32,
initializer=tf.random_normal_initializer())
s = tf.get_variable('s',
shape=self.n_zipfs,
dtype=tf.float32,
initializer=tf.random_uniform_initializer(-2, -.5))
t = tf.get_variable('t',
shape=self.n_zipfs,
dtype=tf.float32,
initializer=tf.random_normal_initializer())
w = tf.expand_dims(tf.nn.softmax(affine(a, b)), axis=1, name='w')
z = tf.expand_dims(affine(s, t), axis=2, name='z')
yhat = tf.squeeze(tf.matmul(w, z), axis=2, name='yhat')
ell = tf.reduce_mean((tf.log(y) - yhat)**2 / 2, name='ell')
ell += tf.reduce_mean((tf.reduce_max(w, axis=0) - 1)**2 / 2)
minimize = radam_optimizer.minimize(ell, name='minimize')
return x, y, ell, minimize
def __init__(self, train=False, *args, **kwargs):
""""""
super(Network, self).__init__(*args, **kwargs)
# hacky!
#hacky_train_files = op.join(self.save_dir, op.basename(self.get("train_files")))
#self._config.set('Configurable', 'train_files', hacky_train_files)
#todo: essayer de reserver tt le gpu avant toute chose, si ça n'est pas possible on passe au prochain gpu
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=1)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
temp_nlp_model = self.nlp_model.from_configurable(self)
print("\nloading word vocab")
word_vocab = WordVocab.from_configurable(self)
print("\nloading pretrained vocab")
pretrained_vocab = PretrainedVocab.from_vocab(word_vocab)
print("\nloading tag representation vocab")
tagrep_vocab = None
if 'tagrep' in temp_nlp_model.input_vocabs:
tagrep_vocab = TagRepVocab.from_vocab(word_vocab)
print("\nloading subtoken vocab")
subtoken_vocab = self.subtoken_vocab.from_vocab(word_vocab)
print("\ncreating multivocab")
vocabs_of_multivocab = [word_vocab, pretrained_vocab]
word_multivocab = Multivocab.from_configurable(self,
vocabs_of_multivocab,
name=word_vocab.name)
if ('chars' in temp_nlp_model.input_vocabs
or word_multivocab.merge_strategy != "only_pretrained"):
vocabs_of_multivocab.append(subtoken_vocab)
word_multivocab = Multivocab.from_configurable(
self, vocabs_of_multivocab, name=word_vocab.name)
print(word_multivocab.merge_strategy)
print('chars' in temp_nlp_model.input_vocabs)
print(word_multivocab.merge_strategy != "only_pretrained")
print("\nloading tag vocab")
tag_vocab = None
if 'tags' in temp_nlp_model.input_vocabs or 'tags' in temp_nlp_model.output_vocabs:
tag_vocab = TagVocab.from_configurable(self)
print("\nloading dep vocab")
dep_vocab = DepVocab.from_configurable(self)
print("\nloading lemma vocab")
lemma_vocab = LemmaVocab.from_configurable(self)
print("\nloading xtag vocab")
xtag_vocab = None
if 'xtags' in temp_nlp_model.input_vocabs or 'xtags' in temp_nlp_model.output_vocabs or 'subxtags' in temp_nlp_model.input_vocabs or 'subxtags' in temp_nlp_model.output_vocabs:
xtag_vocab = XTagVocab.from_configurable(self)
print("\nloading feat vocab")
feat_vocab = None
if 'feats' in temp_nlp_model.input_vocabs or 'feats' in temp_nlp_model.output_vocabs or 'subfeats' in temp_nlp_model.input_vocabs or 'subfeats' in temp_nlp_model.output_vocabs:
feat_vocab = FeatVocab.from_configurable(self)
print("\nloading subfeat vocab")
subfeat_vocabs = None
if 'subfeats' in temp_nlp_model.input_vocabs or 'subfeats' in temp_nlp_model.output_vocabs:
subfeat_vocabs = self.prepare_subpos_vocabs(
feat_vocab, temp_nlp_model)
print("\nloading subxtag vocab")
subxtag_vocabs = None
if 'subxtags' in temp_nlp_model.input_vocabs or 'subxtags' in temp_nlp_model.output_vocabs:
subxtag_vocabs = self.prepare_subpos_vocabs(
xtag_vocab, temp_nlp_model)
head_vocab = HeadVocab.from_configurable(self)
rel_vocab = RelVocab.from_configurable(self)
self._vocabs = [
dep_vocab, word_multivocab, lemma_vocab, tag_vocab, xtag_vocab,
feat_vocab, head_vocab, rel_vocab, tagrep_vocab
]
if subfeat_vocabs != None:
self._vocabs.extend(subfeat_vocabs)
if subxtag_vocabs != None:
self._vocabs.extend(subxtag_vocabs)
self._vocabs = filter(None, self._vocabs)
print(self._vocabs)
#print('subfeat_vocabs:',len(subfeat_vocabs))
#print('subxtag_vocabs:',len(subxtag_vocabs))
self._global_step = tf.Variable(0.,
trainable=False,
name='global_step')
self._global_epoch = tf.Variable(0.,
trainable=False,
name='global_epoch')
self._optimizer = RadamOptimizer.from_configurable(
self, global_step=self.global_step)
if train:
self._global_time = tf.Variable(0.,
trainable=False,
name='global_time')
return