def __init__(self, filters=200, dyn_out_len=100, name: str = ''):
## PARAMETERS
# Embeddings
# vocab_size = 4092
entity_vocab_size = 20
pos_vocab_size = 20
# word_emb_dim = 50
entity_emb_dim = 10
pos_emb_dim = 10
# Conv1D
# filters = 20
conv_kernel_size = 1
conv_strides = 1
# Pooling
# dyn_out_len = 10
self.parameters = (name, entity_vocab_size, pos_vocab_size,
entity_emb_dim, pos_emb_dim, filters,
conv_kernel_size, conv_strides, dyn_out_len)
## LAYERS
# Input
tokens = Input(batch_shape=(1, None, 51))
relpos1 = Input(batch_shape=(1, None, 1))
relpos2 = Input(batch_shape=(1, None, 1))
entity = Input(batch_shape=(1, None))
pos = Input(batch_shape=(1, None))
lemma = Input(batch_shape=(1, None, 51))
# Embedding
# word_emb = Embedding(vocab_size, word_emb_dim)(tokens)
entity_emb = Embedding(entity_vocab_size, entity_emb_dim)(entity)
pos_emb = Embedding(pos_vocab_size, pos_emb_dim)(pos)
# lemma_emb = Embedding(vocab_size, word_emb_dim)(lemma)
# Base-model
# x = concatenate([word_emb, relpos1, relpos2, entity_emb, pos_emb, lemma_emb], axis=2)
x = Concatenate(axis=2)(
[tokens, relpos1, relpos2, entity_emb, pos_emb, lemma])
x = Dropout(0.5)(x)
x = Conv1D(filters=filters,
kernel_size=conv_kernel_size,
strides=conv_strides,
activation='relu')(x)
x = Dynamic_max_pooling(filters, dyn_out_len)(x)
x = Flatten()(x)
x = Dropout(0.5)(x)
x = Dense(4, activation='softmax')(
x) # 4 == n_classes in data (i.e. NOT a parameter)
## Create model
model = Model(inputs=[tokens, relpos1, relpos2, entity, pos, lemma],
outputs=x)
# model.summary()
model.compile(
loss=
categorical_crossentropy, # for onehot, for idx: sparse_categorical_crossentropy
optimizer='adam',
metrics=['accuracy'])
self.model = model
self.base_dir = os.path.join(
os.getcwd(), 'Extra_files/resources/', self.__class__.__name__ +
'_' + str(filters) + '_' + str(dyn_out_len))
os.makedirs(self.base_dir, exist_ok=True)
dir_ = self.base_dir
names = [x for x in os.listdir(dir_)]
names.sort(key=lambda x: int(x.split('_')[0]))
h = hash(self.parameters)
dir_set = False
num = 0
for x in names:
num, hash_ = x.split('_')
if h == int(hash_):
dir_ = os.path.join(dir_, x)
dir_set = True
break
if not dir_set:
dir_ = os.path.join(dir_, str(int(num) + 1) + '_' + str(h))
# print('saving to: {}'.format(dir_))
self.final_dir = dir_
os.makedirs(self.final_dir, exist_ok=True)
