def predict(self, x):
embedding_size = 128
hidden_size = 128
ff_size = 300
output_size = [2, 7, 1, 1, 1, 1, 1, 17, 209, 4, 19, 20, 9, 16, 12, 14]
embedded_features = []
inputs = tf.split(x, num_or_size_splits=16, axis=1)
for i, x in enumerate(inputs): # x: (batch, X)
if i > 1 and i < 7: # regression (#2 ~ #7)
embedded_features.append(tf.cast(x, tf.float32))
else: # classification
embedded_features.append(
tf.reshape(
l.embedding(x, output_size[i] + 1, embedding_size,
'embed_' + str(i)), (-1, embedding_size)))
embedded_features = tf.concat(
embedded_features, axis=1) # (batch_size, 16, embedding_size)
fc1 = l.fc_layer(embedded_features, [embedding_size * 11 + 5, 300],
'fc1',
activation='relu',
batch_normal=True)
fc2 = l.fc_layer(fc1, [300, ff_size],
'fc2',
activation='relu',
batch_normal=True)
preds = []
for i in range(16):
if i > 1 and i < 7: # regression (#2 ~ #7)
preds.append(
l.fc_layer(fc2, [ff_size, 1],
'pred_' + str(i),
activation=None))
else: # classification
preds.append(
l.fc_layer(fc2, [ff_size, output_size[i]],
'pred_' + str(i),
activation='softmax'))
return preds
def predict(self, x):
embedding_size = 300
hidden_size = 128
embedded_features = []
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 0], 2 + 1, embedding_size, 'embed1'),
1)) # 주야 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 1], 7 + 1, embedding_size, 'embed2'),
1)) # 요일 + hole
reg_features = tf.cast(
tf.tile(tf.expand_dims(x[:, 2:7], -1), [1, 1, embedding_size]),
tf.float32)
embedded_features.append(tf.expand_dims(reg_features[:, 0], 1)) # 사망자수
embedded_features.append(tf.expand_dims(reg_features[:, 1], 1)) # 사상자수
embedded_features.append(tf.expand_dims(reg_features[:, 2], 1)) # 중상자수
embedded_features.append(tf.expand_dims(reg_features[:, 3], 1)) # 경상자수
embedded_features.append(tf.expand_dims(reg_features[:, 4],
1)) # 부상신고자수
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 7], 17 + 1, embedding_size, 'embed8'),
1)) # 발생지시도 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 8], 209 + 1, embedding_size, 'embed9'),
1)) # 발생지시군구 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 9], 4 + 1, embedding_size, 'embed10'),
1)) # 사고유형_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 10], 19 + 1, embedding_size, 'embed11'),
1)) # 사고유형_중분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 11], 20 + 1, embedding_size, 'embed12'),
1)) # 법규위반 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 12], 9 + 1, embedding_size, 'embed13'),
1)) # 도로형태_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 13], 16 + 1, embedding_size, 'embed14'),
1)) # 도로형태 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 14], 12 + 1, embedding_size, 'embed15'),
1)) # 당사자종별_1당_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 15], 14 + 1, embedding_size, 'embed16'),
1)) # 당사자종별_2당_대분류 + hole
embedded_features = tf.concat(
embedded_features, axis=1) # (batch_size, 16, embedding_size)
blstm_features = l.blstm(embedded_features, hidden_size,
'blstm') # (batch_size, 16, embedding_size*2)
blstm_features = tf.layers.batch_normalization(blstm_features)
blstm_features = l.blstm(
blstm_features, hidden_size,
'blstm2') # (batch_size, 16, embedding_size*2)
features = tf.split(blstm_features, num_or_size_splits=16, axis=1)
final_features = []
ff_size = 300
for i, f in enumerate(features):
final_features.append(
l.fc_layer(tf.reshape(f, (-1, hidden_size * 2)),
[hidden_size * 2, ff_size], 'refineFC_' + str(i)))
#blstm_features = tf.layers.batch_normalization(blstm_features)
#fc1 = l.fc_layer(tf.cast(blstm_features, tf.float32), [hidden_size*2, 200], 'fc1', activation='relu', dropout=0.5)
#feature = l.fc_layer(fc1, [200, 200], 'fc2', dropout=0.5)
output_size = [2, 7, 1, 1, 1, 1, 1, 17, 209, 4, 19, 20, 9, 16, 12, 14]
preds = []
for i, f in enumerate(final_features):
if i > 1 and i < 7: # regression (#2 ~ #7)
preds.append(
l.fc_layer(f, [ff_size, output_size[i]],
'pred_' + str(i),
activation=None))
else: # classification
preds.append(
l.fc_layer(f, [ff_size, output_size[i]],
'pred_' + str(i),
activation='softmax'))
return preds
def predict(self, x):
fc1 = l.fc_layer(x, [16, 200], 'fc1', activation='relu', dropout=0.5)
feature = l.fc_layer(fc1, [200, 200], 'fc2', dropout=0.5)
preds = []
preds.append(
l.fc_layer(feature, [200, 2], 'pred_1', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 7], 'pred_2', activation='softmax'))
preds.append(l.fc_layer(feature, [200, 5], 'pred_37', activation=None))
preds.append(
l.fc_layer(feature, [200, 17], 'pred_8', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 209], 'pred_9', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 4], 'pred_10', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 19], 'pred_11', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 20], 'pred_12', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 9], 'pred_13', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 16], 'pred_14', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 12], 'pred_15', activation='softmax'))
preds.append(
l.fc_layer(feature, [200, 14], 'pred_16', activation='softmax'))
return preds
def predict(self, x):
embedding_size = 50
hidden_size = 128
embedded_features = []
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 0], 2 + 1, embedding_size, 'embed1'),
1)) # 주야 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 1], 7 + 1, embedding_size, 'embed2'),
1)) # 요일 + hole
#print(tf.tile(tf.expand_dims(x[:, 2], 1), (1, 50)).get_shape())
embedded_features.append(
tf.cast(
tf.expand_dims(
tf.tile(tf.expand_dims(x[:, 2], 1), (1, embedding_size)),
1), tf.float32)) # 사망자수(최대값) + hole
embedded_features.append(
tf.cast(
tf.expand_dims(
tf.tile(tf.expand_dims(x[:, 3], 1), [1, embedding_size]),
1), tf.float32)) # 사상자수(최대값) + hole
embedded_features.append(
tf.cast(
tf.expand_dims(
tf.tile(tf.expand_dims(x[:, 4], 1), [1, embedding_size]),
1), tf.float32)) # 중상자수(최대값) + hole
embedded_features.append(
tf.cast(
tf.expand_dims(
tf.tile(tf.expand_dims(x[:, 5], 1), [1, embedding_size]),
1), tf.float32)) # 경상자수(최대값) + hole
embedded_features.append(
tf.cast(
tf.expand_dims(
tf.tile(tf.expand_dims(x[:, 6], 1), [1, embedding_size]),
1), tf.float32)) # 부상신고자수(최대값) + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 7], 17 + 1, embedding_size, 'embed8'),
1)) # 발생지시도 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 8], 209 + 1, embedding_size, 'embed9'),
1)) # 발생지시군구 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 9], 4 + 1, embedding_size, 'embed10'),
1)) # 사고유형_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 10], 19 + 1, embedding_size, 'embed11'),
1)) # 사고유형_중분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 11], 20 + 1, embedding_size, 'embed12'),
1)) # 법규위반 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 12], 9 + 1, embedding_size, 'embed13'),
1)) # 도로형태_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 13], 16 + 1, embedding_size, 'embed14'),
1)) # 도로형태 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 14], 12 + 1, embedding_size, 'embed15'),
1)) # 당사자종별_1당_대분류 + hole
embedded_features.append(
tf.expand_dims(
l.embedding(x[:, 15], 14 + 1, embedding_size, 'embed16'),
1)) # 당사자종별_2당_대분류 + hole
embedded_features = tf.concat(
embedded_features, axis=1) # (batch_size, 16, embedding_size)
print(embedded_features.get_shape())
blstm_features = l.blstm(embedded_features, hidden_size,
'blstm') # (batch_size, 16, embedding_size*2)
print(blstm_features.get_shape())
blstm_features2 = l.blstm(
blstm_features, hidden_size,
'blstm2') # (batch_size, 16, embedding_size*2)
print(blstm_features2.get_shape())
#fc1 = l.fc_layer(tf.cast(blstm_features, tf.float64), [hidden_size*4, 200], 'fc1', activation='relu', dropout=0.5)
#feature = l.fc_layer(fc1, [200, 200], 'fc2', dropout=0.5)
preds = []
preds.append(
l.fc_layer(blstm_features2[:, 0, :], [hidden_size * 2, 2],
'pred_1',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 1, :], [hidden_size * 2, 7],
'pred_2',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 2, :], [hidden_size * 2, 1],
'pred_3',
activation=None))
preds.append(
l.fc_layer(blstm_features2[:, 3, :], [hidden_size * 2, 1],
'pred_4',
activation=None))
preds.append(
l.fc_layer(blstm_features2[:, 4, :], [hidden_size * 2, 1],
'pred_5',
activation=None))
preds.append(
l.fc_layer(blstm_features2[:, 5, :], [hidden_size * 2, 1],
'pred_6',
activation=None))
preds.append(
l.fc_layer(blstm_features2[:, 6, :], [hidden_size * 2, 1],
'pred_7',
activation=None))
preds.append(
l.fc_layer(blstm_features2[:, 7, :], [hidden_size * 2, 17],
'pred_8',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 8, :], [hidden_size * 2, 209],
'pred_9',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 9, :], [hidden_size * 2, 4],
'pred_10',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 10, :], [hidden_size * 2, 19],
'pred_11',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 11, :], [hidden_size * 2, 20],
'pred_12',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 12, :], [hidden_size * 2, 9],
'pred_13',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 13, :], [hidden_size * 2, 16],
'pred_14',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 14, :], [hidden_size * 2, 12],
'pred_15',
activation='softmax'))
preds.append(
l.fc_layer(blstm_features2[:, 15, :], [hidden_size * 2, 14],
'pred_16',
activation='softmax'))
return preds