我正在通过tensorflow实现定制的成对损失函数。举一个简单的例子,训练数据有5个实例,其标签是
y=[0,1,0,0,0]
假设预测是
y'=[y0',y1',y2',y3',y4']
在这种情况下,可以使用简单的损失函数
min f=(y0'-y1')+(y2'-y1')+(y3'-y1')+(y4'-y1')
自y[1]=1
以来。我只是想确保预测y0',y2',y3',y4'
为"远"为y1'
。
但是,我不知道如何在tensorflow中实现它。在我目前的实现中,我使用迷你批处理并将训练标签设置为占位符,如:
y = tf.placeholder("float", [None, 1])
。在这种情况下,我无法构建损失函数,因为我不知道训练数据的大小以及哪个实例有标签" 1"或" 0"由于"无"。
有谁能建议如何在tensorflow中做到这一点?谢谢!
答案 0 :(得分:2)
您可以预处理数据在模型外
。例如:
首先将正面和负面实例分成两组输入:
# data.py
import random
def load_data(data_x, data_y):
"""
data_x: list of all instances
data_y: list of their labels
"""
pos_x = []
neg_x = []
for x, y in zip(data_x, data_y):
if y == 1:
pos_x.append(x)
else:
neg_x.append(x)
ret_pos_x = []
ret_neg_x = []
# randomly sample k negative instances for each positive one
for x0 in pos_x:
for x1 in random.sample(neg_x, k):
ret_pos_x.append(x0)
ret_neg_x.append(x1)
return ret_pos_x, ret_neg_x
接下来,在您的模型中,定义2个占位符,而不是1:
# model.py
import tensorflow as tf
class Model:
def __init__(self):
# shape: [batch_size, dim_x] (assume x are vectors of dim_x)
self.pos_x = tf.placeholder(tf.float32, [None, dim_x])
self.neg_x = tf.placeholder(tf.float32, [None, dim_x])
# shape: [batch_size]
# NOTE: variables in some_func should be reused
self.pos_y = some_func(self.pos_x)
self.neg_y = some_func(self.neg_x)
# A more generalized form: loss = max(0, margin - y+ + y-)
self.loss = tf.reduce_mean(tf.maximum(0.0, 1.0 - self.pos_y + self.neg_y))
self.train_op = tf.train.AdamOptimizer(learning_rate).minimize(self.loss)
最后遍历您的数据以提供模型:
# main.py
import tensorflow as tf
from model import Model
from data import load_data
data_x, data_y = ... # read from your file
pos_x, neg_x = load_data(data_x, data_y)
model = Model()
with tf.Session() as sess:
# TODO: randomize the order
for beg in range(0, len(pos_x), batch_size):
end = min(beg + batch_size, len(pos_x))
feed_dict = {
model.pos_x: pos_x[beg:end],
model.neg_x: neg_x[beg:end]
}
_, loss = sess.run([model.train_op, model.loss], feed_dict)
print "%s/%s, loss = %s" % (beg, len(pos_x), loss)
答案 1 :(得分:0)
假设我们有标签,例如y=[0,1,0,0,0]
,
将其转换为Y=[-1,1,-1,-1,-1]
。
预测为y'=[y0',y1',y2',y3',y4']
,
所以,目标是min f = -mean(Y*y')
请注意,上述公式相当于您的陈述。