Question

我想将共同出现作为损失函数的先验知识。像这样的共生矩阵， enter image description here

我的网络是这样的，

# x = tf.placeholder(tf.float32, [None, 745])
x = tf.placeholder(tf.float32)
W1 = tf.Variable(tf.truncated_normal([2056, 500], stddev=0.1))
b1 = tf.Variable(tf.zeros([500]))
layer1 = tf.nn.relu(tf.matmul(x, W1) + b1)#隐层
# layer1 = tf.clip_by_value(tf.matmul(x, W1) + b1)#隐层

W2 = tf.Variable(tf.truncated_normal([500,33], stddev=0.1))
b2 = tf.Variable(tf.zeros([33]))
# layer2 = tf.nn.tanh(tf.matmul(layer1, W2) + b2)
# W3 = tf.Variable(tf.truncated_normal([500, 11], stddev=0.1))
# b3 = tf.Variable(tf.zeros([11]))
y_without=tf.matmul(layer1, W2) + b2
y = tf.nn.sigmoid(y_without)#输出层，sigmoid函数
# y_ = tf.placeholder(tf.float32, [None, 101])
y_ = tf.placeholder(tf.float32)

我想计算W2的每两列之间的相似性，计算函数如下，

def frob(w1,w2):
    asa=(tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(w1,w2)))))
    return(asa)
def occu(W2):
    each_mul=[]
#     total_occu=tf.Variable(tf.constant(0.0))
    for i in range(33):
        for j in range(33):
            num=float(co[i,j])
            frob_pair=frob((W2)[:,i],(W2)[:,j])
            mul=tf.multiply(num,frob_pair)
            each_mul.append(mul)
    wer=tf.reduce_sum(each_mul)
    wer/=40000
#     print(each_mul)
    return wer#此时p等于0.01

像这样训练一步，

    cross_entropy=tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=y_,logits=y_without))
coco=occu(W2)
loss = tf.add(cross_entropy , coco)
train_step = tf.train.GradientDescentOptimizer(0.05).minimize(loss)
examples, labels = input_pipeline(200,20)
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess = tf.Session()
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
train_accu=[]
test_accu=[]
train_loss=[]
saver=tf.train.Saver()
now=time.time()
try:
    i = 0
    while not coord.should_stop():
        i = i + 1
        example_batch, label_batch = sess.run([examples, labels])
#         aa=sess.run(example_batch)
#         print(example_batch)
#         break
        print(sess.run(coco))
        sess.run(train_step, feed_dict={x: example_batch, y_: label_batch}) 
        print(sess.run(coco))
        train_loss.append(sess.run(loss,feed_dict={x:example_batch,y_:label_batch}))
        train_accu.append(hit3accu(sess.run(y,feed_dict={x:example_batch}),label_batch))
        test_accu.append(hit3accu(sess.run(y,feed_dict={x:X_test_final_accu}),y_test_accu))
#         print(np.sum((label_batch[:,18])))
        if i % 1 == 0:
#             print(sess.run(occu(W2)))

#             summary = sess.run(merged, feed_dict={x: example_batch, y_: label_batch})

            print('iter:%s,train_loss:%s,train_accu:%s,test_accu:%s'%(i,train_loss[-1],train_accu[-1],test_accu[-1]))
        if i %1000==0:
            model_path = "F:\co-occurence\saved\model{}.ckpt".format(i) 
            saver.save(sess,model_path)
#             trainwriter.add_summary(summary,i)             
#         print(i)       
#         break
except tf.errors.OutOfRangeError:
    print('Done training -- epoch limit reached')
finally:
    # When done, ask the threads to stop.
    coord.request_stop()

# Wait for threads to finish.
coord.join(threads)
duration=time.time()-now
print(duration)

然而，我得到了错误 enter image description here 似乎在第一个W2被正确初始化，但是在一个时代之后W2变成了南。如果有人能给我一些关于这个问题的想法，我将非常感激。

如何使用张量流将共生先验知识添加到nn的损失函数中

0 个答案: