<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[OliverChrist的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/OliverChrist</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; OliverChrist]]></copyright><item><title><![CDATA[深度学习在排序模型的核心操作]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/108183070</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/108183070</guid><author>OliverChrist</author><pubDate>Sun, 23 Aug 2020 14:52:55 +0800</pubDate><description><![CDATA[从2016年google的wide&amp;deep和youtube dnn两篇标志性的paper开始，深度学习在推荐系统中的应用已经遍地开花，个人觉得可以简单粗暴地将其归纳为两大阶段，第一阶段是dnn在推荐系统中的可行性验证并规模化应用，如wide&amp;deep/deepfm/deepcross等的大规模落地；第二阶段是各大公司在dnn的基础上，结合自家的业务特点，做模型二次创新，如阿里结合综合电商场景门类大而全的特点，研发了DIN系列模型。

我们来聊聊第一阶段，dnn在推荐系统落地的核心操作之一。]]></description><category></category></item><item><title><![CDATA[关于wide&deep的再思考]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/108173235</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/108173235</guid><author>OliverChrist</author><pubDate>Sat, 22 Aug 2020 19:31:46 +0800</pubDate><description><![CDATA[如果一个item在全局上被点击(或其他正反馈行为)次数过少，在排序侧把item id粒度特征直接喂入dnn做embedding，由于样本量过少，此类稀疏id特征对应的模型参数很难收敛。

干脆把item id这个field的embedding size调小呢？看似解决了稀疏特征参数收敛的问题，但却引来了新的问题：对于有充足正反馈的那部分id来说，需要用较大容量的隐向量表达语义，现在embedding size被调小了，显然限制了这部分id特征的表达。

如果此时把id特征只放到wide&amp;deep的l]]></description><category></category></item><item><title><![CDATA[spark scala udf使用方法]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/86702882</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/86702882</guid><author>OliverChrist</author><pubDate>Wed, 30 Jan 2019 14:40:28 +0800</pubDate><description><![CDATA[1.初始化dataset

val dataset = Seq((1,&quot;java&quot;),(2,&quot;python&quot;)).toDF(&quot;id&quot;,&quot;text&quot;)

2.定义udf函数

val upper: String=&amp;gt;String=_.toUpperCase

3.注册udf函数

import org.apache.spark.sql.functions.udf

val upperUDF = ...]]></description><category></category></item><item><title><![CDATA[树模型几个知识点]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/84791427</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/84791427</guid><author>OliverChrist</author><pubDate>Tue, 04 Dec 2018 16:54:04 +0800</pubDate><description><![CDATA[ID3无法处理连续特征

C4.5可以处理连续特征，用信息增益率选择分裂特征，遇到连续特征的时候，依次二分样本，根据信息增益获得最佳分割点

ID3和C4.5都只能处理分类问题，CART既能处理分类问题，又能处理回归问题

CART分割点的选择粒度更细，根据GINI系数，选择某个特征的某个值作为分割点（分为等于该值和不等于该值的二叉树）

分类树的叶子节点类别是由数量最多的种类决定的；回归树的叶子...]]></description><category></category></item><item><title><![CDATA[Dataset读取tfrecord]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/81513444</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/81513444</guid><author>OliverChrist</author><pubDate>Wed, 08 Aug 2018 18:40:25 +0800</pubDate><description><![CDATA[
    def tfrecord_pipeline(cls, tfrecord_file,  batch_size, prebatch,
                          epochs, shuffle=True):
        ''获取tfrecord配置文件''
        # tfrecord file should be a text file with abs...]]></description><category></category></item><item><title><![CDATA[基于FaceNet的人脸识别]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/80161811</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/80161811</guid><author>OliverChrist</author><pubDate>Wed, 02 May 2018 09:08:23 +0800</pubDate><description><![CDATA[1.导入包from keras.models import Sequential from keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenatefrom keras.models import Modelfrom keras.layers.normalization import BatchNormali...]]></description><category></category></item><item><title><![CDATA[深度学习的一些优化方法]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/80069648</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/80069648</guid><author>OliverChrist</author><pubDate>Tue, 24 Apr 2018 19:58:42 +0800</pubDate><description><![CDATA[过拟合数据：用更多数据训练、对于图像实施数据扩充正则化方法：L2正则化在cost function增加(lamdb/(2*m))sum(w[l]),整理偏导数得到w=(1-lamdb/m)w+learning_rate*back_prop,相当于对w乘以（1-lamdb/m）,缩小了w的范围，结合tanh的曲线，在w较小的范围里更接近线性缺点：通过调节超参数lamdb需要较高的计算成本Dropou...]]></description><category></category></item><item><title><![CDATA[train yolo on voc遇到的问题]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79990719</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79990719</guid><author>OliverChrist</author><pubDate>Wed, 18 Apr 2018 15:18:02 +0800</pubDate><description><![CDATA[1.按照官方步骤操作，https://pjreddie.com/darknet/yolo/2.提交命令训练YOLOV3 on voc./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74（其他省略的步骤都是按照官方指导做的，而且做了两遍都不行）刚开始只是少量打印nan训练一会儿之后，大量打印nan执行以下...]]></description><category></category></item><item><title><![CDATA[L2正则化的TensorFlow实现]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79659094</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79659094</guid><author>OliverChrist</author><pubDate>Thu, 22 Mar 2018 20:06:52 +0800</pubDate><description><![CDATA[TensorFlow实现L2正则化

转自：https://zhuanlan.zhihu.com/p/29297934

TensorFlow的最优化方法tf.train.GradientDescentOptimizer包办了梯度下降、反向传播，所以基于TensorFlow实现L2正则化，并不能按照上节的算法直接干预权重的更新，而要使用TensorFlow方式：

tf.add_to_collec...]]></description><category></category></item><item><title><![CDATA[【深度学习_4.4】人脸识别算法模型搭建-Face Verification和Face Recognition]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79490094</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79490094</guid><author>OliverChrist</author><pubDate>Thu, 08 Mar 2018 23:07:30 +0800</pubDate><description><![CDATA[人脸识别分为两大类：
Face Verification指的是，这张脸是不是xxx？

Face Recognition指的是，这张脸是谁？
FaceNet是一种神经网络，把一张人脸照片编码成为128组向量，通过比较两张照片的128组向量来确定是不是同一个人。
本实例中使用训练好的ConvNet激活函数，应用的是channel first的形式（m, nc, nw, nh）

将人脸]]></description><category></category></item><item><title><![CDATA[YOLO算法总结]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79472683</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79472683</guid><author>OliverChrist</author><pubDate>Wed, 07 Mar 2018 16:21:28 +0800</pubDate><description><![CDATA[步骤：
1.输入照片（608,608,3）
2.经过CNN处理，输出（19,19,5,85）
3.展开以上矩阵后为（19,,19,425）
每个19*19的网格里有425个数字
425=5*85，每个网格里有5个achor box，achor box中有85个数字
85=5+80，（pc,bx,by,bw,bh,c）,c有80个种类，pc代表这个box里面是否detect到object]]></description><category></category></item><item><title><![CDATA[【深度学习_4.3】构建YOLO物体识别算法]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79466299</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79466299</guid><author>OliverChrist</author><pubDate>Wed, 07 Mar 2018 16:03:25 +0800</pubDate><description><![CDATA[训练集里面的训练图片被标记如下


如果YOLO算法需要识别80种物体，那么c可以是1-80之间的任意整数，也可以是80维的向量，识别出的物体为1，其他均为零。
YOLO算法模型
输入（m,608,608,3）
输出是识别出来的物体被边框（pc,bx,by,bw,bh,c）,加入c是一个80维的向量，则每个边框有80个代表值
示例中将使用5个achors box，因此模型为IMAGE]]></description><category></category></item><item><title><![CDATA[【深度学习_4.2】Keras构建残差神经网络]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79450104</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79450104</guid><author>OliverChrist</author><pubDate>Mon, 05 Mar 2018 20:08:38 +0800</pubDate><description><![CDATA[残差神经网络可帮助避免多层神经网络的梯度消失（主要解决的问题）、梯度爆炸等现象
普通残差块




First component of main path:

The first CONV2D has F1F1 filters
 of shape (1,1) and a stride of (1,1). Its padding is "valid" and its name s]]></description><category></category></item><item><title><![CDATA[【深度学习_4.1_2】在TensorFlow中应用卷积神经网络]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79432892</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79432892</guid><author>OliverChrist</author><pubDate>Sat, 03 Mar 2018 20:56:57 +0800</pubDate><description><![CDATA[1.创建placeholder（暂不定义训练集数量）
def create_placeholders(n_H0, n_W0, n_C0, n_y):
    “”
    Arguments:
    n_H0 -- scalar, height of an input image
    n_W0 -- scalar, width of an input image
    n_C0]]></description><category></category></item><item><title><![CDATA[【深度学习_4.1_1】构建卷积神经网络]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79432305</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79432305</guid><author>OliverChrist</author><pubDate>Sat, 03 Mar 2018 20:43:14 +0800</pubDate><description><![CDATA[构建卷积神经网络整体步骤
一、卷积
1.Zero padding
2.Convole window
3.Convolution forward
4.Convolution backward(Optional)
二、池化
1.Polling forward
2.create mask
3.distribute values
4.Pooling backward(Optional)]]></description><category></category></item><item><title><![CDATA[【深度学习_2.3_2】TensorFlow搭建神经网络模型]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79397406</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79397406</guid><author>OliverChrist</author><pubDate>Wed, 28 Feb 2018 11:01:47 +0800</pubDate><description><![CDATA[模型：通过1-5的手势图片训练
 LINEAR -> RELU -> LINEAR -> RELU -> LINEAR -> SOFTMAX

处理输入数据集
X_train_flatten = X_train_orig.reshape(X_train_orig.shape[0], -1).T
X_test_flatten = X_test_orig.reshape(X_test_ori]]></description><category></category></item><item><title><![CDATA[【深度学习_2.3_1】神经网络之TensorFlow初步应用]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79395809</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79395809</guid><author>OliverChrist</author><pubDate>Wed, 28 Feb 2018 10:01:41 +0800</pubDate><description><![CDATA[导入相关类库
import math
import numpy as np
import h5py
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.python.framework import ops
from tf_utils import load_dataset, random_m]]></description><category></category></item><item><title><![CDATA[【深度学习_2.2】神经网络之算法优化]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79380880</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79380880</guid><author>OliverChrist</author><pubDate>Mon, 26 Feb 2018 20:23:08 +0800</pubDate><description><![CDATA[通过梯度下降可以更新参数并最小化cost函数，通过算法优化可以加速学习并可能会得到更优的cost函数值一、梯度下降L是神经网络的层数，α代表学习率， 所有参数都存在parameters字典里    Arguments:    parameters -- python dictionary containing your parameters to be updated:              ...]]></description><category></category></item><item><title><![CDATA[【深度学习_2.1.3】神经网络之梯度检验]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79370429</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79370429</guid><author>OliverChrist</author><pubDate>Sun, 25 Feb 2018 18:31:10 +0800</pubDate><description><![CDATA[梯度检验用于检验后向传播正常工作
一维数组模型梯度检验
前向传播，计算损失函数：
J = theta * x

后向传播：
dtheta = x

梯度检验：
首先计算gradapprox：


    thetaplus = theta + epsilon                               
    thetaminus = theta - e]]></description><category></category></item><item><title><![CDATA[【深度学习_2.1.2】神经网络正则化]]></title><link>https://blog.youkuaiyun.com/OliverChrist/article/details/79369861</link><guid>https://blog.youkuaiyun.com/OliverChrist/article/details/79369861</guid><author>OliverChrist</author><pubDate>Sun, 25 Feb 2018 17:43:22 +0800</pubDate><description><![CDATA[偏差大（欠拟合）：训练集和测试集的准确率都很低；通常可以增加单层神经元数量、增加层数来优化
方差大（过拟合）：训练集的正确率高，测试集（或dev）的正确率比训练集低一些；通常可以增大正则化lambda的值、增加训练集数据量解决
可以通过在训练模型过程中正则化来解决过拟合问题


l2正则化，在原来的cost function基础上添加l2 regularization cost]]></description><category></category></item></channel></rss>