<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[tjjyqing的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/tjjyqing</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; tjjyqing]]></copyright><item><title><![CDATA[pytorch中WGAN的实战以及代码注释]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/116838339</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/116838339</guid><author>tjjyqing</author><pubDate>Sat, 15 May 2021 11:05:42 +0800</pubDate><description><![CDATA[WGAN
这里是学习过程中WGAN判别八个点高斯分布的实战，代码是照着教学视频抄的，中文注释是自己写的，希望对大家理解有帮助。
import torch
from torch import nn, optim, autograd
import numpy as np
import visdom
from torch.nn import functional as F
from matplotlib import pyplot as plt
import random

h_dim = 400
batchsz]]></description><category></category></item><item><title><![CDATA[pytoch_AE&VAE]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/116785693</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/116785693</guid><author>tjjyqing</author><pubDate>Fri, 14 May 2021 11:50:39 +0800</pubDate><description><![CDATA[AE和VAE重现手写数字数据集AEVAEmain函数
下面分别是AE和VAE的模型，以及训练用的main函数
AE
import torch
from torch import nn


class AE(nn.Module):
    def __init__(self):
        super(AE, self).__init__()
        # [b, 784] mnist数据集像素28*28
        self.encoder = nn.Sequential(
         ]]></description><category></category></item><item><title><![CDATA[pytorch数据预处理和卷积神经网络训练]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/116643917</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/116643917</guid><author>tjjyqing</author><pubDate>Tue, 11 May 2021 10:36:18 +0800</pubDate><description><![CDATA[目录数据集建立神经网络ResNet模型网络训练迁移学习
数据集建立
在实际情况中，我们拥有的图片数据往往不是那么规整的，图片的大小以及图片的格式往往都会不同。我们要将这些在不同文件中间的图片，加载到一个固定的数据集的类别中，以便我们之后调用。
这里利用Pokemon数据，可以从百度上搜到，我们定义一个class类来存放这些数据。具体代码如下所示。
注：main（）函数仅仅用来查看当前class是否有误，检查其运行状况。
注：实例中用到了torch，cuda， visdom等需要手动下载的包； visdom也]]></description><category></category></item><item><title><![CDATA[卷积神经网络-pytorch-lenet5&&resnet]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/116426112</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/116426112</guid><author>tjjyqing</author><pubDate>Wed, 05 May 2021 15:50:07 +0800</pubDate><description><![CDATA[卷积神经网络pytorch实现一、模型建立1 Lenet5卷积神经网络模型2 Resnet卷积神经网络模型二、模型调用
一、模型建立
本代码基于pytorch以及cuda的GPU模式，如果没有GPU的话，把to.devide()这种代码删去就好。
下面仅仅是代码的实现，不对模型建立做解释，相关的代码解释有注释，希望对大家有帮助。
1 Lenet5卷积神经网络模型
需要先运行模型，然后下面调用才能成功。其次，模型代码文件需与运行的代码文件在同一级同一目录下。
from torch import nn


cl]]></description><category></category></item><item><title><![CDATA[pytorch利用梯度下降求方程的最小值_day2]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115695694</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115695694</guid><author>tjjyqing</author><pubDate>Wed, 14 Apr 2021 12:28:06 +0800</pubDate><description><![CDATA[pytorch

我们需要求解上面函数f（x,y）的最小值，方法如下。
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
from matplotlib import pyplot as plt
import torch


def himmelblau(x):  # 利用梯度下降计算这个函数的最小值
    return (x[0] ** 2 + x[1] - 11) ** 2 + (x[0] + x[1] ** 2 - 7) ** 2
]]></description><category></category></item><item><title><![CDATA[Pytorch_gradient descent_day 1]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115589063</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115589063</guid><author>tjjyqing</author><pubDate>Sun, 11 Apr 2021 10:10:44 +0800</pubDate><description><![CDATA[下面演示的是计算梯度下降求解线性拟合的问题
import numpy as np

# y = wx + b


def compute_error_for_line_given_points(b, w, points):
"""这个是用来计算给出w，b的误差（也叫loss function）"""
    totalError = 0
    for i in range(0, len(points)):  # 对每个点都要算
        x = points[i, 0]
        y = po]]></description><category></category></item><item><title><![CDATA[DL_network_compression_Day12]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115281873</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115281873</guid><author>tjjyqing</author><pubDate>Sun, 28 Mar 2021 15:52:31 +0800</pubDate><description><![CDATA[Network Compression
Introduction
神经网络的计算需要大量的内存或者显存，这对运行设备的要求比较高。但在某些小型电子设备上运行时不能满足这些要求，所以我们可以考虑将神经网络压缩的更小一点。
Network Pruning
神经网络中有些参数或者说神经单元是可有可无的，他们对训练的结果没有什么影响，所以扔掉这些神经元可以提高运行速度且不改变结果。下图是那人类成年后神经元数量做对比。

该方法的步骤如下

方法体现

Knowledge Distillation
除了扔掉部分神经元]]></description><category></category></item><item><title><![CDATA[DL_universal_attack_day11]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115229937</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115229937</guid><author>tjjyqing</author><pubDate>Fri, 26 Mar 2021 10:48:12 +0800</pubDate><description><![CDATA[目录Universal AttackHow to attackintroductionloss function for attackconstraintprocessattack approacheswhite or black boxblack box attackuniversal attackattack in the real worldDefensepre-defensepassive defensemore
Universal Attack
How to attack
introduction]]></description><category></category></item><item><title><![CDATA[DL_explainableML_Day10]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115115126</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115115126</guid><author>tjjyqing</author><pubDate>Tue, 23 Mar 2021 12:23:55 +0800</pubDate><description><![CDATA[目录Explainable Machine LearningClassificationLocal ExplanationGlobal ExplanationLocal Explanation: Explain the DecisionIntroductionLimitation of Gradient based ApproachesAttack InterpretationGlobal Explanation: Explain the whole ModelActivation Minimization]]></description><category></category></item><item><title><![CDATA[DL_semi-supervised_Day9]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115079477</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115079477</guid><author>tjjyqing</author><pubDate>Mon, 22 Mar 2021 15:47:22 +0800</pubDate><description><![CDATA[目录Semi-supervised LearningIntroductionSemi-supervised Generative ModelLow-density SeparationIntroductionSelf-trainingEntropy-based RegularizationSemi-supervised SVMSmoothness AssumptionIntroductionCluster and then LabelGraph-based Approach
Semi-supervised ]]></description><category></category></item><item><title><![CDATA[DL_ELMO/BERT/Transformer_Day8笔记]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/115008998</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/115008998</guid><author>tjjyqing</author><pubDate>Fri, 19 Mar 2021 17:16:40 +0800</pubDate><description><![CDATA[ELMO
含义：Embeddings from Language Model
embedding就是对单词的一个向量化
下面是训练ELMO的方法


BERT
Bidirectional Encoder (双向编码)  Representations(表示) from Transformers
BERT的四种使用方法




Transformer
结构

Transformer是第一个用纯attention搭建的模型，不仅计算速度更快，在翻译任务上获得了更好的结果，也为后续的BERT模型做了铺垫
atte]]></description><category></category></item><item><title><![CDATA[DL_RNN_DAY7]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114897884</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114897884</guid><author>tjjyqing</author><pubDate>Wed, 17 Mar 2021 20:46:28 +0800</pubDate><description><![CDATA[RNN
1-of-N encoding
有些类似one-hot编码，将文本的类型转换为能够机器学习的向量类型进行输入。

除了上面的编码方式外，还有其余两种。左边是处理库存里面没有单词的方法，右边是根据单词类型进行编码。

store
循环神经网络RNN在DNN之外加入了可以存储之前输入信息的空间。这种情况下，神经网络会保持对之前数据的记忆。这样的话，输入的顺序会改变训练的效果。处理上下有关联的信息，比如文本信息是很有用的。

example

deep RNN
加深的话就直接加上几个隐含层了，每个神经元都]]></description><category></category></item><item><title><![CDATA[DL_CNN_Day6]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114458637</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114458637</guid><author>tjjyqing</author><pubDate>Sat, 06 Mar 2021 22:59:39 +0800</pubDate><description><![CDATA[CNN
cnn概况
cnn的步骤如下所示，比一般的神经网络多了卷积过程和池化过程。

convolution
卷积过程需要一个filter（过滤器）与原输入数据按部分依次相乘，组成新的矩阵。每次向右移动的格数称为步长。

输入数据可能有多层，如果是彩色照片，一般是三层。阿尔法狗的那个是用了48层来表示某一个落点的特征。

卷积操作和全连接层的关系：
卷积可以看成是神经网络一层的操作，如下所示。

max pooling
池化操作是把数据分割后按照一定原则处理数据，比如取这些组中最大的数据作为他们的代表，这样]]></description><category></category></item><item><title><![CDATA[DL_Backpropagation_tips for DNN_Day 5]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114365691</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114365691</guid><author>tjjyqing</author><pubDate>Thu, 04 Mar 2021 13:05:03 +0800</pubDate><description><![CDATA[目录homework for classificationBackpropagationpurposeprocessTips for training DNNReluMaxoutRegularizationDrop out
homework for classification
原课程是不让用sklearn写的.
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from s]]></description><category></category></item><item><title><![CDATA[DL_Classification、Logistic Regression、Deep Intro_Day4]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114293821</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114293821</guid><author>tjjyqing</author><pubDate>Wed, 03 Mar 2021 00:14:22 +0800</pubDate><description><![CDATA[目录Classificationtwo classesprobability from classGaussian DistributionMaximum LikelihoodModifying ModelLogistic Regressionfunction setevaluationfind the best functionmulti-class classificationlimitationBrief Introduction of DLneural networkgoodness of func]]></description><category></category></item><item><title><![CDATA[DL_Optimization_Day3]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114260239</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114260239</guid><author>tjjyqing</author><pubDate>Mon, 01 Mar 2021 13:17:54 +0800</pubDate><description><![CDATA[目录Optimizationwhat is optimizationcontentsome notationsSGDSGDMRMSPropAdamHow to improveAdamSGDMLookaheadAdamW SGDWMSomething helps optimization
Optimization
what is optimization
content
找到一个合适的参数的θ，使得L(θ)损失函数值最小
some notations
一些参数的含义如下所示

SGD
SGDM
SGD是梯度下]]></description><category></category></item><item><title><![CDATA[DL2020_Day2_Gradient Descent]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114237014</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114237014</guid><author>tjjyqing</author><pubDate>Sun, 28 Feb 2021 19:56:45 +0800</pubDate><description><![CDATA[Gradient Descent
homework 1
题目

代码
下面是利用梯度下降法做线性回归预测，用了三种方法。效果发现直接写梯度下降的代码效果比调用sklearn中线性回归模型以及岭回归模型要好。可能与直接编写的代码中针对学习速率做了调整有关。
import pandas as pd
import numpy as np
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import trai]]></description><category></category></item><item><title><![CDATA[DL2020_regression_Day1]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114193459</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114193459</guid><author>tjjyqing</author><pubDate>Sat, 27 Feb 2021 21:53:14 +0800</pubDate><description><![CDATA[regression
1 Model
Example:   f(input) = output

This is a linear function
2 Evaluation
loss function

Input: a function, output: how bad it is
2.1 High Bias
under fitting: the function is too simple

2.2 High Variance
overfitting: the function is too comp]]></description><category></category></item><item><title><![CDATA[机器学习_逻辑回归、聚类分析、模型调用与保存]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114087644</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114087644</guid><author>tjjyqing</author><pubDate>Thu, 25 Feb 2021 16:09:27 +0800</pubDate><description><![CDATA[一、模型的加载与保存
# 保存训练好的模型
joblib.dump(自己命名分类器名称, "保存地址")
# 比如 lr = LogisticRegression(),那里就写lr

# 调用保存好的模型(训练好的)去做预测
model = joblib.load("地址")
y_predict = model.predict(x_test)
# 当我们的x,y都被正则化的时候
#y_predict = std_y.inverse_transform(model.predict(x_test))
print]]></description><category></category></item><item><title><![CDATA[机器学习_线性回归问题]]></title><link>https://blog.youkuaiyun.com/tjjyqing/article/details/114060444</link><guid>https://blog.youkuaiyun.com/tjjyqing/article/details/114060444</guid><author>tjjyqing</author><pubDate>Thu, 25 Feb 2021 09:36:58 +0800</pubDate><description><![CDATA[线性回归
案例分析
正规方程、梯度下降、岭回归
from sklearn.datasets import load_boston
from sklearn.linear_model import LinearRegression, SGDRegressor, Ridge
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metri]]></description><category></category></item></channel></rss>