当前位置:网站首页>【AI助力科研】loss曲线傻瓜式绘制
【AI助力科研】loss曲线傻瓜式绘制
2022-06-25 12:10:00 【燕策西】
【AI助力科研】loss曲线傻瓜式绘制
前言
俗话说,一图胜千言,在一篇文章里,图是最吸引人的,画得好便会如虎添翼,让读者对文章主要方法的理解更加容易。今日分享和记录一下绘制损失(loss)函数图时发现的一个事后补救、锦上添花的小小小小~小技巧。看一下成品:
一、已具备的loss曲线图
之前通过python的matplotlib或者plotly绘制好部分的loss曲线图如下(肯定是部分图啊,这么单调怎可单独成图?),且不说作为成果展示让人家品评,你就自己看了都想揉一下扔进垃圾桶,shit,啥玩意?真就工科男,直接放原图?不修一修的嘛?
咳咳,当然,这是真实的结果图,反映了network在向好发展,训练过程自然是不稳定,那么真的就这样直接放图?当然不,读者要的肯定有两个信息——趋势和性能,趋势可以用局部均值来描述,性能可以用过程中的局部方差来代替。所以,我们可以采用python重新画图,计算t时刻前后N个点的均值、方差,然后绘图。写到这里,其实还是要提醒下自己:数据保存,过程保存。因为我没有保存训练时的loss下降的数据…再去跑一次真的很煞风景,你说实验都做完了,现在又要跑一次训练过程???人懒,不想做。
写一句营销味儿的广告词:用这个办法,1 min 让图妙起来!
二、AI助力
使用的AI工具就俩:矩形工具和曲率工具
1.为图像建立矩形遮板
首先将原图复制一份作为副本,粘贴到一个新建文件或者空白的地方,防止修改原图。在副本图像上方画一个矩形,如果矩形图层不在最上方,一定要调整。将矩形填充为白色,调整透明度,只要下方曲线可见就行,如图,现在无法选中下方曲线了,因为有矩形遮住了。
2.绘制趋势曲线
使用曲率工具,绘制趋势图,隔几个点过一下中点,画完后退出或按下英文v键即可。
单独调整趋势曲线的颜色、线宽等。
最后将上述趋势曲线移动到原图上方,然后将原图的透明度调一下就可以得到最终效果了。
AI很妙。OK,收工。
边栏推荐
- Disconnected: No supported authentication methods available (server sent: )
- Renrenyue -- renrenyue system development source code sharing
- JS uses the for loop in the function to insert and delete the array at the specified position
- Zhangxiaobai's way of penetration (VIII) - detailed operation steps of SQL injection - Boolean blind injection of blind injection
- Parse JSON format data and save it to entity class
- ThinkPHP upload image compression size
- Zhangxiaobai's road of penetration (IV) -- detailed explanation of XSS cross site script vulnerabilities
- Zhangxiaobai's way of penetration (V) -- detailed explanation of upload vulnerability and parsing vulnerability
- Full nanny tutorial of Market Research Competition (experience sharing)
- Penetration tool environment -- use of cknife Chinese kitchen knife
猜你喜欢

Why do we do autocorrelation analysis? Explain application scenarios and specific operations

Execution order of MySQL query statements join, on and where

(4) Pyqt5 tutorial -- > Custom signal and slot (super winding...)
![按权重随机选择[前缀和+二分+随机target]](/img/84/7f930f55f8006a4bf6e23ef05676ac.png)
按权重随机选择[前缀和+二分+随机target]

Today, I will explain to you what is DFI and its development prospects

Optimal solution for cold start

2021-09-02

Zhangxiaobai's way of penetration (V) -- detailed explanation of upload vulnerability and parsing vulnerability

Guess Tongyuan B

C program linking SQLSERVER database: instance failed
随机推荐
Matlab simulation of m-sequence
2022 Baidu collection batch automatic push assistant
Wechat forbids sharing
GNSS receiver technology and application review
ECSHOP commodity wholesale multi attribute multi specification multi inventory batch purchase ECSHOP wholesale plug-in ECSHOP multi attribute order
Common software numerical filtering methods (I) have been applied
Shell learning notes (latest update: 2022-02-18)
地理空间搜索:kd树的实现原理
Today, I will explain to you what is DFI and its development prospects
Zhangxiaobai's road of penetration (IV) -- detailed explanation of XSS cross site script vulnerabilities
ECSHOP upload video_ ECSHOP video list, video classification, video related product guide
Parse JSON format data and save it to entity class
(4) Pyqt5 tutorial -- > Custom signal and slot (super winding...)
High performance + million level Excel data import and export
High imitation blue playing network disk file sharing to make money network disk PHP system source code
Negative sample image used in yolov5 training
Upgrade opsenssh to 8.8p1
Figure explanation of fiborache sequence
2021-09-30
el-select clear 清空内容时触发事件