建立没有数据集96准确性的辣胡椒分类器

数据科学 ， 机器学习 (Data Science, Machine Learning)

In this article, I will create an AI capable of recognizing a spicy pepper from measurements and color. Because you won’t be able to find any dataset on the measurements of spicy peppers online, I will generate it myself using statistics methodologies. In a second article, I may try to apply regression algorithms to estimate the spiciness of your pepper on the Scoville Scale.

在本文中，我将创建一个能够通过测量和颜色识别辣辣椒的AI。 因为您将无法在线上找到任何关于辣辣椒测量的数据集，所以我将使用统计方法自行生成该数据集。 在第二篇文章中，我可能会尝试使用回归算法以Scoville量表估算您的胡椒的辛辣度。

处理： (Process:)

1. Finding Available Data
   查找可用数据
2. Making Measurements
   进行测量
3. Creating the dataset from distributions
   从分布创建数据集
4. Creating the Model
   创建模型
5. Performance Evaluation
   绩效评估

1.查找可用数据 (1. Finding available data)

As mentioned before, you will unlikely find a dataset for everything you wish to build. In my case, I wanted to build a spicy pepper classifier, which is a difficult task if you have no data to start with. The only thing I could find on the internet was a comparison table of different spicy peppers (hopefully on the same scale).

如前所述，您不太可能找到要构建的所有内容的数据集。 就我而言，我想构建一个辣味分类器，如果没有任何数据开始，这将是一项艰巨的任务。 我在互联网上唯一能找到的是一张不同麻辣胡椒的比较表(希望是相同的比例)。

I will need to transform this data into a digital one. What I can do is take measurements of these images and place them as features in a dataset.

我将需要将此数据转换为数字数据。 我所能做的就是对这些图像进行测量，并将它们作为特征放置在数据集中。

2.进行测量 (2. Making Measurements)

To make measurements I can use pixels. After knowing the rate of conversion of pixels to centimeters I can measure the size of each spicy pepper in pixels and convert it to its real-world scale.

为了进行测量，我可以使用像素。 在了解了像素到厘米的转换率之后，我可以测量每个香辛椒的大小(以像素为单位)，并将其转换为现实世界的比例。

This is the final table with all measurements (name, height, width, and color) converted into features.

这是最终表，其中所有度量(名称，高度，宽度和颜色)均已转换为特征。

#   measurements
pepper_measurements_px = [
                          ['Anaheim', 262, 63, 'Green'],
                          ['Cubanelle', 222, 70, 'Green'],
                          ['Cayenne', 249, 22, 'Red'],
                          ['Shishito', 140, 21, 'Green'],
                          ['Hungarian Wax', 148, 63, 'Orange'],
['Jimmy Nardello', 190, 23, 'Red'],
                          ['Fresno', 120, 43, 'Red'],
                          ['Jalapeno', 106, 40, 'Dark Green'],
                          ['Aji Amarillo', 92, 13, 'Yellow'],
                          ['Aji Dulce', 81, 30, 'Red'],
['Serrano', 74, 14, 'Dark Green'],
                          ['Padron', 62, 38, 'Dark Green'],
                          ['Scotch Bonnet', 37, 42, 'Yellow'],
                          ['Habanero', 67, 21, 'Orange'],
                          ['Cumari', 18, 11, 'Yellow'],
]

I will now generate a dataset of 100.000 samples for spicy peppers.

现在，我将生成一个100.000个辛辣辣椒样本的数据集。

3.从分布创建数据集 (3. Creating datasets from distributions)

Before starting to create distributions, I will first need to convert the pixels into centimeters. Then for both length and width, I will need two separate normal distributions using this data as mean. For a standard deviation, I will use 10% of the mean (in this way I won’t have to Google the details of every spicy pepper).

在开始创建分布之前，我首先需要将像素转换为厘米。 然后，对于长度和宽度，我将需要使用此数据作为均值的两个单独的正态分布。 对于标准差，我将使用平均值的10％(这样，我就不必在Google上搜索每个辛辣胡椒的详细信息)。

创建功能 (Creating Functions)

I am creating a set of functions that will allow the creation of n datasets, inputting the size. I will use 100,000 samples for spicy pepper.

我正在创建一组函数，将允许创建n个数据集，并输入大小。 我将用100,000个样本制作辣胡椒。

#simulated probability distribution of one stock
from scipy.stats import skewnorm
import matplotlib.pyplot as plt
import pandas as pd
import numpy as npdef create_peppers(sd, mean, alfa, size):
  #invertire il segno di alfa
  x = skewnorm.rvs(-alfa, size=size) 
  def calc(k, sd, mean):
    return (k*sd)+mean
  x = calc(x, sd, mean) #standard distribution#graph the distribution
  #pd.DataFrame(x).hist(bins=100)#pick one random number from the distribution
  #formally I would use cdf, but I just have to pick randomly from the 1000000 samples
  df = [np.random.choice(x) for k in range(size)]
  #return the DataFrame
  return pd.DataFrame(df)def cm_converter(px_measurements):
  pc_cm = 0.05725
  for _ in range(len(px_measurements)):
    px_measurements[_][1] *= pc_cm
    px_measurements[_][2] *= pc_cm
  return px_measurements

创建数据集 (Creating the Dataset)

I am now ready to create the datasets. I can specify the use of the 10% of the mean as a standard deviation (I can easily change it from height_sd and widht_sd):

我现在准备创建数据集。 我可以指定使用平均值的10％作为标准偏差(我可以很容易地从height_sd和widht_sd进行更改)：

#   create converted list
pepper_measurements_cm = cm_converter(pepper_measurements_px)#   create final datasets
heigh_sd = 0.1
width_sd = 0.1df = pd.DataFrame()
for _ in pepper_measurements_cm:
  #   create height
  #SD is 10% of the height
  df_height = create_peppers(_[1]*heigh_sd, _[1], 0, 100000)
  #   create width
  #SD is 10% of the width
  df_width = create_peppers(_[2]*width_sd, _[2], 0, 100000)
  #create DataFrame
  df_single = pd.concat([df_height, df_width], axis=1)
  df_single.columns = ['height', 'width']
  #create name
  df_single['name'] = str(_[0])
  df_single['color'] = str(_[3])df = pd.concat([df, df_single], axis=0)
df

Normal distribution of a single generated feature

单个生成特征的正态分布

This is the final result: combined, the dataset counts 1.5 Million samples:

这是最终结果：合并后，数据集计数了150万个样本：

Final Dataset

最终数据集

If we plot height and width in different histograms:

如果我们在不同的直方图中绘制高度和宽度：

height and width in separated histograms

分开的直方图中的高度和宽度

4.创建模型 (4. Creating the Model)

The model I will be using is a Naive Bayes Classifier. Rather than many other models, this one is specialized with data that:

我将使用的模型是朴素贝叶斯分类器。 而不是许多其他模型，该模型专用于以下数据：

• Is independent
  是独立的
• Follows a normal distribution
  服从正态分布

Because I built my dataset following these presuppositions, this classifier is perfect for what I wish to build.

因为我是按照这些前提建立数据集的，所以该分类器非常适合我要构建的内容。

前处理 (Preprocessing)

The only preprocessing step I will have to make is encoding the color with a one_hot encoding algorithm:

我唯一要做的预处理步骤是使用one_hot编码算法对颜色进行编码：

#backup
X = df.copy()def one_hot(df, partitions):
  #togliamo le colonne da X
  for col in partitions:
    k = df.pop(col)
    k = pd.get_dummies(k, prefix=col)
    df = pd.concat([df, k] , axis=1)
  return dfX = one_hot(X, ['color'])
X

选择功能和标签 (Selecting Features and Labels)

y = X.pop('name')
y

Labels

标签

X

Features after one_hot encoding

one_hot编码后的功能

分裂 (Splitting)

from sklearn.model_selection import train_test_splitX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

I will now split Features and Labels randomly, a ratio of 80:20 will suffice.

现在，我将随机分割特征和标签，比率为80:20就足够了。

训练模型 (Training the Model)

from sklearn.naive_bayes import GaussianNB
clf = GaussianNB()
clf.fit(X_train, y_train)

The model has been trained:

该模型已经过训练：

GaussianNB(priors=None, var_smoothing=1e-09)

5.绩效评估 (5. Performance Evaluation)

After training the model, I will test it on the part of the dataset which the AI has never seen during training:

训练完模型后，我将在AI在训练过程中从未见过的数据集部分进行测试：

clf.score(X_test, y_test, sample_weight=None)
0.9659133333333333

The model has reached an outstanding 96% accuracy!

该模型达到了出色的96％精度！

翻译自: https://medium.com/towards-artificial-intelligence/building-a-spicy-pepper-classifier-with-no-datasets-96-accuracy-8262d54a8117