截图为copilot官网

要使用copilot，首先要安装vscode。下载copilot插件。下载完了以后，并不能马上使用。还需要申请内测账号，前几天，我终于收到内测账号通过的邮件。趁着今天在家，连忙打开vscode测试一下。

爬图片

我们的第一个实验是爬csdn的logo。

于是，我新建了一个文件，我开始写注释。当我写道retri的时候，copilot已经知道我要说什么了，直接把一句话补全了。牛！

我于是打了一下tab，这句话就补全了。

接着，我打了import，copilot已经开始建议我导入urllib了。牛！

打完def三个字，整个函数的建议都有了。牛！

接着，打if：

于是，满怀期待的，运行了程序，结果，悲剧了：

AttributeError: module ‘urllib’ has no attribute ‘urlopen’

我还是比较熟悉requests模块，于是，我把第一句代码改成了import requests，重新生成代码。如下：

# find logo in # retrieve the logo and save it in the same directoryimport requestsfrom bs4 import BeautifulSoupimport osdef get_logo(url):r = requests.get(url)soup = BeautifulSoup(r.text, 'html.parser')logo = soup.find('link', rel='shortcut icon')if logo:logo_url = logo['href']r = requests.get(logo_url)if r.status_code == 200:with open(os.path.basename(logo_url), 'wb') as f:f.write(r.content)print('logo saved')else:print('logo not found')else:print('logo not found')if __name__ == "__main__":get_logo('')

运行完了，发现不是logo，是favicon。😦

算了，算了，Copilot找不到logo，还是我自己找吧。

于是，我把注释改成了

# download and save https://img-/images/2024032511.png# to the same folder

生成的代码如下：

# download and save https://img-/images/2024032511.png# to the same folderimport requestsfrom bs4 import BeautifulSoupimport osdef download():url = 'https://img-/images/2024032511.png'root = './img'path = root + url.split('/')[-1]try:if not os.path.exists(root):os.mkdir(root)if not os.path.exists(path):r = requests.get(url)with open(path, 'wb') as f:f.write(r.content)f.close()print('文件保存成功')else:print('文件已存在')except:print('爬取失败')if __name__ == '__main__':download()

这次终于成功的下载了csdn的logo。

我想，这可能是因为csdn使用了反爬技术（logo图片名字不叫logo），所以找不到，于是，我想到了，可以去爬小米的200万logo。

给了注释以后，生成代码如下：

# find logo from /index.html# and download it to localimport osimport requestsfrom bs4 import BeautifulSoupdef download():url = '/index.html'r = requests.get(url)r.raise_for_status()soup = BeautifulSoup(r.text, 'html.parser')img = soup.find('img', class_='logo')print(img)img_url = img['src']img_name = img['alt']img_path = os.path.join('.', img_name)if not os.path.exists(img_path):with open(img_path, 'wb') as f:f.write(requests.get(img_url).content)print('Downloading %s' % img_name)else:print('%s already exists' % img_name)if __name__ == '__main__':download()

运行结果，找不到logo。我打开小米官网，看了半天，我都没看出来logo的地址在哪里。怪不得copilot也找不到。看来，200万的logo，不容屌丝随便盗取呀！

Pandas过滤

第二个例子，我打算写一个过滤程序，导入两个csv，用第一个减去第二个。得到第三个并保存。

生成代码如下

# read two csv files using pandas, # subtract the data in the second file from the first file,# and write the result to a new csv file.import pandas as pdimport csvimport sysdef subtract():# read the two csv filesdf1 = pd.read_csv(sys.argv[1])df2 = pd.read_csv(sys.argv[2])# subtract the second file from the first filedf3 = df1 - df2# write the result to a new csv filedf3.to_csv(sys.argv[3], index=False)if __name__ == '__main__':subtract()

结果，运行出错：

TypeError: unsupported operand type(s) for -: ‘str’ and ‘str’

于是，我修改了注释，再次生成。

# read two csv files, a.csv and b.csv# remove every row in a.csv that appears in b.csv# write the result to a new csv file.import osimport csvimport pandas as pddef main():# get the current working directorycwd = os.getcwd()# get the path to the filesa_path = os.path.join(cwd, "a.csv")b_path = os.path.join(cwd, "b.csv")# read the files into pandas dataframesa_df = pd.read_csv(a_path)b_df = pd.read_csv(b_path)# remove rows from a_df that appear in b_dfa_df = a_df[~a_df.index.isin(b_df.index)]# write the result to a new csv filea_df.to_csv("a_minus_b.csv", index=False)if __name__ == "__main__":main()

这次居然用了index过滤，但是并不是我想要的。算了，我还是去StackOverflow找吧。

Leetcode测试

我用leetcode第14题，测试，结果如下：

# 14. Longest Common Prefix# Easy# 4845# 2366# Add to List# Share# Write a function to find the longest common prefix string amongst an array of strings.# If there is no common prefix, return an empty string "".# Example 1:# Input: strs = ["flower","flow","flight"]# Output: "fl"# Example 2:# Input: strs = ["dog","racecar","car"]# Output: ""# Explanation: There is no common prefix among the input strings.# Constraints:# 1 <= strs.length <= 200# 0 <= strs[i].length <= 200# strs[i] consists of only lower-case English letters.class Solution(object):def longestCommonPrefix(self, strs):""":type strs: List[str]:rtype: str"""# 以上是我复制粘贴自# /problems/longest-common-prefix/# 以下为copilot生成if not strs:return ""if len(strs) == 1:return strs[0]min_len = min([len(s) for s in strs])for i in range(min_len):for j in range(len(strs)):if strs[j][i] != strs[0][i]:return strs[0][:i]return strs[0][:min_len]if __name__ == "__main__":a = Solution()print(a.longestCommonPrefix(["flower","flow","flight"]))print(a.longestCommonPrefix(["dog","racecar","car"]))print(a.longestCommonPrefix(["a","b","c"]))print(a.longestCommonPrefix(["a"]))print(a.longestCommonPrefix([""]))print(a.longestCommonPrefix([]))

打印结果

fl

a

看起来好像很厉害，都能做leetcode的题目了。但是，我大胆怀疑，copilot应该用了leetcode训练模型，所以能做出leetcode并不奇怪。

机器学习代码

同理，我给出注释，然后copilot给出代码，如下：

# read iris datasets from sklearn.datasets# predict the class of the samples in the test set# return the predicted classes# print accuracy, f1 score, recall, precisionfrom sklearn.datasets import load_irisfrom sklearn.model_selection import train_test_splitfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn.metrics import accuracy_score, f1_score, recall_score, precision_scoredef predict(X_train, y_train, X_test, y_test, k):knn = KNeighborsClassifier(n_neighbors=k)knn.fit(X_train, y_train)y_pred = knn.predict(X_test)print("Accuracy: {:.2f}".format(accuracy_score(y_test, y_pred)))print("F1 score: {:.2f}".format(f1_score(y_test, y_pred)))print("Recall: {:.2f}".format(recall_score(y_test, y_pred)))print("Precision: {:.2f}".format(precision_score(y_test, y_pred)))def main():iris = load_iris()X = iris.datay = iris.targetX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)predict(X_train, y_train, X_test, y_test, 5)if __name__ == "__main__":main()

又报错了：

ValueError: Target is multiclass but average=‘binary’. Please choose another average setting, one of [None, ‘micro’, ‘macro’, ‘weighted’].

修改predict函数如下

def predict(X_train, y_train, X_test, y_test, k):knn = KNeighborsClassifier(n_neighbors=k)knn.fit(X_train, y_train)y_pred = knn.predict(X_test)print("Accuracy: {:.2f}".format(accuracy_score(y_test, y_pred)))print("F1 score: {:.2f}".format(f1_score(y_test, y_pred, average='macro')))print("Recall: {:.2f}".format(recall_score(y_test, y_pred, average='macro')))print("Precision: {:.2f}".format(precision_score(y_test, y_pred, average='macro')))

再测试，通过了。

虽然有错误，但是我觉得也挺好，我可以少打不少字了。少数几个错误改一下即可。

总结

要用好copilot，还是需要给他非常明确的指示。比如下载图片，要告诉他图片的URL。很多时候，你忙了半天，或许还不如自己去网上找代码。常见问题能解决，非常见问题，就未必了。不过，这还仅仅是开始，相信未来copilot会越来越准。那时候，善于使用copitlot的程序员，就不用加班了。不用copilot，就要天天加班了。

即使有了copilot，你还是需要自己理解代码。比如上面的pandas过滤程序，代码其实错了，但是正好通过了测试。如果你不知起所以然，直接上生产，后果可想而知。

测试过程中发现，他的速度非常不稳定。经常是半天没反应，不知道是不是网络原因。希望未来copilot能在中国放服务器，这样速度就有保证了。