使用阿里雲ECS建立聊天機器人

初次釋出於2018年1月

前期考慮使用了圖靈機器人，有兩種實現方式繫結公眾號和加入群聊。圖靈機器人可以直接繫結公眾號，使用過程還是不錯的，不用明確@，可以直接進行對話，上下文邏輯比微軟小冰要好，有1000條/天的限制。第二種方案則是使用第三方外掛，繫結微訊號，用機器人執行微訊號，缺點就是需要一直執行微信不能退出，邏輯不明顯。好吧，邏輯都不是很明顯。

接下來考慮了Sam Gu呼叫itchat的方式[1]。Sam 使用谷歌雲平臺整合ItChat微信聊天機器人並整合深度人工智慧應用，首先安裝Jupyter Notebook。

登陸伺服器，可以通過控制檯，或者putty, mobaxterm等工具，前者簡潔後者提供檔案操作介面。

mkdir anaconda #建立安裝目錄

cd anaconda #將進入安裝目錄

wget https://repo.continuum.io/archive/Anaconda3-5.0.1-Linux-x86_64.sh #下載安裝包

bash Anaconda3-5.0.1-Linux-x86_64.sh #安裝另外解除安裝用rm -rf anaconda3

jupyter notebook --generate-config --allow-root 

ipython 

from notebook.auth import passwd

passwd()
#設定密碼會生成hashed password，複製

vi ~/.jupyter/jupyter_notebook_config.py

複製貼上以下幾行，然後執行。具體操作可以參翼起小飛在社群的博文，即[2]。

 

c.NotebookApp.ip=`*`

c.NotebookApp.password = u`把上面的文字貼上到這裡`

c.NotebookApp.open_browser = False

c.NotebookApp.port =8888

nohup jupyter notebook --allow-root

 

然後執行Jupyter Notebook，並在瀏覽器登入。Sam是在谷歌雲臺（第一年免費，贈送300美元，當然阿里雲也有學生優惠）上建立的，基本原理是一致的，他的視訊裡講的也是非常清楚，這裡把這些內容搬運過來。登陸Jupyter Notebook後，建立一個資料夾存放相關檔案，然後建立Python3檔案，在其命令列裡複製並執行以下命令，第二條執行兩次。

 

!git clone https://github.com/telescopeuser/workshop_blog.git

%load workshop_blog/setup_cloud.py

 

執行lesson 1就可以開始了。當然這些命令也可以在伺服器直接呼叫，Jupyter Notebook在這裡更多的是互動介面的作用。接下來的圖形識別要呼叫機器學習，Sam使用了Google Cloud Platform`s Machine Learning APIs，直接呼叫就可以，複製貼上API，然後再執行lesson 2。Lesson3 主要講了語音的識別，翻譯。Lesson4就高深一點，涉及到情感語義分析，接下來的lesson5涉及到從視訊中提取資訊。按照參考文獻[1]的知識進行操作即可。但是，這樣呼叫谷歌API，和之前掛用圖靈機器人，最大的區別就是多了幾行程式碼，Well，很多行程式碼，和我預想的還是有差距的。

然後在warmheartli的專案裡面發現了些有趣的內容，不過我要準備去擠火車了……

 

更新於2018年9月17日星期一：收集聊天語料

莫名其妙的更新被刪除掉，好在以前的都還留著，從新再發語料的收集不再考慮使用字幕，從微博爬取資料，作為日後的語料和機器學習用，爬蟲的方法參考nghuyong，nghuyong的方法在驗證登入的時候存在無法識別。因此滑塊驗證的方法參考LiuXingMing/WeiboSliderCode；Python3WebSpider/CrackWeiboSlide；bone_ace/article/details/71056741，當然他們的方法也並不是有效，嘗試了很多方法後，最簡單粗暴地是把驗證圖形的矩陣直接放在cookies.py裡面。

 

#!/usr/bin/env python

# encoding: utf-8

import datetime

import json

import base64

from time import sleep

import os

import time

import random

import io

from PIL import Image

from math import sqrt

 

import pymongo

from selenium import webdriver

from selenium.webdriver import ActionChains

from selenium.common.exceptions import TimeoutException

from selenium.webdriver import ActionChains

from selenium.webdriver.common.by import By

from selenium.webdriver.support.ui import WebDriverWait

from selenium.webdriver.support import expected_conditions as EC

from selenium.webdriver.remote.command import Command

 

WeiBoAccounts = [

{`username`: `登陸賬號`, `password`: ‘密碼’},

]

 

cookies = []

client = pymongo.MongoClient(“localhost”, 27017)

db = client[“Sina”]

userAccount = db[“userAccount”]

PIXELS = []

 

 

def getExactly(im):

    “”” Precise cutting “””

    imin = -1

    imax = -1

    jmin = -1

    jmax = -1

    row = im.size[0]

    col = im.size[1]

    for i in range(row):

        for j in range(col):

            if im.load()[i, j] != 255:

                imax = i

                break

        if imax == -1:

            imin = i

 

    for j in range(col):

        for i in range(row):

            if im.load()[i, j] != 255:

                jmax = j

                break

        if jmax == -1:

            jmin = j

    return (imin + 1, jmin + 1, imax + 1, jmax + 1)

 

 

def getType(browser):

    “”” Identifying the graphic path “””

    ttype = “

    time.sleep(3.5)

    im0 = Image.open(io.BytesIO(browser.get_screenshot_as_png()))

    box = browser.find_element_by_id(`patternCaptchaHolder`)

    im = im0.crop((int(box.location[`x`]) + 10, int(box.location[`y`]) + 100, int(box.location[`x`]) + box.size[`width`] – 10, int(box.location[`y`]) + box.size[`height`] – 10)).convert(`L`)

    newBox = getExactly(im)

    im = im.crop(newBox)

    width = im.size[0]

    height = im.size[1]

    for png in ims.keys():

        isGoingOn = True

        for i in range(width):

            for j in range(height):

                if ((im.load()[i, j] >= 245 and ims[png][i][j] < 245) or (im.load()[i, j] < 245 and ims[png][i][j] >= 245)) and abs(ims[png][i][j] – im.load()[i, j]) > 10: # With 245 as the threshold, about 245 as the blank, less than 245 as the line; the difference between two pixels is about 10 to remove the error on the 245 boundary

                    isGoingOn = False

                    break

            if isGoingOn is False:

                ttype = “

                break

            else:

                ttype = png

        else:

            break

    px0_x = box.location[`x`] + 40 + newBox[0]

    px1_y = box.location[`y`] + 130 + newBox[1]

    PIXELS.append((px0_x, px1_y))

    PIXELS.append((px0_x + 100, px1_y))

    PIXELS.append((px0_x, px1_y + 100))

    PIXELS.append((px0_x + 100, px1_y + 100))

    return ttype

 

 

def move(browser, coordinate, coordinate0):

    “”” Move from coordinate0 to coordinate “””

    time.sleep(0.05)

    length = sqrt((coordinate[0] – coordinate0[0]) ** 2 + (coordinate[1] – coordinate0[1]) ** 2)  # Two point line distance

    if length < 4:  # If the distance between two points is less than 4PX, go straight ahead.

        ActionChains(browser).move_by_offset(coordinate[0] – coordinate0[0], coordinate[1] – coordinate0[1]).perform()

        return

    else:  # Recursion, sliding to the end

        step = random.randint(3, 5)

        x = int(step * (coordinate[0] – coordinate0[0]) / length)  # In proportion

        y = int(step * (coordinate[1] – coordinate0[1]) / length)

        ActionChains(browser).move_by_offset(x, y).perform()

        move(browser, coordinate, (coordinate0[0] + x, coordinate0[1] + y))

 

 

def draw(browser, ttype):

    “”” Slide “””

    if len(ttype) == 4:

        px0 = PIXELS[int(ttype[0]) – 1]

        login = browser.find_element_by_id(`loginAction`)

        ActionChains(browser).move_to_element(login).move_by_offset(px0[0] – login.location[`x`] – int(login.size[`width`] / 2), px0[1] – login.location[`y`] – int(login.size[`height`] / 2)).perform()

        browser.execute(Command.MOUSE_DOWN, {})

 

        px1 = PIXELS[int(ttype[1]) – 1]

        move(browser, (px1[0], px1[1]), px0)

 

        px2 = PIXELS[int(ttype[2]) – 1]

        move(browser, (px2[0], px2[1]), px1)

 

        px3 = PIXELS[int(ttype[3]) – 1]

        move(browser, (px3[0], px3[1]), px2)

        browser.execute(Command.MOUSE_UP, {})

    else:

        print(`Sorry! Failed! Maybe you need to update the code.`)

                                    

def get_cookie_from_weibo(username, password):

    browser = webdriver.Chrome()

    browser.set_window_size(1050, 840)

    browser.get(`https://weibo.cn/`)

    time.sleep(1)

    assert “微博” in browser.title

    login_link = browser.find_element_by_link_text(`登入`)

    ActionChains(browser).move_to_element(login_link).click().perform()

    login_name = WebDriverWait(browser, 10).until(EC.visibility_of_element_located((By.ID, “loginName”)))

    login_password = browser.find_element_by_id(“loginPassword”)

    login_name.send_keys(username)

    login_password.send_keys(password)

    login_button = browser.find_element_by_id(“loginAction”)

    login_button.click()

    try:

        img = WebDriverWait(browser, 20).until(EC.presence_of_element_located((By.CLASS_NAME, `patt-shadow`)))

    except TimeoutException:

        print(`No verification codes`)

        self.open()

    # Here you stay for 10 seconds to see if the Chrome is successfully logged in.

    ttype = getType(browser)  # Identifying the graphic path

    print(`Result: %s!` % ttype)

    draw(browser, ttype)  # Slide

    time.sleep(10)

    cookie = browser.get_cookies()

    browser.close()

    return cookie

 

 

def init_cookies():

    for cookie in userAccount.find():

        cookies.append(cookie[`cookie`])

 

 

if __name__ == “__main__”:

    try:

        userAccount.drop()

    except Exception as e:

        pass

    for account in WeiBoAccounts:

        cookie = get_cookie_from_weibo(account[“username”], account[“password”])

        userAccount.insert_one({“_id”: account[“username”], “cookie”: cookie})

 

TOBECONTINUED

 

 

 

參考文獻：

[1]         Telescopeuser, workshop_blog, (n.d.). https://github.com/telescopeuser/workshop_blog.

[2]         翼起小飛, 如何在阿里ECS雲端執行Jupyter Notebook進行機器/深度學習？-部落格-雲棲社群-阿里雲, (n.d.). https://yq.aliyun.com/articles/98527 (accessed February 5, 2018).