GPT2 Telegram Chatbot
1.0.0
Чат-бот Telegram GPT-2, относительно настроенный для общения. Не стесняйтесь делать мне пиар, и я проверю ваш код! Бот не всегда точен на 100 % (почему я написал функцию /retry).
Поскольку бот потребляет очень много памяти, я запрограммировал его в циклическом режиме. При каждом вводе таймер на идентификаторе вашей учетной записи сбрасывается. Как только таймер истечет, бот станет бесплатным для использования другими пользователями. Вы будете уведомлены, когда таймер истечет, и другие пользователи смогут увидеть, сколько времени осталось и используется ли бот.
Краткая инструкция по установке на Ubuntu 20/WSL.
Я настоятельно рекомендую вместо этого взглянуть на блокнот Jupyter/ipynb в Google Collab.
Установите python3.7 (думаю, 3.6 тоже подойдет, но не 3.8):
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.7
Установите pip на Python 3.7:
wget https://bootstrap.pypa.io/get-pip.py
python3.7 get-pip.py
Установите требования внутри папки бота после клонирования репозитория:
python3.7 -m pip install -r requirements.txt
Примечание. На самом деле вам понадобится 16 ГБ оперативной памяти или видеокарта на 8 ГБ. В противном случае вы будете ждать вечно. Вы можете использовать функции графического процессора как минимум с видеокартой объемом 8 ГБ, которая поддерживает cuda takelit 10.0 и cudnn для cuda Toolkit 10. Эта установка также работает в Windows с Python 3.7 и nvidia, вы должны запустить командную строку от имени администратора, запустив Python 3.7 в Windows.
Загрузите модель:
python3.7 download_model.py 1558M
Установите ключ API своего бота Telegram в src/GPT2-Learning.py.
Replace "BOTKEY" with telegram bot token i.e. "1827396499:AAHifc06oS31oQ9L3TuCiZxD9EIfKPi0oWQ"
Запустите бота:
Если вы используете команду python3:
./start.sh
Если вы используете команду python3.7:
python3.7 src/GPT2-Learning.py
Вы можете отредактировать start.sh, чтобы он соответствовал вашей команде Python, по вашему желанию.
start - Basic start command.
help - Show quick help command.
chatbot - Chatbot mode, me+you string edit and open-ended questions.
finish - Default finish sentence command counting words as output length.
learnon - Turn on sentence compilation with Me+you string for conversation.
learnoff - Turn off the sentence compilation.
learnreset - Reset the learned conversation.
retry - Retry the current computation.
Вот пример режима обучения чат-бота с последней базой данных gpt-2 и значением top_p, равным 0,83:
Me: Hello! How are you today? Are you feeling alright? I hope you are doing well and not getting into any trouble.
Bot: Yeah, I'm fine. Do you want to make something to eat?
Me: I already made some food, but thanks! I had some jalapeno poppers. I made a ton of them last night after I sold a lawnmower to some folks I got for free earlier in the day. So what are your plans for the day?
Bot: Nothing special. Nothing exciting in particular. I just wanna chill out for a bit.
Me: Just a typical day of kicking back and relaxing huh? Are you gonna watch cartoons or any movies in particular?
Bot: If it's not sports I'll probably just watch something on the TV.
Me: Sounds alright. I coded on my chatbot project most of the morning. Was waiting on someone to come buy a projector but they haven't shown up yet unfortunately. What do you do for work if you don't mind me asking?
Bot: I'm a web developer and engineer. It's fun.
Me: Cool! I'm also a coder and an ethical hacker. I mostly know how to code in Python but I've been unemployed for some time now after my spine injury. I wish I could have joined the military when I was younger. Were you ever in the military?
Bot: No. I'm not really too proud of my military career. It ended up taking a huge toll on my back.
Me: What happened to your back if you don't mind me asking?
Bot: First of all, it's better than before.
Me: Well that's good to hear. Did you injure your back in the military or doing something else?
Bot: Nope. It's an old injury.
Me: Ah I see- Well I'll talk to you later. Sorry about your back! I hope we can catch up tomorrow!
Bot: Yes, of course. Thanks again for chatting!
INFO:tensorflow:Restoring parameters from models/1558M/model.ckpt
2020-05-10 14:19:43,794 - tensorflow - INFO - Restoring parameters from models/1558M/model.ckpt
==========
Before splitlines: Nothing special. Nothing exciting in particular. I just wanna chill out for a bit. You: I'll try to keep my spirits up. Just keep in mind that today's the 10th anniversary of the citywide strike
==========
==========
Mode: True
Learn: True
Length: 44
==========
Before regex: Nothing special. Nothing exciting in particular. I just wanna chill out for a bit. You: I'll try to keep my spirits up. Just keep in mind that today's the 10th anniversary of the citywide strike
==========
Output: Nothing special. Nothing exciting in particular. I just wanna chill out for a bit.
==========
Raw_text or Original: You: Hello! How are you today? Are you feeling alright? I hope you are doing well and not getting into any trouble. Me: Yeah, I'm fine. Do you want to make something to eat? You: I already made some food, but thanks! I had some jalapeno poppers. I made a ton of them last night after I sold a lawnmower to some folks I got for free earlier in the day. So what are your plans for the day? Me:
==========
Learning text or Next: You: Hello! How are you today? Are you feeling alright? I hope you are doing well and not getting into any trouble. Me: Yeah, I'm fine. Do you want to make something to eat? You: I already made some food, but thanks! I had some jalapeno poppers. I made a ton of them last night after I sold a lawnmower to some folks I got for free earlier in the day. So what are your plans for the day? Me: Nothing special. Nothing exciting in particular. I just wanna chill out for a bit.
==========
top_p out: 0.8338636363636364
==========
top_p in: 0.83
==========
Список оценок по грамматике можно найти на сайте /SCORES.MD.
Кончик:
0.77 top_p может звучать эмоционально, сбивчиво и подражательно.
0.66 top_p может звучать продуманно и буквально, но может содержать ошибки ascii и обрезки.
