This project aims to implement a simple interface to use Facebook ParlAI's chatbot based on the paper Beyond Goldfish Memory: Long-Term Open-Domain Conversation.
It provides multiple scripts to discuss with Summary-Memory model and BlenderBot2 or eventually train them.
The ParlAI repository recquires a lot of requirements. 😨
It is recommended to use a Conda environment this time because there can be some bugs on Windows with pip or on Google Colab. It is also highly recommended to be on a Linux setup (WSL or whatsoever) for some installations like fairseq or FAISS. Just execute these commands
git clone https://github.com/Rubiksman78/Memory_chatbot.git
If you want to use virtualenv:
pip install virtualenv
virtualenv venv
source venv/bin/activate #or source venv/Scripts/activate on Windows
If you want to use conda venvs:
conda create -n venv python=3.9
conda activate venv
And now install all the requirements for ParlAI (the setup can be quite long).
bash setup.sh
If you want to chat with the model on the browser_chat API provided by FB you can execute
bash scripts/chat.sh
You can now connect from your browser on "IP:8080". You can get the IP of your remote machine with:
hostname -I
There might be some issues if the download of the model is interrupted during the process (be sure to have a stable internet connection or you would have to download the model several times). Just delete the folder ParlAI/data/models/problematic_model and relaunch the command.
You can specify which way of chatting you want to use in chat_config.py. You can also take a look at the opt files in the configs folders which can allow you to custom the models (for inference or to fine tune them).
You can store history conversations in the text_files folder and change the related paths in the configs. You can also create a database of personas in a .csv file to load them at the end of the context.
There is a script to interact directly with the model from a simple python script
python summsc_chat.py
Or if you want to use BB2 model directly without search engine
python blender_chat.py
You may have to delete models in ParlAI/data if there are some downloading issues.
To use the gradio chat interface launch the server on the remote machine with
python summsc_gradio.py
Or for the BB2 model
python blender_gradio.py
You can now access to the public server just by clicking the link.
You can also use localhost: On your local machine create a tunnel to this server with
ssh -L local_port:remote_server_ip:remote_port remote_name@remote_host
ssh -L 8080:192.168.10.40:7860 remote_name@remote_host
The local terminal will ask you your ID for your remote account and you can now access to the chatbot API on localhost:8080.