documentation and tutorial on adding three capabilities to the neural chatbot agent (NCA) previously described here chat-transformer
There are 3 major improvements described here:
-
Methods to keep an elastic vocabulary embedding, allowing the number of tokens understood by the model as input or output to expand and contract throughout the lifetime of the NCA
-
Methods to quickly encode a memory at test time using few-shot meta-learning, allowing the model to remember concepts that it can use immediately without time-consuming batch gradient descent
-
Methods to train the model in parallel on multiple GPUs to speed up batch gradient descent, which is still required to learn how to quickly remember concepts
Typical seq2seq models have a fixed vocabulary, ie the number of words or tokens that can be used in the input and output are fixed. We have developed a method here for keeping a vocabulary is constantly growing and occasionally pruned.
This repository builds on the chat-transformer chatbot by adding a memory mechinism to the transformer sequence2sequence chatbot based on this paper by Tsendsuren Munkhdalai et al called Metalearned Neural Memory: Teaching neural networks how to remember,here are some Google Slides from an MLCOllective Research Jam
the high level gist of it is this: the controller in the diagram above is the sequence2sequence chatbot. In addition to taking a sequence as input and outputting a sequence as output, it also sends a signal to the memory network, a key vector, this is the input to a neural network telling it to retrieve a memory. That memory recall is returned to the chatbot in the form of the value vector which the chatbot can use to inform its next response.
We explain the lines of code that recruit and parallelize training across multiple GPUs
if you already have python 3.6 and virtual environments, create a python 3.6 virtual environment, here i used env36 for python3.6 but you can use anything
python3 -m venv env
or
$ python3.6 -m venv env
or
virtualenv --python=/usr/bin/python3.6 env
if python3.6 is your default version, then when you type python into your terminal then it should say python version 3.6.x, and for you making the virtual environment is as simple as
$ python -m venv env
otherwise it is simple to get python3.6 and virtual environments
how to install Python 3.6 on ubuntu
install virtual environment then
how to specify the Python executable you want to use
sudo add-apt-repository ppa:jonathonf/python-3.6
sudo apt-get update
sudo apt-get install python3.6
virtualenv --python=/usr/bin/python3.6 env36
When you want to run the code activate the virtual environment inside the same folder as your environment env using
$ source env/bin/activate
install dependencies
$ pip3 install -r requirements.txt
even with virtual environments, some troubleshoot might be needed
with enough google searches you can find an answer for almost any problem
save new dependences to requirements
$ pip3 freeze > requirements.txt
You can deactivate the virtual environment using the following command in your terminal:
$ deactivate
if you get a ImportError: No module named while at the same time in your Terminal you get
pip3 install import-ipynb
Requirement already satisfied: import-ipynb in /path/to/env/lib/python3.6/site-packages (0.1.3)
This can be fixed by providing your python interpreter with the path-to-your-module,the path
import sys
sys.path.append('/path/to/env/lib/python3.6/site-packages')
you can find this path listed here
$ python3 -m site
Start with the Minimal Working Example to get the gist. Then to learn the details, start with notebook who's name starts with part_1 and follow the directions inside, moving to part_2 etc.