Natacha Bot

The natacha bot generates posts in the style of an individual, Natacha. It is derived from the Trump bot using Facebook posts and comments using a word-level RNN and pre-trained GloVe word vectors. This repo corresponds to this Eight Portions blog post.

Development

The following steps are what I did to generate the container and model. I don't typically do this interactively, but since the training would take days I chose this approach. First, I parsed raw data using data/natacha/generate_input.py. I went through this after generating the file to still remove people's full names and any address information.

Then I pull the base image that has python and dependencies installed.

$ docker pull rtlee/t-bot:sample

Run the container and bind the new natacha data directory.

docker run -it -v $PWD/data/natacha:/root/Trump-bot/data/natacha rtlee/t-bot:sample

Copy the script to train the natacha data:

cp data/natacha/train_*.sh .

And then run the training after updating the repository:

git pull origin master
./train_script.sh # I'm not sure if we need to run this
./train_words.sh
./train_char.sh

Identify the best word and character level models and move to /cv/, replacing the existing files corresponding to the appropriate model type.

cp cv_char/lm_lstm_epoch30.00_1.3580.t7 cv/char-rnn-trained.t7
cp cv_word_caps_256_2/lm_lstm_epoch40.00_4.4779.t7 cv/word-rnn-trained.t7

Then you can run the script to generate output!

python sample.py "I will build a"

When it was working to my liking, I could commit the current container to my local machine to be able to push it to Docker Hub.

$ docker commit 49c6efad77dd vanessa/natacha-bot:latest

And then test running the container to generate new text.

$ docker run --rm -it vanessa/natacha-bot:latest python sample.py "I want to"

I want to be the power syndicate that they are the many with the profit present that she is a significant of the new the controlled profit since they are state the corruption of the state of the artist sender that is the organization and a corruption when it is a thing at the media many are they have the sen. Something can't go from your mind. I am not going to find the more again and not not sure that now isn't a but like a form of a break. If you will know the facts any of your evidence ? You can

Note that this shows manual generation of the model using the training container, which is very large (~15GB). To distribute you likely want to use the multistage build, discussed next.

Build

The included Dockerfile can be used as a start for generating the same image.

docker build -t vanessa/natacha-bot .

But then you likely want to build a smaller container using a multistage build to just copy the finished models from cv, and then be done!

docker build -f Dockerfile.bot -t natacha-bot .

Note that I name it something different, even though ultimately I want to tag it as vanessa/natacha-bot. I do this in case there is a mistake - I don't want to lose the original container. You can then test:

docker run -it natacha-bot

And then tag the final container to save it!

docker tag natacha-bot vanessa/natacha-bot

Credits, inspiration and similar projects

• The glorious, original Trump-bot.
• Lars Hiller Eidnes word-rnn based on Andrej Karpathy's char-rnn.
• Auto-Generating Clickbait With Recurrent Neural Networks
• DeepDrumpf: Twitterbot
• RoboTrumpDNN: Generating Donald Trump Speeches with Word2Vec and LSTM