RNNs with Darknet

Installing Darknet

• We need to install our modified version of Darknet:
• Darknet Installation

Producing text with a pre-trained model

• Download some pre-trained models or calculated weights from several writers:

wget https://pjreddie.com/media/files/shakespeare.weights
wget https://pjreddie.com/media/files/kant.weights
wget https://pjreddie.com/media/files/tolstoy.weights

• Example: Generating a text based on the pre-trained model with books written by Shakespeare with Darknet:
• The machine learning framework Darknet, is installed in the black PCs of the Interaction Station. We will use the operating system Ubuntu (Linux):
• Login & password: interactionstation
• We first need to open the terminal
• Then we need to go to the darknet directory. Try typing:

cd darknet

• and press the Enter key.

In some computers is installed in another directory and you need to type:

cd /media/Machine_Learning/darknet

• Then we are going to type:

./darknet rnn generate cfg/rnn.cfg shakespeare.weights -srand 0

• Parameters:
• -srand N -> N determines the random seed.
• To be able to generate new texts, instead of 0 use another number.
• -len N -> change the length of text generated, default 1,000
• -seed Word -> It sets the first word of the generated text

Train the RNNs with another corpus of text

• Download a text file. Here you can find some corpus of text online:

George Meredith
http://www.gutenberg.org/ebooks/4500
Mark Twain
http://www.gutenberg.org/ebooks/3200
More
http://www.gutenberg.org/ebooks/search/?query=The+Complete+Works+of

• We train the neural network with the new text (This might take a while):

./darknet rnn train cfg/rnn.train.cfg -file filename.txt

• Generating new text with our new model

./darknet rnn generate cfg/rnn.cfg filename.weights -srand 0

• Parameters:
• -srand N -> N determines the random seed.
• To be able to generate new texts, instead of 0 use another number.
• -len N -> change the length of text generated, default 1,000
• -seed Word -> It sets the first word of the generated text

• Optional: We can modify file cfg/rnn.train.cfg file (batch = 50 is the value we set by default, the original value was 256)

More information

• RNNs with Darknet:
• https://pjreddie.com/darknet/rnns-in-darknet/
• Terminal basic tutorial:
• https://maker.pro/linux/tutorial/basic-linux-commands-for-beginners
• Recurrent Neural Networks:
• https://skymind.ai/wiki/lstm