Char RNN using TensorFlow 2.0 via Docker with GPU acceleration
Based on Official TF2 Text Generation Tutorial
Trained on some cleaned up text from the Stith Thompson's Motif-Index of Folk-Literature
Build and tag the docker image to run locally
docker build . --tag thompson:v1.0.0
Start the Docker image using a local volume and the auth data for Mastodon and Twitter.
docker run --name thompson --rm --gpus all -it -v "thompson:/tf/thompson" --env CLIENT_KEY="..." --env CLIENT_SECRET="..." --env ACCESS_TOKEN="..." --env TWITTER_CONSUMER_KEY="..." --env TWITTER_CONSUMER_SECRET="..." --env TWITTER_TOKEN_KEY="..." --env TWITTER_TOKEN_SECRET="..." thompson:v1.0.0 "/tf/thompson/rnn_folkmotif.py"
Starting a container and launching a shell to train or sample data manually.
docker run --name thompson --rm --gpus all -it -v "thompson:/tf/thompson" thompson:v1.0.0 bash
Return text sample output from the model. The higher the temperature number the weirder the results get.
python eval.py --gen_len=2000 --save_path=/tf/thompson --temperature=1
Notes: Not required, but if you want to retrain using the source input.txt or train a new model using your own text.
python train.py --data_path=/tf/thompson/input.txt --save_path=/tf/thompson --epochs=500 --n_layers=3 --n_embedding=128
The past_motifs.txt file is used to store previously generated motifs to check for uniqueness. Newly generated motifs are appended to the file.
Example Output:
Son blinds messenger brings about robbery.
Lover's gift regained: borrowing from the husband and rescued.
Beast as helper.
Axe magically guarded by bird who loves to dance.
Drunken dancers and ashes.
Bleeding rock.
Men middle: what is costliest? The earth.
Riddle: what is most general? Hope.
Saint's inexhaustible rice.
Magic wheat.
Ugly picture of saint speaks so that he does not die.
Working examples running the bot can be found here:
Twitter and Mastodon