Status: Reference code only. No updates expected.
Implementation and visualization of a custom RNN layer with attention in Keras for translating dates.
This repository comes with a tutorial found here: https://medium.com/datalogue/attention-in-keras-1892773a4f22
Make sure you have Python 3.4+ installed.
Clone this repository to your local system
git clone https://github.com/datalogue/keras-attention.git
We recommend using GPU's otherwise training might be prohbitively slow:
pip install -r requirements-gpu.txt
If you do not have a GPU or want to prototype on your local machine:
pip install -r requirements.txt
data and run
This will create 4 files:
training.csv- data to train the model
validation.csv- data to evaluate the model and compare performance
human_vocab.json- vocabulary for the human dates
machine_vocab.json- vocabulary for the machine dates
We highly recommending having a machine with a GPU to run this software, otherwise training might be prohibitively slow. To see what arguments are accepted you can run
python run.py -h from the main directory:
usage: run.py [-h] [-e |] [-g |] [-p |] [-t |] [-v |] [-b |] optional arguments: -h, --help show this help message and exit named arguments: -e |, --epochs | Number of Epochs to Run -g |, --gpu | GPU to use -p |, --padding | Amount of padding to use -t |, --training-data | Location of training data -v |, --validation-data | Location of validation data -b |, --batch-size | Location of validation data
All parameters have default values, so if you want to just run it, you can type
python run.py. You can always stop running the model early using
You can use the script
visualize.py to visualize the attention map. We have provided sample weights and vocabularies in
weights/ so that this script can run automatically using just an example. Run with the
-h argument to see what is accepted:
usage: visualize.py [-h] -e | [-w |] [-p |] [-hv |] [-mv |] optional arguments: -h, --help show this help message and exit named arguments: -e |, --examples | Example string/file to visualize attention map for If file, it must end with '.txt' -w |, --weights | Location of weights -p |, --padding | Length of padding -hv |, --human-vocab | Path to the human vocabulary -mv |, --machine-vocab | Path to the machine vocabulary
padding parameters correspond between
visualize.py and therefore, if you change this make sure to note it. You must supply the path to the weights you want to use and an example/file of examples. An example file is provided in
Here are some example visuals you can obtain:
The model has learned that “Saturday” has no predictive value!
We can see the weirdly formatted date “January 2016 5” is incorrectly translated as 2016–01–02 where the “02” comes from the “20” in 2016
Start an issue and we will do our best to help!
As with all open source code, we could not have built this without other code out there. Special thanks to:
Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural machine translation by jointly learning to align and translate." arXiv preprint arXiv:1409.0473 (2014).