This is an implementation of a transformer based language model using Pytorch-Lightning framework. The model consists of two transformer layers with 2 heads each combined with embedding and linear layers. The code uses pytorch-lightning framework that is built for fast research prototyping. Although the code does not leverage all the features of the lightning framework, but it has achieved good result in the language modelling task. For learning purposes, a Pytorch classical version from this project is also provided here (inspired from Pytorch team ❤️).
The code is using pipenv
as a virtual environment and package manager. To run the code, all you need is to install the necessary dependencies. open the terminal and type:
$ git clone https://github.com/Khamies/Transformer_Lightning.git
$ cd Transformer_Lightning
$ pipenv install
or
$ pip install requirements.txt
And you should be ready to go to play with code and build upon it!
-
To train the model, run:
python main.py
-
To train the model with specific arguments, run:
python main.py --batch_size=64
. The following command-line arguments are available:- Train batch size:
--bsz_train
- Test batch size:
--bsz_test
- bptt:
--bptt
- Learning rate:
--lr
- Embedding size:
--embed_size
- Size of FeedForward Neural Network (1st layer):
--ffnn_size
- Attention Heads:
--nhead
- Transformer Layers:
--nlayers
- Train batch size:
The model is trained on 10 epochs
using Adam as an optimizer with a learning rate = 0.001
and batch size = 32
, you can find all the model settings in settings.py. Here is the loss curve for the training step:
Here are some generated samples from the model:
<time> warner owns n million francs.
<media> concern said it is a share n million.
@misc{Khamies2021Transformer_Lightning, author = {Khamies, Waleed}, title = {A pytorch implementation of transformer based language model using Pytorch-Lightning framework.}, year = {2021}, publisher = {GitHub}, journal = {GitHub repository}, howpublished = {\url{ https://github.com/Khamies/Transformer_Lightning}}, }
For any question or a collaboration, drop me a message here
Follow me on Linkedin!
Thank you ❤️