lab_transformer

Install instructions

install with:

conda env create --file environment.yml
conda activate pytorch_transformer

and if data does not exist run the dataset (en-de, de-en) download:

# cd is important, otherwise data goes into wrong folder
cd helpers 
python helpers/iwslt_setup.py 

Data:

Please Check before running training: The data should be in the same directory as the python script in which you are executing torchtext.datasets.IWSLT.splits, e.g.:

transformer-annotated
    |- transformer.py 
    |
    |- .data
        |- iwslt
            |- en-de.tgz
            |- de-en.tgz    

The Annotated Transformer:

1. blog post
2. Code from

Project specific

• 2021 DL Lab Project Group "Different Functionality, Different Optimization" - Google Slides

Notes:

copy files from remote to local:

scp -r [email protected]:/work/ws/nemo/fr_as1464-transformer_work-0/transformer-main/experiments_save/<NAME_OF_EXPERIMENT> /home/mrrobot/PycharmProjects/transformer-main/experiments_save

Start TENSORBOARD

# Within the project main folder:
tensorboard --logdir experiments_save/runs

Bleu Score

• https://machinelearningmastery.com/calculate-bleu-score-for-text-python/

Transformer

• Paper link: https://arxiv.org/pdf/1706.03762.pdf
• Suggested theory: https://jalammar.github.io/illustrated-transformer/

changes to Attention is all you need

Contains the implementation of the original transformer paper "Attention is all you need".

Paper link: https://arxiv.org/pdf/1706.03762.pdf

Certain modifications:

1. LayerNorm (before instead of after)
2. Dropout (Added additionally to attention weights and point-wise feed-forward net sublayer