Transformer Save

A TensorFlow Implementation of the Transformer: Attention Is All You Need

Project README

[UPDATED] A TensorFlow Implementation of Attention Is All You Need

When I opened this repository in 2017, there was no official code yet. I tried to implement the paper as I understood, but to no surprise it had several bugs. I realized them mostly thanks to people who issued here, so I'm very grateful to all of them. Though there is the official implementation as well as several other unofficial github repos, I decided to update my own one. This update focuses on:

  • readable / understandable code writing
  • modularization (but not too much)
  • revising known bugs. (masking, positional encoding, ...)
  • updating to TF1.12. (, ...)
  • adding some missing components (bpe, shared weight matrix, ...)
  • including useful comments in the code.

I still stick to IWSLT 2016 de-en. I guess if you'd like to test on a big data such as WMT, you would rely on the official implementation. After all, it's pleasant to check quickly if your model works. The initial code for TF1.2 is moved to the tf1.2_lecacy folder for the record.


  • python==3.x (Let's move on to python 3 if you still use python 2)
  • tensorflow==1.12.0
  • numpy>=1.15.4
  • sentencepiece==0.1.8
  • tqdm>=4.28.1



It should be extracted to iwslt2016/de-en folder automatically.

  • STEP 2. Run the command below to create preprocessed train/eval/test data.

If you want to change the vocabulary size (default:32000), do this.

python --vocab_size 8000

It should create two folders iwslt2016/prepro and iwslt2016/segmented.

  • STEP 3. Run the following command.

Check to see which parameters are possible. For example,

python --logdir myLog --batch_size 256 --dropout_rate 0.5
  • STEP 3. Or download the pretrained models.
wget; unzip; rm

Training Loss Curve

Learning rate

Bleu score on devset

Inference (=test)

  • Run
python --ckpt log/1/iwslt2016_E19L2.64-29146 (OR yourCkptFile OR yourCkptFileDirectory)


  • Typically, machine translation is evaluated with Bleu score.
  • All evaluation results are available in eval/1 and test/1.
tst2013 (dev) tst2014 (test)
28.06 23.88


  • Beam decoding will be added soon.
  • I'm going to update the code when TF2.0 comes out if possible.
Open Source Agenda is not affiliated with "Transformer" Project. README Source: Kyubyong/transformer