TEAMNAME team:
- @github_profile_1
- @github_profile_2
- @github_profile_3
(provide links to your github profiles only)
(Screenshot of your project board with tasks is here; keep it compact please)
Starting from raw data, we applied following preprocessing steps:
- corpora concatenating (link to github SLURM script/s): we got one big parallel text corpus of N lines
- data shuffling (link to github SLURM script/s): to feed sentences to NMT system later in the random order
- data splitting (link to github SLURM script/s): X training examples, Y test examples, and Z development examples
...
... (fill in here)
- BPE (link to github SLURM script/s): we used joint vocabulary of size N
We used Moses scripts to do basic preprocessing, and BPE for the subword segmentation.
We used 1 Tesla P100 GPU Machine provided by HPC center of the University of Tartu to train our model with vocablary of size N. The model we trained is the default OpenNMT-py model, which consists of a 2-layer LSTM with 500 hidden units on both the encoder/decoder.
We had trained our best model for ~N days, M epochs. Development set perplexity was X. We performed early stopping / waited to the end to stop the training process. You can find the script we used to run training here (link to SLURM script).
We performed an inference and got unpostprocessed English hyps file.
We used this file, processed reference file, and BLEU metric to evaluate the translation performance of our model, and got X points.
For the next milestone, we will focus on some more advanced evaluation and error analysis technics.