Closed RamoramaInteractive closed 2 years ago
What was your expected output? It's difficult to say in general without access to the training data/setup. From a Sockeye perspective it's just tokens in and tokens out and it will probably depend on the training data and whether there's a training/test mismatch.
I want a target output like this (BPE removed):
82 Gefäßerkrankungen 82 Gefäßerkrankungen 82 Gefäßerkrankungen legt fest, ob das Bewegen von Karten animiert wird legt fest, ob die Werkzeugleiste angezeigt wird ohne einen König gibt es keinen Herrschermord. vermeiden Sie sie. geschrieben von David Rogers geschrieben von James LewisMoss geschrieben von Matthew Wilcox geschrieben von W. Borgert geschrieben von Jack LewisMoss geschrieben von James Wilcox Sie suchen nach einer Vier. Sie suchen nach einer Sechs. Sie suchen nach einer Drei. Sie suchen nach einem Ass. Sie suchen nach einer Acht.
I want to replicate the model from Training Neural Machine Translation to Apply Terminology Constraints](https://aclanthology.org/P19-1294/).
I'm training with the WMT17 dataset from FairSeq. I inserted target words after source words in 5% of random lines. After that I created their corresponding source factors file by replacing all source and target terms with the numbers 1 and 2. Default words were replaced with 0. The target terms are words from the IT and Medical Vocabulary.
Do you need further information? How can I check if there's a mismatch in train/test?
Found the problem! I forgot to change the min-epoch to 50 and max-epoch to 100. No wonder my trained model worked so badly.
For my translation I created an own test data set, which includes German target words
and it's corresponding token parallel file (German words replaced with 2, source terms 1 and every default word is replaced with 0).
However, my output is not as expected:
The source terms were not replaced by the target terms. I have two assumptions: Either I have annotated far too less terms in the training set or 50 epochs of training time is too little. What do you think could be the source of the erroneous translation?