I’m trying to use NeMo’s NMT to train a translation model from language A to language B. Employed languages present similar set of words. I am following the sample in NeMo’s github but after the training phase the model always predicts an empty string. Looking at logs I found that after the first training epoch, the output at the epoch end validation is always the following:
[NeMo I 2022-09-10 15:26:47 mt_enc_dec_model:477] Input:
[NeMo I 2022-09-10 15:26:47 mt_enc_dec_model:478] Prediction:
[NeMo I 2022-09-10 15:26:47 mt_enc_dec_model:479] Ground Truth:
but the first validation (before the first training epoch) contains a non-empty prediction.
I also tried to train a translation model based on the wmt16 dataset with the same configuration, which seems to work.
Any idea on how to solve or diagnose this problem?
The last try on an Ubuntu20.04 system resulted in the same problem also with the wmt16.