Managed alignments

In previosly posts you suggess to watch this 2 links for alignment features supported by OpenNMT

if i’m right the pytorch have more complicated options for alignment, like -lambda_align, -alignment_layer, -alignment_heads, -full_context_alignment. But for tf just setup additional dataset block & thats all.

Can i manage alignment training in tf like in pt?


The implementations are based on different papers so the options are also different.

OpenNMT-tf implements the original guided alignment paper:

while OpenNMT-py implements a more recent paper:

Thx for papers.

Is someone retrain unaligned model on data with alignment for get it alingned? Does it make sense …

I personally don’t understand your question.

Best regards,

I have model trained without alignment data. Can I retrain model with alignment data?, without training model from the begining.


I think it should work but I haven’t tested this myself.