Hello, I would like to train a system to translate from Italian to English. At the moment I have a parallel corpus made by 2.300.000 sentences by many corpus (Tatoeba, Europarlament, …). I am training with CPU (1300 tokens/second). I am training with the default parameters, and the accuracy stop the increasing at 58% after 40000 steps. Do I need more sentences for the training? Any tips? Since the training is very long, I would like to hear some suggestions in order to try with the best parameter setting as possible. Thanks
You should probably train a bigger model than the one used with the default parameters. See for example:
However, this will be somewhat difficult to achieve on a CPU.
perfect. Thank you. I post another question (more code related) in a new topic. Hope you can help me again. Thanks.