Issues related to validation about dl4mt-multi HOT 3 CLOSED

HI @jayparks , thank you for pointing these out. Let me answer them one by one:

1. bleu_val_freq is actually a legacy option, which is ignored completely (as you've observed). The reason why we do not evaluate BLEU during training is because it takes quite a bit of time and training pauses during decoding. So I removed this functionality to make code less dilated. What I do for BLEU evaluation is, calling evaluate_models.py script while the model is actually training. It runs on CPU and evaluates models (on BLEU) as they are saved by the training main loop. By doing so, you do not pause training to do evaluation. I'd suggest you to take a look at evaluate_models.py script. If you still want to use BLEU validation functionality during training, you can take a look at this example and adapt it respectively (let me know if you decide to do so, i can outline more instructions).
2. This is an old bug that i fixed but apparently forgot to push to this repo. The problem is about the batch sizes during validation. log_prob_bs should divide the number of lines in the validation set without remainder, else, it raises this error :) The fix is very easy, and i will push as soon as i can (i'm travelling atm).

from dl4mt-multi.

Ah i see. I've tried evaluate_models.py and it works like a charm!
Appreciate your quick reply during holiday :)

from dl4mt-multi.

@jayparks, now we allow arbitrary batch sizes for logprob streams, with c0342a0 and by default batch size is 100.

from dl4mt-multi.