I've trained an NMT system with (say) 50,000 training steps initially. Now, I would like to train it for (say) 75000 steps. If I change the num_train_steps=75000 and re-run the training code, it is just loading the old hparams and not training further.
In other implementations like nmt.matlab and nmt.hybrid, I used to do the similar thing and it worked.
@ssokhey In additional to your original training command, adding --override_loaded_hparams=True. I don't think this feature is available at tf-1.2 branch though.
I've trained an NMT system with (say) 50,000 training steps initially. Now, I would like to train it for (say) 75000 steps. If I change the
num_train_steps=75000
and re-run the training code, it is just loading the oldhparams
and not training further.In other implementations like
nmt.matlab
andnmt.hybrid
, I used to do the similar thing and it worked.What am I missing here?