adding features to Gaussian Normalization #479
Conversation
|
|
||
| def __init__(self): | ||
| @serializable_init | ||
| def __init__(self, src_file, trg_file): |
There was a problem hiding this comment.
Could you document the arguments? Another thing to consider is that this design will cause the statistics to be recomputed even when loading a saved model. If this takes some amount of time or is otherwise inconvenient (e.g. because it requires keeping the files around at the same location), it would be possible to do something similar to the vocabs that only load the vocab from file if the i2w argument is not set, and then use saved_processed_arg() to store the result so that the vocab file will not need to be opened when loading the model: https://github.com/neulab/xnmt/blob/master/xnmt/vocab.py#L25
There was a problem hiding this comment.
I agree with @msperber here. @cindyxinyiwang, could you take a look at it?
|
@cindyxinyiwang Ping? |
|
sorry I forgot about this...will work on it soon! |
modified length stats object so that it can be passed to Gaussian Normalization; add two other conditions for using Gaussian Normalization.