Hi, is the case on the target implemented in the pytorch version? I have successfully trained a model with case feature on both source and target. However while translating I get the error :
AssertionError: Not all arguments have the same value: (1, 2)
I checked in the source code that the scripts capture the case feature in the test data (source). I tried translating a one line file and when I log
onmt/modules/Embeddings.py, I see
nfeat, in_length, in_batch 2 17 1
nfeat, in_length, in_batch 1 1 5
The first line is correct : the input sentence has a length of 17 with the case feature (2) and the batch size is 1 (because I feed the model only one sentence to translate). However, why there is only 1 as number of features for the target?
logs of training ==> I do have the case feature on both source and target
src feature 0 size = 77
tgt feature 0 size = 82
FYI, a model with the case on the source only (not on target) decreases the performance.