DIC_EVAL_PATH in word_translation sticks in the default path about muse HOT 7 CLOSED

This problem has been solved, thank you！！！

from muse.

Yes, having an optional path for the dictionary is a good idea, thank you for the suggestion! We will add this soon.

from muse.

If you make DIC_EVAL_PATH configurable it would be great if you could also allow dictionaries to contain upper-case words. I have manipulated the dictionaries to fit my word embeddings (which are not lower-cased), and get an error due to an assertion in word_translation.py:

/opt/muse/src/evaluation/word_translation.py in load_dictionary(path, word2id1, word2id2)
     56     with open(path, 'r') as f:
     57         for _, line in enumerate(f):
---> 58             assert line == line.lower()
     59             word1, word2 = line.rstrip().split()
     60             if word1 in word2id1 and word2 in word2id2:

AssertionError: 

from muse.

python unsupervised.py --src_lang zh --tgt_lang mn --src_emb ~/models/vecs/zh-1220.vec.zh --tgt_emb ~/models/vecs/mn.vec.mn --dico_method nn
when I use this method, there is an error: assert mf <= min(len(self.src_dico), len(self.tgt_dico))

from muse.

Now there is an option to specify another evaluation dictionary: 27e193a

from muse.

If you make DIC_EVAL_PATH configurable it would be great if you could also allow dictionaries to contain upper-case words. I have manipulated the dictionaries to fit my word embeddings (which are not lower-cased), and get an error due to an assertion in word_translation.py:

/opt/muse/src/evaluation/word_translation.py in load_dictionary(path, word2id1, word2id2)
     56     with open(path, 'r') as f:
     57         for _, line in enumerate(f):
---> 58             assert line == line.lower()
     59             word1, word2 = line.rstrip().split()
     60             if word1 in word2id1 and word2 in word2id2:

AssertionError: 

could you tell me how to solve this problem?

from muse.

The lowercased dictionaries can handle the non-lowercased embeddings. The idea is the following: if for instance "london" is in the dictionary, and both "london" and "London" are in the embeddings, the model will use the embedding of the most frequent word between "london" and "London". In fastText, the words are sorted by decreasing frequency, so "London" will be the selected embeddings, which is probably what we want to have.

from muse.