[WIP] add text augmentor #8
Conversation
PhilipMay
force-pushed
the
main
branch
3 times, most recently
from
858aadd to
8861279
Compare
|
Very nice. Thanks @sitongye |
|
Perhaps you can put all code in a file called |
|
The .md file (documentation) should be at |
|
|
sitongye
force-pushed
the
text_augmentation
branch
from
4a29ae6 to
722cf20
Compare
|
@sitongye I made some fixes |
|
now there are "only" pylint issues remaining. |
transformer_tools/augmentation.py
Outdated
| def _swap_with_weights(self, ori_word, prob): | ||
| # swap a word with candidates or stay the same depending on given probability | ||
| # prob: probability of being swapped | ||
| swap = self.candidate_dict.get(ori_word) # FIXME: where os candidate_dict set? |
There was a problem hiding this comment.
@sitongye can you please check this. I think candidate_dict is not available.
TODO
|
transformer_tools/augmentation.py
Outdated
| self.mid2ori_model = self._load_transmodel( | ||
| self.mid_ori_model_path, self.mid_ori_checkpoints | ||
| ) |
There was a problem hiding this comment.
@sitongye this needs a 3rd param
transformer_tools/augmentation.py
Outdated
| self.model = AutoModelForMaskedLM.from_pretrained(local_model_path) | ||
| self.nr_candidates = nr_candidates | ||
|
|
||
| def _generate(self, sequence): |
There was a problem hiding this comment.
@sitongye just rename sequence to sent?
|
Hey @sitongye |
transformer_tools/augmentation.py
Outdated
| def map_apostrophe(string): | ||
| """Replace special short forms in german back to original forms.""" | ||
| # rule based | ||
| mapping = { | ||
| "'s": " es", | ||
| "'nem": " einem", | ||
| "'ne": " eine", | ||
| "'ner": " einer", | ||
| "'nen'": " einen", | ||
| "'n": " ein", | ||
| } | ||
| for key, value in mapping.items(): | ||
| string = re.sub(key, value, string) | ||
| return string |
There was a problem hiding this comment.
please check if we realy need this
PhilipMay
force-pushed
the
text_augmentation
branch
2 times, most recently
from
1be9435 to
f83c233
Compare
|
rebased on main |
PhilipMay
force-pushed
the
text_augmentation
branch
from
77116ba to
4c88aa2
Compare
|
resolved merge conflict |
|
I think the tests fail because it does not manage to install some pip packages. Maybe fasttext.... |
PhilipMay
force-pushed
the
text_augmentation
branch
from
01d0e7d to
7e80452
Compare
|
@PhilipMay I think it's that I imported "math" but did not use it. last time the imports of libraries did not throw a problem. I will use this afternoon to check it and also test the functionalities |
For some reason spacy has problems to be installed. Now that I removed it the tests pass but only the linting has some issues. When spacy is in optional dependencies it does not even install in the CI pipeline. |
sitongye
force-pushed
the
text_augmentation
branch
from
7e80452 to
95deded
Compare
PhilipMay
force-pushed
the
text_augmentation
branch
from
95deded to
6f33414
Compare
| out = " ".join(aug_text).strip() | ||
| out = re.sub(" +", " ", out) | ||
| out = re.sub(" ,", ",", out) | ||
| out = re.sub(r" \.", ".", out) |
There was a problem hiding this comment.
@sitongye instead of this we should be able to use the whitespace_ attribute of the spacy token.
That would be a cleaner way to concatenate the text.
See: https://spacy.io/api/token#attributes
There was a problem hiding this comment.
@PhilipMay do you think it's worth it if I transform the fasttext embedding to be in the spacy format so that we can simply swap the tokens? https://spacy.io/usage/linguistic-features#vectors-similarity
There was a problem hiding this comment.
Not sure. Maybe we should write an abstraction to change the way how we swap token?
|
I think fairseq has issues with py 3.9. |
|
@PhilipMay Hi philip, seems the transformers import failed, since naming of the "transformers.py" under "transformer_tools" conflicts with the "transformers" library, how shall I resolve that? |
Do you mean the failure in the CI pipeline here in GH? our transformers module should be used as transformer.transformers while the other is just transformers. |
add initial version of text augmentor + documentation as readme.md