Readme and others
Browse files- README.md +19 -0
- inference.py +111 -0
README.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
## Model description
|
| 7 |
+
|
| 8 |
+
This is a model based on Google's MT5 finetuned for paraphrasing in Catalan language
|
| 9 |
+
|
| 10 |
+
Sample:
|
| 11 |
+
|
| 12 |
+
```
|
| 13 |
+
original: Aquesta és una associació sense ànim de lucre amb la missió de fomentar la presència i l'ús del català.
|
| 14 |
+
Aquesta és una organització sense ànim de lucre amb la finalitat de promoure la presència i l'ús del català.
|
| 15 |
+
Aquesta és una organització sense ànim de lucre que té com a objectiu promoure la presència i l'ús del català.
|
| 16 |
+
|
| 17 |
+
```
|
| 18 |
+
|
| 19 |
+
To run inference check the inference.py file in the repository.
|
inference.py
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import T5ForConditionalGeneration, T5Tokenizer
|
| 2 |
+
from transformers import AutoTokenizer, MT5Model
|
| 3 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
| 4 |
+
from transformers import MT5ForConditionalGeneration, AutoTokenizer
|
| 5 |
+
from simpletransformers.t5 import T5Model
|
| 6 |
+
import datetime
|
| 7 |
+
import logging
|
| 8 |
+
import os
|
| 9 |
+
import sys
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
class Inference:
|
| 13 |
+
def _discard_recommendations(self, original, proposal):
|
| 14 |
+
proposal = proposal.lower()
|
| 15 |
+
original = original.lower()
|
| 16 |
+
if proposal == original:
|
| 17 |
+
return True
|
| 18 |
+
|
| 19 |
+
chars = [".", "!", " ", "?", ","]
|
| 20 |
+
_proposal = proposal
|
| 21 |
+
_original = original
|
| 22 |
+
for char in chars:
|
| 23 |
+
proposal = proposal.replace(char, "")
|
| 24 |
+
original = original.replace(char, "")
|
| 25 |
+
|
| 26 |
+
if proposal == original:
|
| 27 |
+
return True
|
| 28 |
+
|
| 29 |
+
return False
|
| 30 |
+
|
| 31 |
+
# https://github.com/Vamsi995/Paraphrase-Generator/blob/master/evaluate.py
|
| 32 |
+
def get_paraphrases(
|
| 33 |
+
self,
|
| 34 |
+
model_name,
|
| 35 |
+
sentence,
|
| 36 |
+
temperature,
|
| 37 |
+
prefix="paraphrase: ",
|
| 38 |
+
n_predictions=2,
|
| 39 |
+
top_k=120,
|
| 40 |
+
max_length=256,
|
| 41 |
+
device="cpu",
|
| 42 |
+
):
|
| 43 |
+
|
| 44 |
+
model = MT5ForConditionalGeneration.from_pretrained(model_name)
|
| 45 |
+
tokenizer = T5Tokenizer.from_pretrained(model_name)
|
| 46 |
+
|
| 47 |
+
discaded = 0
|
| 48 |
+
text = prefix + sentence + " </s>"
|
| 49 |
+
encoding = tokenizer.encode_plus(
|
| 50 |
+
text, pad_to_max_length=True, return_tensors="pt"
|
| 51 |
+
)
|
| 52 |
+
input_ids, attention_masks = encoding["input_ids"].to(device), encoding[
|
| 53 |
+
"attention_mask"
|
| 54 |
+
].to(device)
|
| 55 |
+
|
| 56 |
+
do_sample = True if temperature > 0 else False
|
| 57 |
+
print(f"do_sample: {do_sample}")
|
| 58 |
+
print(f"temperature: {temperature}")
|
| 59 |
+
# https://huggingface.co/blog/how-to-generate
|
| 60 |
+
# https://huggingface.co/transformers/v3.2.0/_modules/transformers/generation_utils.html
|
| 61 |
+
model_output = model.generate(
|
| 62 |
+
input_ids=input_ids,
|
| 63 |
+
attention_mask=attention_masks,
|
| 64 |
+
do_sample=do_sample,
|
| 65 |
+
max_length=max_length,
|
| 66 |
+
top_k=top_k,
|
| 67 |
+
num_beams=n_predictions * 2, ## ask for twice since some will be discarted
|
| 68 |
+
top_p=0.98,
|
| 69 |
+
temperature=temperature,
|
| 70 |
+
early_stopping=True,
|
| 71 |
+
num_return_sequences=n_predictions * 2,
|
| 72 |
+
)
|
| 73 |
+
logging.debug(f"{len(model_output)} predictions for {sentence}")
|
| 74 |
+
outputs = []
|
| 75 |
+
for output in model_output:
|
| 76 |
+
generated_sent = tokenizer.decode(
|
| 77 |
+
output, skip_special_tokens=True, clean_up_tokenization_spaces=True
|
| 78 |
+
)
|
| 79 |
+
if (
|
| 80 |
+
self._discard_recommendations(sentence, generated_sent) is False
|
| 81 |
+
and generated_sent not in outputs
|
| 82 |
+
):
|
| 83 |
+
generated_sent = generated_sent.replace("’", "'")
|
| 84 |
+
outputs.append(generated_sent)
|
| 85 |
+
else:
|
| 86 |
+
logging.debug(f"Discarded: {generated_sent} - source:{sentence}")
|
| 87 |
+
discaded = +1
|
| 88 |
+
|
| 89 |
+
if len(outputs) == n_predictions:
|
| 90 |
+
break
|
| 91 |
+
|
| 92 |
+
return outputs
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
def main():
|
| 96 |
+
i = Inference()
|
| 97 |
+
sentence = "Aquesta és una associació sense ànim de lucre amb la missió de fomentar la presència i l'ús del català."
|
| 98 |
+
#sentence = "Softcatalà és una associació sense ànim de lucre amb la missió de fomentar la presència i l'ús del català en tots els àmbits de les noves tecnologies."
|
| 99 |
+
model = os.getcwd()
|
| 100 |
+
options = i.get_paraphrases(
|
| 101 |
+
model,
|
| 102 |
+
sentence,
|
| 103 |
+
1.0
|
| 104 |
+
)
|
| 105 |
+
print(f"original: {sentence}")
|
| 106 |
+
for option in options:
|
| 107 |
+
print(f" {option}")
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
if __name__ == "__main__":
|
| 111 |
+
main()
|