aiisc-watermarking-modelv3 / paraphraser.py
jgyasu's picture
Upload folder using huggingface_hub
63b3783 verified
raw
history blame
1.59 kB
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# Function to Initialize the Model
def init_model():
para_tokenizer = AutoTokenizer.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")
para_model = AutoModelForSeq2SeqLM.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")
return para_tokenizer, para_model
# Function to Paraphrase the Text
def paraphrase(question, para_tokenizer, para_model, num_beams=10, num_beam_groups=10, num_return_sequences=10, repetition_penalty=10.0, diversity_penalty=3.0, no_repeat_ngram_size=2, temperature=0.7, max_length=64):
input_ids = para_tokenizer(
f'paraphrase: {question}',
return_tensors="pt", padding="longest",
max_length=max_length,
truncation=True,
).input_ids
outputs = para_model.generate(
input_ids, temperature=temperature, repetition_penalty=repetition_penalty,
num_return_sequences=num_return_sequences, no_repeat_ngram_size=no_repeat_ngram_size,
num_beams=num_beams, num_beam_groups=num_beam_groups,
max_length=max_length, diversity_penalty=diversity_penalty
)
res = para_tokenizer.batch_decode(outputs, skip_special_tokens=True)
return res
def generate_paraphrase(question):
para_tokenizer, para_model = init_model()
res = paraphrase(question, para_tokenizer, para_model)
return res
# print(generate_paraphrase("Donald Trump said at a campaign rally event in Wilkes-Barre, Pennsylvania, that there has “never been a more dangerous time 5since the Holocaust” to be Jewish in the United States."))