Chirayu's picture
Update README.md
2823503
|
raw
history blame
1.38 kB
# This model predicts the sentiment('Negative'/'Positive') for the input sentence. It is fine-tuned mt5-small
The present model supports 6 languages -
1) English
2) Hindi
3) German
4) Korean
5) Japanese
6) Portuguese
Here is how to use this model
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
model = AutoModelForSeq2SeqLM.from_pretrained("Chirayu/mt5-multilingual-sentiment")
tokenizer = AutoTokenizer.from_pretrained("Chirayu/mt5-multilingual-sentiment")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
def get_sentiment(text, num_beams=2,max_length=512, repetition_penalty=2.5, length_penalty=1, early_stopping=True,top_p=.95, top_k=50, num_return_sequences=1):
input_ids = tokenizer.encode(
text, return_tensors="pt", add_special_tokens=True
)
input_ids = input_ids.to(device)
generated_ids = model.generate(
input_ids=input_ids,
num_beams=num_beams,
max_length=max_length,
repetition_penalty=repetition_penalty,
length_penalty=length_penalty,
early_stopping=early_stopping,
top_p=top_p,
top_k=top_k,
num_return_sequences=num_return_sequences,
)
sentiment = [tokenizer.decode(generated_id,skip_special_tokens=True,clean_up_tokenization_spaces=True,) for generated_id in generated_ids]
return sentiment
```