bleurt-pytorch

Use BLEURT models in native PyTorch with Transformers.

Getting started

Install with:

pip install git+https://github.com/lucadiliello/bleurt-pytorch.git

Now load your favourite model with:

import torch
from bleurt_pytorch import BleurtConfig, BleurtForSequenceClassification, BleurtTokenizer

config = BleurtConfig.from_pretrained('lucadiliello/BLEURT-20-D12')
model = BleurtForSequenceClassification.from_pretrained('lucadiliello/BLEURT-20-D12')
tokenizer = BleurtTokenizer.from_pretrained('lucadiliello/BLEURT-20-D12')

references = ["a bird chirps by the window", "this is a random sentence"]
candidates = ["a bird chirps by the window", "this looks like a random sentence"]

model.eval()
with torch.no_grad():
    inputs = tokenizer(references, candidates, padding='longest', return_tensors='pt')
    res = model(**inputs).logits.flatten().tolist()
print(res)
# [0.9604414105415344, 0.8080050349235535]

You can find all BLUERT models adapted for PyTorch here. The recommended model is lucadiliello/BLEURT-20, however this model is very large and may require too much resources. BLEURT-20-D12 is smaller but works well enough for most comparisons.

Credits

• Google original BLEURT implementation
• Transformers project
• Users of this issue, from which I took inspiration.