Zero-Shot Classification

Installation

git clone https://github.com/pbmstrk/zeroshot.git
cd zeroshot
pip install .

Building a Pipeline

from transfomers import AutoModel, AutoTokenizer
from zeroshot import ZeroShotPipeline

tokenizer = AutoTokenizer.from_pretrained("deepset/sentence_bert")
model = AutoModel.from_pretrained("deepset/sentence_bert")

pipeline = ZeroShotPipeline(tokenizer, model)
pipeline.add_labels(labels)

Can also optionally add a projection matrix,

pipeline.add_projection_matrix(projection_matrix)

Example

import torch
from transformers import AutoTokenizer, AutoModel
from zeroshot import ZeroShotPipeline

tokenizer = AutoTokenizer.from_pretrained("deepset/sentence_bert")
model = AutoModel.from_pretrained("deepset/sentence_bert")

phrase = "Who are you voting for in 2020?"
labels = ['Politics', "Sports", "Fashion"]

pipeline = ZeroShotPipeline(tokenizer, model)
pipeline.add_labels(labels)

predictions = pipeline(phrase)
print(f"The phrase is about: {labels[torch.argmax(predictions)]}")
# This phrase is about: Politics