app.py

import json
import torch
from transformers import AutoTokenizer
from transformers import AutoModelForSequenceClassification
import gradio as gr

with open('tag_to_name.json', 'r') as fin:
    tag_to_name = json.load(fin)
    id_to_name = dict(zip(range(len(tag_to_name)), tag_to_name.values()))

tokenizer = AutoTokenizer.from_pretrained('checkpoints/checkpoint-5000/')
model = AutoModelForSequenceClassification.from_pretrained(
    'checkpoints/checkpoint-5000/')
model.eval()

TITLE_DEFAULT = "Attention Is All You Need"


@torch.no_grad()
def predict(title, abstract, top_p, top_k):
    if abstract:
        text = title + "[SEP]" + abstract
    else:
        text = title
    tokenized = tokenizer(text, truncation=True,
                          return_tensors='pt', max_length=512)
    probs = model(**tokenized).logits[0].softmax(0)
    top_probs, top_inds = probs.sort(descending=True)
    mask = top_probs.cumsum(0) <= top_p
    if not mask.all():
        mask[mask.sum()] = True
    mask[top_k:] = False
    mask[0] = True
    predicted_ids = top_inds[mask].tolist()
    predicted_probs = top_probs[mask].tolist()
    predicted_names = [id_to_name[id_] for id_ in predicted_ids]
    return {name: prob for name, prob in zip(predicted_names, predicted_probs)}


def inference(
    title,
    abstract,
    top_p,
    top_k,
):
    if not title:
        title = TITLE_DEFAULT
    return predict(title, abstract, top_p, top_k)


g = gr.Interface(
    fn=inference,
    inputs=[
        gr.components.Textbox(
            lines=2, label="Title", placeholder=TITLE_DEFAULT
        ),
        gr.components.Textbox(lines=4, label="Abstract", placeholder=""),
        gr.components.Slider(minimum=0, maximum=1, value=0.95, label="Top p"),
        gr.components.Slider(minimum=1, maximum=len(tag_to_name),
                             step=1, value=10, label="Top n"),
    ],
    outputs=gr.outputs.Label(label="Predicted categories"),
    title="🪄 arXiv classifier 🪄",
)
g.launch()