app.py

import spaces
import gradio
import json
import torch
from transformers import AutoTokenizer
from transformers import pipeline
from optimum.onnxruntime import ORTModelForSequenceClassification
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from setfit import SetFitModel, SetFitTrainer, Trainer, TrainingArguments
from setfit.exporters.utils import mean_pooling
from setfit import get_templated_dataset
from datasets import load_dataset, Dataset

# CORS Config
app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["https://statosphere-3704059fdd7e.c5v4v4jx6pq5.win","https://crunchatize-77a78ffcc6a6.c5v4v4jx6pq5.win","https://crunchatize-2-2b4f5b1479a6.c5v4v4jx6pq5.win","https://tamabotchi-2dba63df3bf1.c5v4v4jx6pq5.win","https://lord-raven.github.io"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

print(f"Is CUDA available: {torch.cuda.is_available()}")
print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")

# "xenova/mobilebert-uncased-mnli" "typeform/mobilebert-uncased-mnli" Fast but small--same as bundled in Statosphere
# "xenova/deberta-v3-base-tasksource-nli" Not impressed
# "Xenova/bart-large-mnli" A bit slow
# "Xenova/distilbert-base-uncased-mnli" "typeform/distilbert-base-uncased-mnli" Bad answers
# "Xenova/deBERTa-v3-base-mnli" "MoritzLaurer/DeBERTa-v3-base-mnli" Still a bit slow and not great answers
# "xenova/nli-deberta-v3-small" "cross-encoder/nli-deberta-v3-small" Was using this for a good while and it was...okay
model_name = "MoritzLaurer/deberta-v3-base-zeroshot-v2.0"
file_name = "onnx/model.onnx"
tokenizer_name = "MoritzLaurer/deberta-v3-base-zeroshot-v2.0"
model = ORTModelForSequenceClassification.from_pretrained(model_name, file_name=file_name, provider="CUDAExecutionProvider")
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, model_max_length=512)
classifier = pipeline(task="zero-shot-classification", model=model, tokenizer=tokenizer, device="cuda:0")

def classify(data_string, request: gradio.Request):
    if request:
        if request.headers["origin"] not in ["https://statosphere-3704059fdd7e.c5v4v4jx6pq5.win", "https://crunchatize-77a78ffcc6a6.c5v4v4jx6pq5.win", "https://crunchatize-2-2b4f5b1479a6.c5v4v4jx6pq5.win", "https://tamabotchi-2dba63df3bf1.c5v4v4jx6pq5.win", "https://ravenok-statosphere-backend.hf.space", "https://lord-raven.github.io"]:
            return "{}"
    data = json.loads(data_string)
    # if 'task' in data and data['task'] == 'few_shot_classification':
    #     return few_shot_classification(data)
    # else:
    return zero_shot_classification(data)

@spaces.GPU
def zero_shot_classification(data):
    results = classifier(data['sequence'], candidate_labels=data['candidate_labels'], hypothesis_template=data['hypothesis_template'], multi_label=data['multi_label'])
    response_string = json.dumps(results)
    return response_string

def create_sequences(data):
    # return ['###Given:\n' + data['sequence'] + '\n###End Given\n###Hypothesis:\n' + data['hypothesis_template'].format(label) + "\n###End Hypothesis" for label in data['candidate_labels']]
    return [data['sequence'] + '\n' + data['hypothesis_template'].format(label) for label in data['candidate_labels']]

# def few_shot_classification(data):
#     sequences = create_sequences(data)
#     print(sequences)
#     # results = onnx_few_shot_model(sequences)
#     probs = onnx_few_shot_model.predict_proba(sequences)
#     scores = [true[0] for true in probs]

#     composite = list(zip(scores, data['candidate_labels']))
#     composite = sorted(composite, key=lambda x: x[0], reverse=True)

#     labels, scores = zip(*composite)

#     response_dict = {'scores': scores, 'labels': labels}
#     print(response_dict)
#     response_string = json.dumps(response_dict)
#     return response_string

gradio_interface = gradio.Interface(
    fn = classify,
    inputs = gradio.Textbox(label="JSON Input"),
    outputs = gradio.Textbox()
)
gradio_interface.launch()