import gradio as gr import os import torch import numpy as np from transformers import AutoModelForSequenceClassification from transformers import AutoTokenizer from huggingface_hub import HfApi from label_dicts import MANIFESTO_LABEL_NAMES HF_TOKEN = os.environ["hf_read"] languages = [ "English", "French", "German", "Hungarian", "Polish" ] domains = { "parliamentary speech": "parlspeech", } def build_huggingface_path(language: str): return "poltextlab/xlm-roberta-large-pooled-sentiment" def predict(text, model_id, tokenizer_id): device = torch.device("cpu") model = AutoModelForSequenceClassification.from_pretrained(model_id, low_cpu_mem_usage=True, device_map="auto", offload_folder="offload", token=HF_TOKEN) tokenizer = AutoTokenizer.from_pretrained(tokenizer_id) model.to(device) inputs = tokenizer(text, max_length=256, truncation=True, padding="do_not_pad", return_tensors="pt").to(device) model.eval() with torch.no_grad(): logits = model(**inputs).logits probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten() output_pred = {model.config.id2label[i]: probs[i] for i in np.argsort(probs)[::-1]} output_info = f'

Prediction was made using the {model_id} model.

' return output_pred, output_info def predict_cap(text, language, domain): model_id = build_huggingface_path(language) tokenizer_id = "xlm-roberta-large" return predict(text, model_id, tokenizer_id) demo = gr.Interface( fn=predict_cap, inputs=[gr.Textbox(lines=6, label="Input"), gr.Dropdown(languages, label="Language"), gr.Dropdown(domains.keys(), label="Domain")], outputs=[gr.Label(num_top_classes=3, label="Output"), gr.Markdown()])