Spaces:
Running
Running
| import gradio as gr | |
| import os | |
| import torch | |
| import numpy as np | |
| import pandas as pd | |
| from transformers import AutoModelForSequenceClassification | |
| from transformers import AutoTokenizer | |
| from huggingface_hub import HfApi | |
| from collections import defaultdict | |
| from label_dicts import ( | |
| CAP_NUM_DICT, | |
| CAP_LABEL_NAMES, | |
| CAP_MIN_NUM_DICT, | |
| CAP_MIN_LABEL_NAMES, | |
| ) | |
| from .utils import is_disk_full, release_model | |
| HF_TOKEN = os.environ["hf_read"] | |
| languages = [ | |
| "Multilingual", | |
| ] | |
| domains = { | |
| "media": "media", | |
| "social media": "social", | |
| "parliamentary speech": "parlspeech", | |
| "legislative documents": "legislative", | |
| "executive speech": "execspeech", | |
| "executive order": "execorder", | |
| "party programs": "party", | |
| "judiciary": "judiciary", | |
| "budget": "budget", | |
| "public opinion": "publicopinion", | |
| "local government agenda": "localgovernment", | |
| } | |
| def get_label_name(idx): | |
| minor_code = CAP_MIN_NUM_DICT[idx] | |
| minor_label_name = CAP_MIN_LABEL_NAMES[minor_code] | |
| major_code = minor_code // 100 if minor_code not in [99, 999, 9999] else 999 | |
| major_label_name = CAP_LABEL_NAMES[major_code] | |
| return f"[{major_code}] {major_label_name} [{minor_code}] {minor_label_name}" | |
| def check_huggingface_path(checkpoint_path: str): | |
| try: | |
| hf_api = HfApi(token=HF_TOKEN) | |
| hf_api.model_info(checkpoint_path, token=HF_TOKEN) | |
| return True | |
| except: | |
| return False | |
| def build_huggingface_path(language: str, domain: str): | |
| if domain in ["social"]: | |
| return "poltextlab/xlm-roberta-large-twitter-cap-minor" | |
| return "poltextlab/xlm-roberta-large-pooled-cap-minor-v3" | |
| def predict(text, model_id, tokenizer_id): | |
| device = torch.device("cpu") | |
| # Load JIT-traced model | |
| jit_model_path = f"/data/jit_models/{model_id.replace('/', '_')}.pt" | |
| model = torch.jit.load(jit_model_path).to(device) | |
| model.eval() | |
| # Load tokenizer (still regular HF) | |
| tokenizer = AutoTokenizer.from_pretrained(tokenizer_id) | |
| # Tokenize input | |
| inputs = tokenizer( | |
| text, max_length=64, truncation=True, padding=True, return_tensors="pt" | |
| ) | |
| inputs = {k: v.to(device) for k, v in inputs.items()} | |
| with torch.no_grad(): | |
| output = model(inputs["input_ids"], inputs["attention_mask"]) | |
| print(output) # debug | |
| logits = output["logits"] | |
| release_model(model, model_id) | |
| probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten() | |
| output_pred = {get_label_name(i): probs[i] for i in np.argsort(probs)[::-1]} | |
| output_info = f'<p style="text-align: center; display: block">Prediction was made using the <a href="https://huggingface.co/{model_id}">{model_id}</a> model.</p>' | |
| return output_pred, output_info | |
| def predict_cap(text, language, domain): | |
| domain = domains[domain] | |
| model_id = build_huggingface_path(language, domain) | |
| tokenizer_id = "xlm-roberta-large" | |
| if is_disk_full(): | |
| os.system("rm -rf /data/models*") | |
| os.system("rm -r ~/.cache/huggingface/hub") | |
| return predict(text, model_id, tokenizer_id) | |
| demo = gr.Interface( | |
| title="CAP Minor Topics Babel Demo", | |
| fn=predict_cap, | |
| inputs=[ | |
| gr.Textbox(lines=6, label="Input"), | |
| gr.Dropdown(languages, label="Language", value=languages[0]), | |
| gr.Dropdown(domains.keys(), label="Domain", value=list(domains.keys())[0]), | |
| ], | |
| outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()], | |
| ) | |