app.py

"""
Hugging Face Space — Vietnamese Sign Language Recognition
Upload file .npz -> nhận diện từ ký hiệu -> dịch sang câu tiếng Việt
"""
import os
import sys
import json
import torch
import numpy as np
import gradio as gr
from huggingface_hub import hf_hub_download, snapshot_download
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

MODEL_REPO = "huy00001/vsl-recognition"

sys.path.insert(0, os.path.join(os.path.dirname(__file__), "model"))
from stgcn_ctc import STGCNCTCModel

DEVICE = "cpu"
_pipeline = None


def load_pipeline():
    global _pipeline
    if _pipeline is not None:
        return _pipeline

    # label_map
    label_map_path = hf_hub_download(repo_id=MODEL_REPO, filename="label_map.json")
    with open(label_map_path, "r", encoding="utf-8") as f:
        data = json.load(f)
    classes     = data["classes"]
    num_classes = len(classes)

    # CTC model
    ctc_path  = hf_hub_download(repo_id=MODEL_REPO, filename="checkpoints/best_ctc.pt")
    ctc_model = STGCNCTCModel(num_classes=num_classes)
    ckpt      = torch.load(ctc_path, map_location=DEVICE)
    ctc_model.load_state_dict(ckpt["model_state"])
    ctc_model.eval()

    # seq2seq model
    seq2seq_dir  = snapshot_download(repo_id=MODEL_REPO, allow_patterns="checkpoints/best_seq2seq/*")
    seq2seq_path = os.path.join(seq2seq_dir, "checkpoints", "best_seq2seq")
    tokenizer    = AutoTokenizer.from_pretrained(seq2seq_path)
    seq2seq      = AutoModelForSeq2SeqLM.from_pretrained(seq2seq_path)
    seq2seq.eval()

    _pipeline = {"ctc": ctc_model, "tokenizer": tokenizer,
                 "seq2seq": seq2seq, "classes": classes, "num_classes": num_classes}
    return _pipeline


def greedy_decode(log_probs, blank=0):
    indices = log_probs.argmax(dim=-1).tolist()
    decoded, prev = [], blank
    for idx in indices:
        if idx != blank and idx != prev:
            decoded.append(idx)
        prev = idx
    return decoded


def predict(npz_file):
    if npz_file is None:
        return "Chưa có file", ""
    try:
        p   = load_pipeline()
        seq = np.load(npz_file, allow_pickle=True)["sequence"].astype(np.float32)
        x   = torch.from_numpy(seq).unsqueeze(0)

        with torch.no_grad():
            log_probs, _ = p["ctc"](x)
            log_probs    = log_probs[:, 0, :]

        indices = greedy_decode(log_probs)
        words   = [p["classes"][i - 1] for i in indices if 1 <= i <= p["num_classes"]]

        if not words:
            return "Không nhận diện được từ nào", ""

        text    = " | ".join(words)
        enc     = p["tokenizer"](text, return_tensors="pt", truncation=True, max_length=64)
        with torch.no_grad():
            out_ids  = p["seq2seq"].generate(**enc, max_new_tokens=128, num_beams=4)
        sentence = p["tokenizer"].decode(out_ids[0], skip_special_tokens=True)

        return text, sentence
    except Exception as e:
        return f"Lỗi: {str(e)}", ""


with gr.Blocks(title="Nhận diện ngôn ngữ ký hiệu tiếng Việt") as demo:
    gr.Markdown("## 🤟 Nhận diện ngôn ngữ ký hiệu tiếng Việt")
    gr.Markdown("Upload file `.npz` chứa skeleton sequence để nhận diện.")
    file_input   = gr.File(label="File .npz", file_types=[".npz"])
    btn          = gr.Button("Nhận diện", variant="primary")
    words_out    = gr.Textbox(label="Từ ký hiệu nhận diện được")
    sentence_out = gr.Textbox(label="Câu tiếng Việt")
    btn.click(fn=predict, inputs=file_input, outputs=[words_out, sentence_out])

if __name__ == "__main__":
    demo.launch()