app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
import torch
from sacremoses import MosesPunctNormalizer
import re
import unicodedata

device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the big model
big_tokenizer = NllbTokenizer.from_pretrained("hunterschep/amis-zh-3.3B")
big_model = AutoModelForSeq2SeqLM.from_pretrained("hunterschep/amis-zh-3.3B").to(device)

# Load the small model
small_tokenizer = NllbTokenizer.from_pretrained("hunterschep/amis-zh-600M")
small_model = AutoModelForSeq2SeqLM.from_pretrained("hunterschep/amis-zh-600M").to(device)

# Fix tokenizers
def fix_tokenizer(tokenizer, new_lang='ami_Latn'):
    old_len = len(tokenizer) - int(new_lang in tokenizer.added_tokens_encoder)
    tokenizer.lang_code_to_id[new_lang] = old_len - 1
    tokenizer.id_to_lang_code[old_len - 1] = new_lang
    tokenizer.fairseq_tokens_to_ids["<mask>"] = len(tokenizer.sp_model) + len(tokenizer.lang_code_to_id) + tokenizer.fairseq_offset
    tokenizer.fairseq_tokens_to_ids.update(tokenizer.lang_code_to_id)
    tokenizer.fairseq_ids_to_tokens = {v: k for k, v in tokenizer.fairseq_tokens_to_ids.items()}
    if new_lang not in tokenizer._additional_special_tokens:
        tokenizer._additional_special_tokens.append(new_lang)
    tokenizer.added_tokens_encoder = {}
    tokenizer.added_tokens_decoder = {}

fix_tokenizer(big_tokenizer)
fix_tokenizer(small_tokenizer)

# Translation function
def translate(text, model_type, src_lang, tgt_lang):
    tokenizer, model = (big_tokenizer, big_model) if model_type == "Large" else (small_tokenizer, small_model)
    if src_lang == "zho_Hant":
        text = preproc_chinese(text)
    tokenizer.src_lang = src_lang
    tokenizer.tgt_lang = tgt_lang
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=1024)
    model.eval()
    result = model.generate(
        **inputs.to(model.device),
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_lang),
        max_new_tokens=256,
        num_beams=4
    )
    return tokenizer.batch_decode(result, skip_special_tokens=True)[0]

# Preprocessing for Chinese
mpn_chinese = MosesPunctNormalizer(lang="zh")
mpn_chinese.substitutions = [(re.compile(r), sub) for r, sub in mpn_chinese.substitutions]

def get_non_printing_char_replacer(replace_by=" "):
    non_printable_map = {ord(c): replace_by for c in (chr(i) for i in range(sys.maxunicode + 1)) if unicodedata.category(c) in {"C", "Cc", "Cf", "Cs", "Co", "Cn"}}
    return lambda line: line.translate(non_printable_map)

replace_nonprint = get_non_printing_char_replacer(" ")

def preproc_chinese(text):
    clean = text
    for pattern, sub in mpn_chinese.substitutions:
        clean = pattern.sub(sub, clean)
    clean = replace_nonprint(clean)
    return unicodedata.normalize("NFKC", clean)

# Gradio interface
def switch_direction(src_lang, tgt_lang):
    return tgt_lang, src_lang

with gr.Blocks() as demo:
    gr.Markdown("# AMIS - Chinese Translation Tool")
    model_type = gr.Radio(choices=["Small", "Large"], value="Small", label="Model Type")
    src_lang = gr.Radio(choices=["zho_Hant", "ami_Latn"], value="zho_Hant", label="Source Language")
    tgt_lang = gr.Radio(choices=["ami_Latn", "zho_Hant"], value="ami_Latn", label="Target Language")
    input_text = gr.Textbox(label="Input Text", placeholder="Enter text here...")
    output_text = gr.Textbox(label="Translated Text", interactive=False)
    translate_btn = gr.Button("Translate")
    switch_btn = gr.Button("Switch Direction")

    translate_btn.click(translate, inputs=[input_text, model_type, src_lang, tgt_lang], outputs=output_text)
    switch_btn.click(switch_direction, inputs=[src_lang, tgt_lang], outputs=[src_lang, tgt_lang])

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