Update app

app.py

@@ -1,64 +1,86 @@
 import gradio as gr
-from huggingface_hub import InferenceClient
+from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
+import torch
+from sacremoses import MosesPunctNormalizer
+import re
+import unicodedata
 
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
+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)
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
+# Load the small model
+small_tokenizer = NllbTokenizer.from_pretrained("hunterschep/amis-zh-600M")
+small_model = AutoModelForSeq2SeqLM.from_pretrained("hunterschep/amis-zh-600M").to(device)
 
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+# 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 = {}
 
-    messages.append({"role": "user", "content": message})
+fix_tokenizer(big_tokenizer)
+fix_tokenizer(small_tokenizer)
 
-    response = ""
+# 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]
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
+# Preprocessing for Chinese
+mpn_chinese = MosesPunctNormalizer(lang="zh")
+mpn_chinese.substitutions = [(re.compile(r), sub) for r, sub in mpn_chinese.substitutions]
 
-        response += token
-        yield response
+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(" ")
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
+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()