use run inference from Space

app.py

@@ -1,6 +1,7 @@
 import requests
 import os
 
+import fasttext
 import gradio as gr
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
 import torch
@@ -11,11 +12,8 @@ When comments are created in the community tab, detect the language of the conte
 Then, if the detected language is different from the user's language, display an option to translate it.
 """
 
-
-TRANSLATION_API_URL = "https://api-inference.huggingface.co/models/t5-base"
-LANG_ID_API_URL = "https://noe30ht5sav83xm1.us-east-1.aws.endpoints.huggingface.cloud"
+LANG_ID_API_URL = "https://q5esh83u7boq5qwd.us-east-1.aws.endpoints.huggingface.cloud"
 ACCESS_TOKEN = os.environ.get("ACCESS_TOKEN")
-ACCESS_TOKEN = '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'
 headers = {"Authorization": f"Bearer {ACCESS_TOKEN}"}
 
 model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
@@ -32,18 +30,22 @@ language_code_map = {
     "Japanese": "jpn_Jpan"
 }
 
-
-def translate_from_api(text):
-    response = requests.post(TRANSLATION_API_URL, headers=headers, json={
-        "inputs": text, "wait_for_model": True, "use_cache": True})
-
-    return response.json()[0]['translation_text']
+def identify_language(text):
+    model_file = "lid218e.bin"
+    model_full_path = os.path.join(os.path.dirname(__file__), model_file)
+    model = fasttext.load_model(model_full_path)
+    predictions = model.predict(text, k=1) # e.g., (('__label__eng_Latn',), array([0.81148803]))
+    
+    PREFIX_LENGTH = 7 # To strip away '__label__' from language code
+    language_code = predictions[0][0][PREFIX_LENGTH:]
+    return language_code
+    
 
 
 def translate(text, src_lang, tgt_lang):
     src_lang_code = language_code_map[src_lang]
     tgt_lang_code = language_code_map[tgt_lang]
-    print(f"src: {src_lang_code} tgt: {tgt_lang_code}")
+
     translation_pipeline = pipeline(
         "translation", model=model, tokenizer=tokenizer, src_lang=src_lang_code, tgt_lang=tgt_lang_code, device=device)
     result = translation_pipeline(text)
@@ -55,8 +57,10 @@ def query(text, src_lang, tgt_lang):
     lang_id_response = requests.post(LANG_ID_API_URL, headers=headers, json={
         "inputs": text, "wait_for_model": True, "use_cache": True})
     lang_id = lang_id_response.json()[0]
+    
+    language_code = identify_language(text)
 
-    return [lang_id, translation]
+    return [language_code, translation]
 
 
 examples = [

lid.176.ftz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f3472cfe8738a7b6099e8e999c3cbfae0dcd15696aac7d7738a8039db603e83
+size 938013

lid218e.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ded5749a2ad79ae9ab7c9190c7c8b97ff20d54ad8b9527ffa50107238fc7f6a
+size 1176355829

requirements.txt

@@ -1,2 +1,3 @@
+fasttext
 torch
 transformers