amine_dubs
commited on
Commit
·
c38e2fa
1
Parent(s):
7dfe957
Implement transformers library with T5 model and custom Arabic prompt
Browse files- backend/main.py +85 -62
- backend/requirements.txt +3 -0
backend/main.py
CHANGED
@@ -9,6 +9,10 @@ import json
|
|
9 |
import traceback
|
10 |
import io
|
11 |
|
|
|
|
|
|
|
|
|
12 |
# --- Configuration ---
|
13 |
# Determine the base directory of the main.py script
|
14 |
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
|
@@ -37,18 +41,58 @@ LANGUAGE_MAP = {
|
|
37 |
"it": "Italian"
|
38 |
}
|
39 |
|
40 |
-
# ---
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
# --- Translation Function ---
|
48 |
def translate_text_internal(text: str, source_lang: str, target_lang: str = "ar") -> str:
|
49 |
"""
|
50 |
-
Translate text using
|
51 |
"""
|
|
|
|
|
52 |
if not text.strip():
|
53 |
return ""
|
54 |
|
@@ -57,8 +101,15 @@ def translate_text_internal(text: str, source_lang: str, target_lang: str = "ar"
|
|
57 |
# Get full language name for prompt
|
58 |
source_lang_name = LANGUAGE_MAP.get(source_lang, source_lang)
|
59 |
|
60 |
-
#
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
Focus on conveying the meaning elegantly using proper Balagha (Arabic eloquence).
|
63 |
Adapt any cultural references or idioms appropriately rather than translating literally.
|
64 |
Ensure the translation reads naturally to a native Arabic speaker.
|
@@ -66,62 +117,34 @@ Ensure the translation reads naturally to a native Arabic speaker.
|
|
66 |
Text to translate:
|
67 |
{text}"""
|
68 |
|
69 |
-
|
70 |
-
|
71 |
-
"facebook/m2m100_418M", # Very reliable multilingual model
|
72 |
-
"Helsinki-NLP/opus-mt-tc-big-en-ar" # Good for English to Arabic
|
73 |
-
]
|
74 |
-
|
75 |
-
for model in hf_models:
|
76 |
-
try:
|
77 |
-
print(f"Attempting translation via Hugging Face Inference API: {model}")
|
78 |
-
api_url = f"https://api-inference.huggingface.co/models/{model}"
|
79 |
-
|
80 |
-
# Different payloads based on model architecture
|
81 |
-
if "m2m" in model:
|
82 |
-
payload = {
|
83 |
-
"inputs": text,
|
84 |
-
"parameters": {
|
85 |
-
"src_lang": source_lang.upper() if source_lang != "zh" else "ZH",
|
86 |
-
"tgt_lang": "AR"
|
87 |
-
}
|
88 |
-
}
|
89 |
-
elif "opus-mt" in model:
|
90 |
-
payload = {"inputs": text}
|
91 |
-
else:
|
92 |
-
payload = {"inputs": prompt}
|
93 |
-
|
94 |
-
# No auth header for public models on free tier
|
95 |
-
response = requests.post(api_url, json=payload, timeout=30)
|
96 |
-
|
97 |
-
if response.status_code == 200:
|
98 |
-
result = response.json()
|
99 |
-
translated_text = None
|
100 |
-
|
101 |
-
# Extract text from various response formats
|
102 |
-
if isinstance(result, list) and len(result) > 0:
|
103 |
-
if isinstance(result[0], dict):
|
104 |
-
translated_text = result[0].get("translation_text") or result[0].get("generated_text")
|
105 |
-
else:
|
106 |
-
translated_text = str(result[0])
|
107 |
-
elif isinstance(result, dict):
|
108 |
-
translated_text = result.get("translation_text") or result.get("generated_text")
|
109 |
-
|
110 |
-
if translated_text:
|
111 |
-
print(f"Translation successful using {model}")
|
112 |
-
return culturally_adapt_arabic(translated_text)
|
113 |
-
|
114 |
-
print(f"Unexpected response format: {response.text}")
|
115 |
-
else:
|
116 |
-
print(f"API error: {response.status_code}")
|
117 |
|
118 |
-
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
-
|
122 |
-
for endpoint in LIBRE_TRANSLATE_ENDPOINTS:
|
123 |
try:
|
124 |
-
print(f"Attempting translation using LibreTranslate: {endpoint}")
|
125 |
payload = {
|
126 |
"q": text,
|
127 |
"source": source_lang if source_lang != "auto" else "auto",
|
|
|
9 |
import traceback
|
10 |
import io
|
11 |
|
12 |
+
# Import transformers for local model inference
|
13 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
|
14 |
+
import torch
|
15 |
+
|
16 |
# --- Configuration ---
|
17 |
# Determine the base directory of the main.py script
|
18 |
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
|
|
|
41 |
"it": "Italian"
|
42 |
}
|
43 |
|
44 |
+
# --- Set cache directory to a writeable location ---
|
45 |
+
# This is crucial for Hugging Face Spaces where /app/.cache is not writable
|
46 |
+
# Using /tmp which is typically writable in most environments
|
47 |
+
os.environ['TRANSFORMERS_CACHE'] = '/tmp/transformers_cache'
|
48 |
+
os.environ['HF_HOME'] = '/tmp/hf_home'
|
49 |
+
os.environ['XDG_CACHE_HOME'] = '/tmp/cache'
|
50 |
+
|
51 |
+
# --- Global model and tokenizer variables ---
|
52 |
+
translator = None
|
53 |
+
tokenizer = None
|
54 |
+
|
55 |
+
# --- Model initialization function ---
|
56 |
+
def initialize_model():
|
57 |
+
"""Initialize the translation model and tokenizer."""
|
58 |
+
global translator, tokenizer
|
59 |
+
|
60 |
+
try:
|
61 |
+
print("Initializing model and tokenizer...")
|
62 |
+
|
63 |
+
# Use a smaller model that works well for instruction-based translation
|
64 |
+
model_name = "google/flan-t5-small"
|
65 |
+
|
66 |
+
# Load the model and tokenizer with explicit cache directory
|
67 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
68 |
+
model_name,
|
69 |
+
cache_dir="/tmp/transformers_cache"
|
70 |
+
)
|
71 |
+
|
72 |
+
# Create a pipeline for text2text generation
|
73 |
+
translator = pipeline(
|
74 |
+
"text2text-generation",
|
75 |
+
model=model_name,
|
76 |
+
tokenizer=tokenizer,
|
77 |
+
device=-1, # Use CPU for compatibility (-1) or GPU if available (0)
|
78 |
+
cache_dir="/tmp/transformers_cache",
|
79 |
+
max_length=512
|
80 |
+
)
|
81 |
+
|
82 |
+
print(f"Model {model_name} successfully initialized")
|
83 |
+
return True
|
84 |
+
except Exception as e:
|
85 |
+
print(f"Error initializing model: {e}")
|
86 |
+
traceback.print_exc()
|
87 |
+
return False
|
88 |
|
89 |
# --- Translation Function ---
|
90 |
def translate_text_internal(text: str, source_lang: str, target_lang: str = "ar") -> str:
|
91 |
"""
|
92 |
+
Translate text using local T5 model with prompt engineering
|
93 |
"""
|
94 |
+
global translator
|
95 |
+
|
96 |
if not text.strip():
|
97 |
return ""
|
98 |
|
|
|
101 |
# Get full language name for prompt
|
102 |
source_lang_name = LANGUAGE_MAP.get(source_lang, source_lang)
|
103 |
|
104 |
+
# Initialize the model if it hasn't been loaded yet
|
105 |
+
if translator is None:
|
106 |
+
success = initialize_model()
|
107 |
+
if not success:
|
108 |
+
return fallback_translate(text, source_lang, target_lang)
|
109 |
+
|
110 |
+
try:
|
111 |
+
# Construct our eloquent Arabic translation prompt
|
112 |
+
prompt = f"""Translate the following {source_lang_name} text into Modern Standard Arabic (Fusha).
|
113 |
Focus on conveying the meaning elegantly using proper Balagha (Arabic eloquence).
|
114 |
Adapt any cultural references or idioms appropriately rather than translating literally.
|
115 |
Ensure the translation reads naturally to a native Arabic speaker.
|
|
|
117 |
Text to translate:
|
118 |
{text}"""
|
119 |
|
120 |
+
# Generate translation using the model
|
121 |
+
outputs = translator(prompt, max_length=512, do_sample=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
+
if outputs and len(outputs) > 0:
|
124 |
+
translated_text = outputs[0]['generated_text']
|
125 |
+
print(f"Translation successful using transformers model")
|
126 |
+
return culturally_adapt_arabic(translated_text)
|
127 |
+
else:
|
128 |
+
print("Model returned empty output")
|
129 |
+
return fallback_translate(text, source_lang, target_lang)
|
130 |
+
|
131 |
+
except Exception as e:
|
132 |
+
print(f"Error in model translation: {e}")
|
133 |
+
traceback.print_exc()
|
134 |
+
return fallback_translate(text, source_lang, target_lang)
|
135 |
+
|
136 |
+
def fallback_translate(text: str, source_lang: str, target_lang: str = "ar") -> str:
|
137 |
+
"""Fallback to online translation APIs if local model fails."""
|
138 |
+
# Try LibreTranslate
|
139 |
+
libre_translate_endpoints = [
|
140 |
+
"https://translate.terraprint.co/translate",
|
141 |
+
"https://libretranslate.de/translate",
|
142 |
+
"https://translate.argosopentech.com/translate"
|
143 |
+
]
|
144 |
|
145 |
+
for endpoint in libre_translate_endpoints:
|
|
|
146 |
try:
|
147 |
+
print(f"Attempting fallback translation using LibreTranslate: {endpoint}")
|
148 |
payload = {
|
149 |
"q": text,
|
150 |
"source": source_lang if source_lang != "auto" else "auto",
|
backend/requirements.txt
CHANGED
@@ -5,3 +5,6 @@ PyMuPDF
|
|
5 |
requests
|
6 |
python-multipart
|
7 |
jinja2
|
|
|
|
|
|
|
|
5 |
requests
|
6 |
python-multipart
|
7 |
jinja2
|
8 |
+
transformers
|
9 |
+
torch
|
10 |
+
sentencepiece
|