Changes from Kinyarwanda MT project

Dockerfile

@@ -1,12 +1,25 @@
-FROM python:3.8-slim
-
+#FROM python:3.8-slim
+FROM nvidia/cuda:11.5.2-devel-ubuntu20.04
 # Project setup
 
 ENV VIRTUAL_ENV=/opt/venv
 
 RUN apt-get update && apt-get clean
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        python3-dev \
+        python3-pip \
+        wget \
+        && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+#RUN apt install apt install nvidia-cuda-toolkit
+
+#RUN apt-get install cuda-cudart-11-8
+RUN apt-get update &&  apt-get install -y --no-install-recommends\
+    python3.8-venv
 
-RUN python -m venv "$VIRTUAL_ENV"
+RUN python3 -m venv "$VIRTUAL_ENV"
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 RUN pip install  --quiet --upgrade pip && \
     pip install  --quiet pip-tools
@@ -23,4 +36,4 @@ COPY . /app
 WORKDIR /app
 
 COPY ./app/nltk_pkg.py /app/nltk_pkg.py
-RUN python /app/nltk_pkg.py
+RUN python3 /app/nltk_pkg.py

Dockerfile.gpu

@@ -0,0 +1,32 @@
+# If you intend to run your MT API on a GPU server it is advisable to using this Dockerfile 
+FROM nvidia/cuda:11.5.2-devel-ubuntu20.04
+# Project setup
+
+ENV VIRTUAL_ENV=/opt/venv
+
+RUN apt-get update && apt-get clean
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        python3-dev \
+        python3-pip \
+        wget \
+        && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN apt-get update &&  apt-get install -y --no-install-recommends\
+    python3.8-venv
+
+RUN python3 -m venv "$VIRTUAL_ENV"
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+RUN pip install  --quiet --upgrade pip && \
+    pip install  --quiet pip-tools
+COPY ./requirements.txt /app/requirements.txt
+RUN pip install -r /app/requirements.txt \
+    && rm -rf /root/.cache/pip
+
+COPY . /app
+WORKDIR /app
+
+COPY ./app/nltk_pkg.py /app/nltk_pkg.py
+RUN python3 /app/nltk_pkg.py

app/__init__.py

@@ -15,8 +15,14 @@ def create_app() -> FastAPI:
 )
 
     from app.views.v1.translate import translate_v1
+    from app.views.v1.translate_html import translate_html
+    from app.views.v1.translate_markdown import translate_markdown
+
 
     app.include_router(translate_v1)
+    app.include_router(translate_html)
+    app.include_router(translate_markdown)
+
 
     @app.on_event('startup')
     async def startup_event() -> None:

app/utils/translators.py

@@ -39,7 +39,7 @@ def get_ctranslator(ctranslator_model_path: str) -> Callable:
     # ]  
 
     def translator(text, src=None, tgt=None):
-        return ctranslator.translate_batch([text])[0][0]['tokens']
+        return ctranslator.translate_batch([text],disable_unk=True)[0][0]['tokens']
 
     return translator
 
@@ -62,7 +62,9 @@ def translator(src_texts, src=None, tgt=None):
             target_prefix = [[tgt]] * len(src_texts)
             src_texts = [sent + ["</s>", src] for sent in src_texts]
 
-            translations = ctranslator.translate_batch(src_texts, target_prefix=target_prefix)
+            #translations = ctranslator.translate_batch([src_texts],disable_unk=True, target_prefix=target_prefix)
+            #translations = ctranslator.translate_batch([src_texts], target_prefix=target_prefix)
+            translations = ctranslator.translate_batch(src_texts,disable_unk=True, target_prefix=target_prefix)
             translations = [translation.hypotheses[0][1:] for translation in translations]
         else:
             translations = [s.hypotheses[0] for s in ctranslator.translate_batch(src_texts)]

app/views/v1/translate.py

@@ -12,6 +12,10 @@
 )
 from app.utils.translate import translate_text
 from app.constants import MULTIMODALCODE
+import re
+from app.views.v1.translate import *
+from transformers import pipeline
+
 
 translate_v1 = APIRouter(prefix='/api/v1/translate')
 
@@ -108,3 +112,4 @@ async def languages() -> LanguagesResponse:
     return LanguagesResponse(
         languages=config.language_codes, models=config.languages_list
     )
+

app/views/v1/translate_html.py

@@ -0,0 +1,34 @@
+from app.views.v1.translate import *
+from bs4 import BeautifulSoup, NavigableString
+import random
+import os
+
+translate_html = APIRouter(prefix='/api/v1/translate_html')
+
+@translate_html.post('/translate_page', status_code=status.HTTP_200_OK)
+async def modify_html_content(request: TranslationRequest):
+
+    model_id, src, tgt = fetch_model_data_from_request(request)
+    # Parse the HTML content
+    soup = BeautifulSoup(request.text, 'html.parser')
+
+
+    def edit_text(element,model_id,src,tgt):
+        """
+        Recursively edits the text of the given BeautifulSoup element and its children.
+        """
+        if isinstance(element, NavigableString):
+            return
+        if element.name in ['script', 'style']:
+            return
+
+        for child in element.children:
+            if isinstance(child, NavigableString):
+                edited_text = translate_text(model_id,child,src,tgt)
+                child.replace_with(edited_text)
+            else:
+                edit_text(element=child,model_id=model_id,src=src,tgt=tgt)
+
+    edit_text(element=soup,model_id=model_id,src=src,tgt=tgt)
+
+    return TranslationResponse(translation=soup)

app/views/v1/translate_markdown.py

@@ -0,0 +1,113 @@
+from fastapi import APIRouter, status
+from app.views.v1.translate import TranslationRequest, TranslationResponse, fetch_model_data_from_request, translate_text
+import re
+
+translate_markdown = APIRouter(prefix='/api/v1/translate_markdown')
+
+def remove_markdown(text):
+    """Remove markdown formatting from text while preserving positions."""
+    segments = []
+    current_pos = 0
+
+    # Define markdown patterns
+    patterns = [
+        (r'(\*\*|__)(.*?)\1', 'bold'),  # Bold
+        (r'(\*|_)(.*?)\1', 'italic'),  # Italics
+        (r'\[(.*?)\]\((.*?)\)', 'link'),  # Links
+        (r'`(.*?)`', 'code'),  # Inline code
+        (r'~~(.*?)~~', 'strikethrough')  # Strikethrough
+    ]
+
+    # Find all markdown segments
+    all_matches = []
+    for pattern, mark_type in patterns:
+        for match in re.finditer(pattern, text):
+            content = match.group(2) if mark_type != 'link' else match.group(1)
+            all_matches.append({
+                'start': match.start(),
+                'end': match.end(),
+                'content': content,
+                'original': match.group(0),
+                'type': mark_type,
+                'url': match.group(2) if mark_type == 'link' else None
+            })
+
+    # Sort matches by position
+    all_matches.sort(key=lambda x: x['start'])
+
+    # Split text into markdown and non-markdown segments
+    for match in all_matches:
+        if current_pos < match['start']:
+            # Add non-markdown text segment
+            segments.append({
+                'type': 'text',
+                'content': text[current_pos:match['start']],
+                'original': text[current_pos:match['start']]
+            })
+        segments.append(match)
+        current_pos = match['end']
+
+    # Add remaining text if any
+    if current_pos < len(text):
+        segments.append({
+            'type': 'text',
+            'content': text[current_pos:],
+            'original': text[current_pos:]
+        })
+
+    # Extract plain text for translation
+    plain_text = ''.join([seg['content'] for seg in segments])
+
+    return plain_text, segments
+
+def reapply_markdown(segments, translated_text):
+    """Reapply markdown formatting to translated text while handling different language lengths."""
+    result = ""
+    translated_segments = []
+    current_pos = 0
+
+    # Split translated text into segments based on original segment lengths and positions
+    for segment in segments:
+        # Get the next chunk of translated text
+        if segment['type'] == 'text':
+            # For non-markdown text, take the next portion of translated text
+            translated_content = translated_text[current_pos:current_pos + len(segment['content'])]
+            current_pos += len(segment['content'])
+            translated_segments.append(translated_content)
+        else:
+            # For markdown text, translate the content separately to maintain formatting
+            translated_content = translate_text(model_id, segment['content'], src, tgt)
+            translated_segments.append(translated_content)
+
+    # Reconstruct the markdown text
+    for i, segment in enumerate(segments):
+        if segment['type'] == 'text':
+            result += translated_segments[i]
+        elif segment['type'] == 'link':
+            result += f"[{translated_segments[i]}]({segment['url']})"
+        elif segment['type'] == 'bold':
+            result += f"**{translated_segments[i]}**"
+        elif segment['type'] == 'italic':
+            result += f"*{translated_segments[i]}*"
+        elif segment['type'] == 'code':
+            result += f"`{translated_segments[i]}`"
+        elif segment['type'] == 'strikethrough':
+            result += f"~~{translated_segments[i]}~~"
+
+    return result
+
+@translate_markdown.post('/markdown', status_code=status.HTTP_200_OK)
+async def translate_markdown_text(request: TranslationRequest) -> TranslationResponse:
+    global model_id, src, tgt
+    model_id, src, tgt = fetch_model_data_from_request(request)
+
+    # Split text into segments and get plain text
+    plain_text, segments = remove_markdown(request.text)
+
+    # Translate the main text
+    translated_text = translate_text(model_id, plain_text, src, tgt)
+
+    # Reapply markdown with proper translations
+    translated_markdown = reapply_markdown(segments, translated_text)
+
+    return TranslationResponse(translation=translated_markdown)

config.json

@@ -2,25 +2,16 @@
   "languages": {
     "en": "English",
     "fr": "French",
-    "de": "German",
-    "ha": "Hausa",
-    "apc":"Levantine Arabic",
-    "sw_cd":"Congolese Swahili",
-    "ti":"Tigrinya",
-    "uk":"Ukranian",
-    "kr": "Kanuri",
-    "ff": "Fulfulde",
-    "tr": "Turkish"
+    "rw": "Kinyarwanda"
   },
   "models": [
     {
       "model_type": "nllb",
       "checkpoint_id": "nllb-200-distilled-600M",
       "multilingual": true,
-      "alt": "nllb",
       "load": false,
       "sentence_split": "nltk",
-      "supported_pairs": ["en-kr", "en-fr", "en-ff", "en-ha"],
+      "supported_pairs": ["en-rw", "en-fr", "fr-rw", "fr-en","rw-en","rw-fr"],
       "pipeline": {
           "translate": true
       },
@@ -31,7 +22,6 @@
       "model_type": "m2m100",
       "checkpoint_id": "m2m100_418M",
       "multilingual": true,
-      "alt":"m2m",
       "load": false,
       "sentence_split": "nltk",
       "supported_pairs": ["en-tr"],
@@ -41,8 +31,8 @@
     },
     {
       "model_type": "ctranslator2",
-      "model_path": "nllb-200-distilled-600M-int8",
-      "alt": "education",
+      "model_path": "DigitalUmuganda/Quantized_Mbaza_MT_v1",
+      "alt": "",
       "src_sentencepiece_model": "flores200_sacrebleu_tokenizer_spm.model",
       "tgt_sentencepiece_model": "flores200_sacrebleu_tokenizer_spm.model",
       "multilingual": true,
@@ -56,7 +46,7 @@
         "translate": true,
         "recase": true
       },
-      "lang_code_map": {"en": "eng_Latn", "tr": "tur_Latn", 
+      "lang_code_map": {"en": "eng_Latn", "tr": "tur_Latn",
                         "fr": "fra_Latn", "rw": "kin_Latn"}
     },
     {
@@ -66,7 +56,7 @@
       "src_sentencepiece_model": "flores200_sacrebleu_tokenizer_spm.model",
       "tgt_sentencepiece_model": "flores200_sacrebleu_tokenizer_spm.model",
       "multilingual": true,
-      "load": true,
+      "load": false,
       "sentence_split": "nltk",
       "supported_pairs": ["en-rw", "rw-en"],
       "pipeline": {

docker-compose.yml

@@ -11,7 +11,7 @@ services:
       - 8001:8000
     volumes:
       - .:/app
-      - ../translation-models:/app/models
+      - ../../translation-models:/app/models
       - ./config.json:/app/config.json
     security_opt:
       - seccomp:unconfined
@@ -23,3 +23,9 @@ services:
       - NVIDIA_VISIBLE_DEVICES=all
       - NVIDIA_DRIVER_CAPABILITIES=all
 
+    networks:
+      - infrastructure_default
+
+networks:
+  infrastructure_default: # external network
+    external: true

main.py

@@ -1,3 +1,5 @@
 from app import create_app
-
+from prometheus_fastapi_instrumentator import Instrumentator
 app = create_app()
+#Monitoring
+Instrumentator().instrument(app).expose(app)

requirements.txt

@@ -7,3 +7,6 @@ nltk==3.8.1
 sentencepiece==0.1.99
 torch==2.1.0
 transformers==4.34.1
+tenacity==8.2.2
+prometheus-fastapi-instrumentator==5.9.1
+beautifulsoup4