Fix HuggingFace cache permissions

.env

@@ -1,19 +0,0 @@
-APP_NAME=AI Moderation Service
-
-APP_VERSION=1.0.0
-
-HOST=0.0.0.0
-
-PORT=8000
-
-DEBUG=true
-
- # ===== AI =====
-
-TEXT_MODEL=textdetox/bert-multilingual-toxicity-classifier
-
-MODEL_CACHE_DIR=./models
-
-DEVICE=auto
-
-HF_TOKEN: str = "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"

.env.example

@@ -0,0 +1,30 @@
+APP_NAME=AI Moderation Service
+
+APP_VERSION=1.0.0
+
+HOST=0.0.0.0
+
+PORT=8000
+
+DEBUG=true
+
+# ===== AI =====
+
+TEXT_MODEL=textdetox/bert-multilingual-toxicity-classifier
+TEXT_TOXIC_THRESHOLD=0.90
+
+IMAGE_MODEL=Falconsai/nsfw_image_detection
+NSFW_THRESHOLD=0.85
+
+IMAGE_CLIP_MODEL=openai/clip-vit-base-patch32
+CLIP_THRESHOLD=0.75
+
+WEAPON_MODEL=google/owlv2-base-patch16-ensemble
+WEAPON_THRESHOLD=0.40
+WEAPON_MIN_AREA_FRACTION=0.005
+
+MODEL_CACHE_DIR=./models
+
+DEVICE=auto
+
+HF_TOKEN=<ваш-токен-hugging-face>

.gitea/workflows/deploy.yml

@@ -0,0 +1,16 @@
+name: deploy-ai
+on:
+  push:
+    branches: [main]
+
+jobs:
+  deploy:
+    runs-on: host
+    steps:
+      - uses: actions/checkout@v4
+      - name: Build image
+        run: docker build -t nakhodka-ai:latest -t nakhodka-ai:${{ github.sha }} ai-moderation
+      - name: Restart ai-service
+        run: docker compose -f /opt/nakhodka/docker-compose.yml up -d ai-service
+      - name: Cleanup
+        run: docker image prune -f

.gitignore

@@ -0,0 +1,7 @@
+.env
+.idea/
+__pycache__/
+*.pyc
+venv/
+.venv/
+ai-moderation/models/

.idea/.gitignore

@@ -1,8 +0,0 @@
-# Default ignored files
-/shelf/
-/workspace.xml
-# Editor-based HTTP Client requests
-/httpRequests/
-# Datasource local storage ignored files
-/dataSources/
-/dataSources.local.xml

.idea/amplicode-jpa.xml

@@ -1,6 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<project version="4">
-  <component name="AmplicodeJpaIdeaProjectConfig">
-    <option name="renamerInitialized" value="true" />
-  </component>
-</project>

.idea/misc.xml

@@ -1,12 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<project version="4">
-  <component name="Black">
-    <option name="sdkName" value="Python 3.10 (moderation-post)" />
-  </component>
-  <component name="ProjectRootManager">
-    <output url="file://$PROJECT_DIR$/out" />
-  </component>
-  <component name="ProjectType">
-    <option name="id" value="jpab" />
-  </component>
-</project>

.idea/modules.xml

@@ -1,8 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<project version="4">
-  <component name="ProjectModuleManager">
-    <modules>
-      <module fileurl="file://$PROJECT_DIR$/moderation-post.iml" filepath="$PROJECT_DIR$/moderation-post.iml" />
-    </modules>
-  </component>
-</project>

.idea/vcs.xml

@@ -1,6 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<project version="4">
-  <component name="VcsDirectoryMappings">
-    <mapping directory="$PROJECT_DIR$" vcs="Git" />
-  </component>
-</project>

ai-moderation/.dockerignore

@@ -0,0 +1,8 @@
+__pycache__
+*.pyc
+.env
+venv
+.venv
+models
+.idea
+.git

ai-moderation/Dockerfile

@@ -0,0 +1,23 @@
+FROM python:3.12-slim
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PIP_NO_CACHE_DIR=1
+WORKDIR /app
+
+# torch ставим отдельно в облегчённой версии для процессора (без CUDA)
+RUN pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
+
+COPY requirements.txt .
+RUN pip install -r requirements.txt
+
+COPY app ./app
+
+RUN useradd -r -m -u 1001 app \
+ && mkdir -p /models \
+ && chown app /models
+USER app
+
+ENV MODEL_CACHE_DIR=/models
+ENV HF_HOME=/models/hf
+EXPOSE 8000
+CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]

ai-moderation/app/ml/image/weapon_detector.py

@@ -0,0 +1,109 @@
+import torch
+from PIL import Image
+from transformers import Owlv2ForObjectDetection, Owlv2Processor
+
+from app.config.settings import settings
+from app.ml.image.weapon_prediction_result import WeaponPredictionResult
+
+
+class WeaponDetector:
+
+  QUERIES = [
+    "a gun",
+    "a pistol",
+    "a revolver",
+    "a rifle",
+    "a shotgun",
+    "an assault rifle",
+    "a knife",
+    "a machete",
+    "a sword",
+    "a bomb",
+    "a grenade",
+    "a crossbow",
+    "a bow and arrow",
+    "brass knuckles",
+    "a nunchaku"
+  ]
+
+  def __init__(self):
+    self.device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Loading OWLv2 weapon detector on {self.device}")
+
+    self.model = Owlv2ForObjectDetection.from_pretrained(
+        settings.WEAPON_MODEL
+    )
+    self.processor = Owlv2Processor.from_pretrained(
+        settings.WEAPON_MODEL
+    )
+
+    self.model.to(self.device)
+    self.model.eval()
+
+  def predict(
+      self,
+      image: Image.Image
+  ) -> WeaponPredictionResult:
+    inputs = self.processor(
+        text=self.QUERIES,
+        images=image,
+        return_tensors="pt"
+    )
+
+    inputs = {
+      key: value.to(self.device)
+      for key, value in inputs.items()
+    }
+
+    with torch.no_grad():
+      outputs = self.model(**inputs)
+
+    target_sizes = torch.tensor([image.size[::-1]])
+
+    results = self.processor.post_process_grounded_object_detection(
+        outputs,
+        threshold=0.0,
+        target_sizes=target_sizes
+    )[0]
+
+    scores = results["scores"].tolist()
+    labels = results["labels"].tolist()
+    boxes = results["boxes"].tolist()
+
+    image_area = image.width * image.height
+    label_scores = {}
+
+    for score, label_index, box in zip(scores, labels, boxes):
+      box_area = (box[2] - box[0]) * (box[3] - box[1])
+
+      if box_area < image_area * settings.WEAPON_MIN_AREA_FRACTION:
+        continue
+
+      if score >= settings.WEAPON_THRESHOLD:
+        query = self.QUERIES[label_index]
+        label_scores[query] = max(
+            label_scores.get(query, 0.0),
+            score
+        )
+
+    if not label_scores:
+      return WeaponPredictionResult(
+          label="normal",
+          score=0.0
+      )
+
+    best_label = max(
+        label_scores,
+        key=label_scores.get
+    )
+
+    print("====================")
+    print("WEAPON RESULT")
+    print("Scores:", label_scores)
+
+    return WeaponPredictionResult(
+        label=best_label,
+        score=label_scores[best_label],
+        detected_labels=sorted(label_scores.keys()),
+        box_count=len(label_scores)
+    )

ai-moderation/app/ml/image/weapon_prediction_result.py

@@ -0,0 +1,14 @@
+from dataclasses import dataclass, field
+
+
+@dataclass(slots=True)
+class WeaponPredictionResult:
+  label: str
+
+  score: float
+
+  detected_labels: list[str] = field(
+      default_factory=list
+  )
+
+  box_count: int = 0