feat: TensorRT FP16 depth estimation backend

SKILL.md

@@ -65,9 +65,12 @@ When used for **privacy mode**, the `depth_only` blend mode fully anonymizes the
 | Platform | Backend | Runtime | Model |
 |----------|---------|---------|-------|
 | **macOS** | CoreML | Apple Neural Engine | `apple/coreml-depth-anything-v2-small` (.mlpackage) |
-| Linux/Windows | PyTorch | CUDA / CPU | `depth-anything/Depth-Anything-V2-Small` (.pth) |
+| Windows/Linux | **TensorRT** | NVIDIA TRT FP16 | Auto-built `.trt` engine (from ONNX) |
+| Windows/Linux | PyTorch | CUDA / CPU | `depth-anything/Depth-Anything-V2-Small` (.pth) |
 
-On macOS, CoreML runs on the Neural Engine, leaving the GPU free for other tasks. The model is auto-downloaded from HuggingFace and stored at `~/.aegis-ai/models/feature-extraction/`.
+On macOS, CoreML runs on the Neural Engine, leaving the GPU free for other tasks.
+On Windows/Linux with NVIDIA GPUs, TensorRT FP16 provides 2-4x speedup over vanilla PyTorch CUDA. Engines are auto-built on first run and cached at `~/.aegis-ai/models/feature-extraction/trt_engines/`.
+The model is auto-downloaded from HuggingFace and stored at `~/.aegis-ai/models/feature-extraction/`.
 
 ## What You Get
 

skills.json

@@ -137,7 +137,7 @@
       "id": "depth-estimation",
       "name": "Depth Estimation (Privacy)",
       "description": "Privacy-first depth map transforms — anonymize camera feeds with Depth Anything v2 while preserving spatial awareness.",
-      "version": "1.1.0",
+      "version": "1.2.0",
       "category": "privacy",
       "path": "skills/transformation/depth-estimation",
       "tags": [

skills/transformation/depth-estimation/deploy.bat

@@ -0,0 +1,66 @@
+@echo off
+REM deploy.bat — Windows deployment for Depth Estimation (Privacy) skill
+REM Creates venv, installs PyTorch + CUDA dependencies, verifies GPU detection.
+REM
+REM The Aegis deployment agent calls this on Windows instead of deploy.sh.
+
+setlocal enabledelayedexpansion
+
+set "SCRIPT_DIR=%~dp0"
+set "VENV_DIR=%SCRIPT_DIR%.venv"
+set "MODELS_DIR=%USERPROFILE%\.aegis-ai\models\feature-extraction"
+
+echo === Depth Estimation (Privacy) — Windows Setup ===
+
+REM ── Create venv ───────────────────────────────────────────────────
+if not exist "%VENV_DIR%" (
+    echo Creating virtual environment...
+    python -m venv "%VENV_DIR%"
+    if errorlevel 1 (
+        echo ERROR: Failed to create virtual environment. Is Python installed?
+        exit /b 1
+    )
+)
+
+set "PIP=%VENV_DIR%\Scripts\pip.exe"
+set "PYTHON=%VENV_DIR%\Scripts\python.exe"
+
+REM Upgrade pip
+"%PIP%" install --upgrade pip --quiet
+
+echo.
+echo === Windows — PyTorch backend (CUDA/CPU) ===
+echo Installing PyTorch dependencies...
+"%PIP%" install --quiet -r "%SCRIPT_DIR%requirements.txt"
+
+if errorlevel 1 (
+    echo ERROR: pip install failed. Check requirements.txt and network connectivity.
+    exit /b 1
+)
+
+echo [OK] PyTorch dependencies installed
+
+REM ── Verify installation ───────────────────────────────────────────
+"%PYTHON%" -c "import torch, cv2, numpy, PIL; from depth_anything_v2.dpt import DepthAnythingV2; cuda = 'YES' if torch.cuda.is_available() else 'NO'; gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A'; print(f'[OK] Verified: torch={torch.__version__}, CUDA={cuda}, GPU={gpu}')"
+
+if errorlevel 1 (
+    echo WARNING: Verification failed. Some packages may not be installed correctly.
+    echo Trying minimal verification...
+    "%PYTHON%" -c "import torch; print(f'torch={torch.__version__}, CUDA={torch.cuda.is_available()}')"
+)
+
+REM ── TensorRT verification (optional) ──────────────────────────────
+echo.
+echo === TensorRT Check ===
+"%PYTHON%" -c "import tensorrt; print(f'[OK] TensorRT={tensorrt.__version__}')" 2>nul
+if errorlevel 1 (
+    echo [INFO] TensorRT not available — will use PyTorch CUDA backend
+    echo [INFO] To enable TensorRT: pip install tensorrt
+) else (
+    echo [OK] TensorRT FP16 acceleration available
+)
+
+echo.
+echo === Setup complete ===
+
+endlocal

skills/transformation/depth-estimation/models.json

@@ -0,0 +1,123 @@
+{
+  "studio": {
+    "title": "3D Depth Vision Studio",
+    "subtitle": "Convert 2D video to 3D depth maps • Privacy-first scene understanding",
+    "icon": "layers"
+  },
+  "models_dir": "~/.aegis-ai/models/feature-extraction",
+  "models": {
+    "depth-anything-v2-small": {
+      "name": "Depth Anything V2 Small",
+      "type": "depth_estimation",
+      "input_size": [518, 392],
+      "description": "Monocular depth estimation — fast, lightweight model",
+      "platforms": {
+        "darwin": {
+          "repository": "apple/coreml-depth-anything-v2-small",
+          "format": "mlpackage",
+          "variants": {
+            "DepthAnythingV2SmallF16": {
+              "precision": "float16",
+              "size_mb": 49.8,
+              "description": "Float16 — Optimized for Neural Engine"
+            },
+            "DepthAnythingV2SmallF32": {
+              "precision": "float32",
+              "size_mb": 99.2,
+              "description": "Float32 — Highest precision"
+            }
+          }
+        },
+        "win32": {
+          "repository": "depth-anything/Depth-Anything-V2-Small",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vits": {
+              "precision": "float32",
+              "size_mb": 99,
+              "description": "PyTorch ViT-S — CUDA/CPU/MPS compatible"
+            },
+            "depth_anything_v2_vits_trt_fp16": {
+              "precision": "float16",
+              "format": "trt",
+              "size_mb": 25,
+              "description": "TensorRT FP16 — Fastest (requires NVIDIA GPU + TensorRT)",
+              "requires": "tensorrt"
+            }
+          }
+        },
+        "linux": {
+          "repository": "depth-anything/Depth-Anything-V2-Small",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vits": {
+              "precision": "float32",
+              "size_mb": 99,
+              "description": "PyTorch ViT-S — CUDA/CPU compatible"
+            }
+          }
+        }
+      }
+    },
+    "depth-anything-v2-base": {
+      "name": "Depth Anything V2 Base",
+      "type": "depth_estimation",
+      "input_size": [518, 392],
+      "description": "Monocular depth estimation — balanced speed and accuracy",
+      "platforms": {
+        "win32": {
+          "repository": "depth-anything/Depth-Anything-V2-Base",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vitb": {
+              "precision": "float32",
+              "size_mb": 390,
+              "description": "PyTorch ViT-B — CUDA/CPU/MPS compatible"
+            }
+          }
+        },
+        "linux": {
+          "repository": "depth-anything/Depth-Anything-V2-Base",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vitb": {
+              "precision": "float32",
+              "size_mb": 390,
+              "description": "PyTorch ViT-B — CUDA/CPU compatible"
+            }
+          }
+        }
+      }
+    },
+    "depth-anything-v2-large": {
+      "name": "Depth Anything V2 Large",
+      "type": "depth_estimation",
+      "input_size": [518, 392],
+      "description": "Monocular depth estimation — highest accuracy, more VRAM required",
+      "platforms": {
+        "win32": {
+          "repository": "depth-anything/Depth-Anything-V2-Large",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vitl": {
+              "precision": "float32",
+              "size_mb": 1340,
+              "description": "PyTorch ViT-L — CUDA recommended (1.3 GB)"
+            }
+          }
+        },
+        "linux": {
+          "repository": "depth-anything/Depth-Anything-V2-Large",
+          "format": "pth",
+          "variants": {
+            "depth_anything_v2_vitl": {
+              "precision": "float32",
+              "size_mb": 1340,
+              "description": "PyTorch ViT-L — CUDA recommended (1.3 GB)"
+            }
+          }
+        }
+      }
+    }
+  }
+}

skills/transformation/depth-estimation/requirements.txt

@@ -1,10 +1,17 @@
 # Depth Estimation — Privacy Transform Skill
 # CoreML-first on macOS (Neural Engine), PyTorch fallback on other platforms.
 #
+# INSTALL WITH: pip install --ignore-requires-python -r requirements.txt
+# The depth-anything-v2 PyPI wheel declares python_requires>=3.12 in its
+# metadata, but is pure Python (py3-none-any) and works on Python 3.11+.
+#
 # macOS: coremltools loads .mlpackage models — fast, leaves GPU free.
 # Other: PyTorch + depth-anything-v2 pip package + HF weights.
 # Common: opencv, numpy, pillow, huggingface_hub for model download.
 
+# ── CUDA (NVIDIA systems get CUDA wheels, CPU-only falls back) ───────
+--extra-index-url https://download.pytorch.org/whl/cu126
+
 # ── CoreML (macOS only) ──────────────────────────────────────────────
 coremltools>=8.0; sys_platform == "darwin"
 
@@ -20,3 +27,10 @@ numpy>=1.24.0
 opencv-python-headless>=4.8.0
 Pillow>=10.0.0
 matplotlib>=3.7.0
+
+# ── TensorRT acceleration (optional, NVIDIA only) ───────────────────
+# Provides ~2-4x speedup over vanilla PyTorch CUDA via FP16 inference.
+# Requires NVIDIA GPU with Compute Capability >= 7.0 (Turing+).
+# If not installed, the skill gracefully falls back to PyTorch.
+tensorrt>=10.0; sys_platform != "darwin"
+onnxruntime-gpu>=1.17.0; sys_platform != "darwin"