feat: DepthAnythingV1 and V3

CHANGELOG.rst

@@ -1,3 +1,14 @@
+Unreleased
+----------
+
+Added
+^^^^^
+* ``models.DepthAnythingV1`` and ``models.DepthAnythingV3`` aliases for the
+  ``DepthEstimator`` pipeline, mirroring the existing ``DepthAnythingV2``
+* Download configs and ONNX export functions for Depth-Anything-V1
+  (small / base / large) and Depth-Anything-V3 (mono-large, metric-large)
+
+
 0.6.1 (2025-06-17)
 ------------------
 
@@ -139,7 +150,7 @@ Added
 
 Added
 ^^^^^
-* Variations of ``impls.yolo.YOLO``: YOLO7, YOLO8, YOLO9, YOLO10, and YOLOX
+* Variations of ``impls.yolo.YOLO``: YOLOv7, YOLOv8, YOLOv9, YOLOv10, and YOLOX
 
 Changed
 ^^^^^^^

src/trtutils/download/_download.py

@@ -16,7 +16,9 @@
 from ._models import (
     export_deim,
     export_deimv2,
+    export_depth_anything_v1,
     export_depth_anything_v2,
+    export_depth_anything_v3,
     export_dfine,
     export_rfdetr,
     export_rtdetrv1,
@@ -180,8 +182,12 @@ def download_model(
         export_func = export_dfine
     elif "rfdetr" in model:
         export_func = export_rfdetr
+    elif "depth_anything_v1" in model:
+        export_func = export_depth_anything_v1
     elif "depth_anything_v2" in model:
         export_func = export_depth_anything_v2
+    elif "depth_anything_v3" in model:
+        export_func = export_depth_anything_v3
 
     # Single call site
     if export_func is None:

src/trtutils/download/_models/__init__.py

@@ -4,7 +4,11 @@
 from __future__ import annotations
 
 from ._deim import export_deim, export_deimv2
-from ._depth_anything import export_depth_anything_v2
+from ._depth_anything import (
+    export_depth_anything_v1,
+    export_depth_anything_v2,
+    export_depth_anything_v3,
+)
 from ._dfine import export_dfine
 from ._rfdetr import export_rfdetr
 from ._rtdetr import export_rtdetrv1, export_rtdetrv2, export_rtdetrv3
@@ -22,7 +26,9 @@
 __all__ = [
     "export_deim",
     "export_deimv2",
+    "export_depth_anything_v1",
     "export_depth_anything_v2",
+    "export_depth_anything_v3",
     "export_dfine",
     "export_rfdetr",
     "export_rtdetrv1",

src/trtutils/download/_models/_depth_anything.py

@@ -3,6 +3,7 @@
 # MIT License
 from __future__ import annotations
 
+import os
 import shutil
 from typing import TYPE_CHECKING
 
@@ -20,6 +21,91 @@
     from pathlib import Path
 
 
+_DA_V1_COMMIT = "1d03336771fe09c5398ffdd211441e33941a97dc"
+_DA_V3_COMMIT = "41736238f5bced4debf3f2a12375d2466874866d"
+_DA_V3_MIN_OPSET = 20
+
+
+def export_depth_anything_v1(
+    directory: Path,
+    config: dict[str, str],
+    python_path: Path,
+    bin_path: Path,
+    model: str,
+    opset: int,
+    imgsz: int | None = None,
+    *,
+    no_cache: bool | None = None,
+    no_uv_cache: bool | None = None,
+    no_warn: bool | None = None,
+    verbose: bool | None = None,
+) -> Path:
+    if not no_warn:
+        LOG.warning(
+            "DepthAnythingV1 is a Apache-2.0 licensed model, be aware of license restrictions"
+        )
+    imgsz = handle_imgsz(imgsz, 518, "DepthAnythingV1", enforce=True)
+    git_clone(
+        "https://github.com/LiheYoung/Depth-Anything",
+        directory,
+        _DA_V1_COMMIT,
+        no_cache=no_cache,
+        verbose=verbose,
+    )
+    da_v1_dir = directory / "Depth-Anything"
+    run_uv_pip_install(
+        da_v1_dir,
+        bin_path.parent,
+        "depth_anything_v1",
+        no_cache=no_uv_cache,
+        verbose=verbose,
+    )
+
+    encoder = config["encoder"]
+
+    program = f"""
+import torch
+
+from depth_anything.dpt import DepthAnything
+
+encoder = "{encoder}"
+
+depth_anything = DepthAnything.from_pretrained(f"LiheYoung/depth_anything_{{encoder}}14")
+depth_anything = depth_anything.to("cpu").eval()
+
+dummy_input = torch.ones((3, {imgsz}, {imgsz})).unsqueeze(0)
+
+example_output = depth_anything.forward(dummy_input)
+
+onnx_path = f"depth_anything_v1_{{encoder}}.onnx"
+
+torch.onnx.export(
+    depth_anything,
+    dummy_input,
+    onnx_path,
+    opset_version={opset},
+    input_names=["input"],
+    output_names=["output"],
+    dynamic_axes={{"input": {{0: "batch_size"}}, "output": {{0: "batch_size"}}}},
+    dynamo=False,
+)
+"""
+    run_cmd(
+        [
+            python_path,
+            "-c",
+            program,
+        ],
+        cwd=da_v1_dir,
+        verbose=verbose,
+    )
+
+    model_path = da_v1_dir / f"depth_anything_v1_{encoder}.onnx"
+    new_model_path = model_path.with_name(model + model_path.suffix)
+    shutil.move(model_path, new_model_path)
+    return new_model_path
+
+
 def export_depth_anything_v2(
     directory: Path,
     config: dict[str, str],
@@ -128,3 +214,165 @@ def export_depth_anything_v2(
     new_model_path = model_path.with_name(model + model_path.suffix)
     shutil.move(model_path, new_model_path)
     return new_model_path
+
+
+def export_depth_anything_v3(
+    directory: Path,
+    config: dict[str, str],
+    python_path: Path,
+    bin_path: Path,
+    model: str,
+    opset: int,
+    imgsz: int | None = None,
+    *,
+    no_cache: bool | None = None,
+    no_uv_cache: bool | None = None,
+    no_warn: bool | None = None,
+    verbose: bool | None = None,
+) -> Path:
+    if not no_warn:
+        LOG.warning(
+            "DepthAnythingV3 monocular variants are Apache-2.0 licensed, be aware of license "
+            "restrictions. ONNX export uses CUDA when available and falls back to CPU; the CPU "
+            "path works but is significantly slower for the LARGE checkpoints."
+        )
+    # ros2-depth-anything-v3-trt pins opset 20 for V3.
+    opset = max(opset, _DA_V3_MIN_OPSET)
+    imgsz = handle_imgsz(imgsz, 518, "DepthAnythingV3", adjust_div=14)
+    git_clone(
+        "https://github.com/ByteDance-Seed/depth-anything-3",
+        directory,
+        _DA_V3_COMMIT,
+        no_cache=no_cache,
+        verbose=verbose,
+    )
+    da_v3_dir = directory / "depth-anything-3"
+
+    # Two patches to V3's api.py:
+    #   1. Force fp16 autocast (ONNX does not fully support bfloat16). See
+    #      https://github.com/ika-rwth-aachen/ros2-depth-anything-v3-trt/blob/main/onnx/README.md
+    #   2. Drop the eager `from depth_anything_3.utils.export import export` import. The
+    #      monocular forward path does not need it, but it transitively imports moviepy,
+    #      pycolmap, plyfile, etc. — including pycolmap, which has no aarch64 wheels.
+    api_path = da_v3_dir / "src" / "depth_anything_3" / "api.py"
+    api_src = api_path.read_text(encoding="utf-8")
+    bf16_line = (
+        "autocast_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16"
+    )
+    fp16_line = "autocast_dtype = torch.float16"
+    export_import_line = "from depth_anything_3.utils.export import export"
+    for required in (bf16_line, export_import_line):
+        if required not in api_src:
+            err_msg = (
+                f"Failed to apply DepthAnythingV3 ONNX patch: expected line not found "
+                f"in {api_path}: {required!r}. The pinned commit ({_DA_V3_COMMIT}) may "
+                f"need to be updated."
+            )
+            raise RuntimeError(err_msg)
+    api_src = api_src.replace(bf16_line, fp16_line)
+    api_src = api_src.replace(export_import_line, f"# patched out: {export_import_line}")
+    api_path.write_text(api_src, encoding="utf-8")
+
+    run_uv_pip_install(
+        da_v3_dir,
+        bin_path.parent,
+        "depth_anything_v3",
+        no_cache=no_uv_cache,
+        verbose=verbose,
+    )
+    # Use PYTHONPATH to expose the cloned package without `pip install -e .`. The V3
+    # pyproject pulls heavy deps (pycolmap, open3d, xformers, etc.) which are not all
+    # needed for ONNX export and lack wheels on some platforms (e.g. aarch64).
+    pythonpath = str((da_v3_dir / "src").resolve())
+
+    # Cache the HF snapshot under our weights cache so re-runs are fast.
+    weights_cache_dir = get_weights_cache_dir()
+    snapshot_cache = weights_cache_dir / model
+    repo_id = config["repo_id"]
+    snapshot_dir = directory / model
+
+    if snapshot_cache.exists() and not no_cache:
+        LOG.info(f"Using cached DepthAnythingV3 snapshot: {model}")
+        shutil.copytree(snapshot_cache, snapshot_dir)
+    else:
+        LOG.info(f"Downloading DepthAnythingV3 snapshot from HF Hub: {repo_id}")
+        snapshot_dir.mkdir(parents=True, exist_ok=True)
+        run_cmd(
+            [
+                python_path,
+                "-c",
+                (
+                    "from huggingface_hub import snapshot_download; "
+                    f'snapshot_download(repo_id="{repo_id}", local_dir="{snapshot_dir}")'
+                ),
+            ],
+            cwd=da_v3_dir,
+            verbose=verbose,
+        )
+        if snapshot_cache.exists():
+            shutil.rmtree(snapshot_cache)
+        shutil.copytree(snapshot_dir, snapshot_cache)
+        LOG.info(f"Cached DepthAnythingV3 snapshot: {model}")
+
+    onnx_filename = f"{model}.onnx"
+    program = f"""
+import os
+os.environ["TORCHDYNAMO_DISABLE"] = "1"
+
+import torch
+from torch import nn
+
+from depth_anything_3.api import DepthAnything3
+
+
+class DepthAnything3OnnxWrapper(nn.Module):
+    def __init__(self, api_model):
+        super().__init__()
+        self.model = api_model
+
+    def forward(self, image):
+        out = self.model(
+            image.unsqueeze(1),
+            extrinsics=None,
+            intrinsics=None,
+            export_feat_layers=[],
+            infer_gs=False,
+        )
+        return out["depth"]
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+api_model = DepthAnything3.from_pretrained(r"{snapshot_dir}")
+api_model = api_model.to(device).eval()
+
+wrapper = DepthAnything3OnnxWrapper(api_model).to(device)
+dummy_input = torch.zeros(1, 3, {imgsz}, {imgsz}, device=device)
+
+with torch.no_grad():
+    _ = wrapper(dummy_input)
+    torch.onnx.export(
+        wrapper,
+        dummy_input,
+        r"{onnx_filename}",
+        export_params=True,
+        opset_version={opset},
+        do_constant_folding=True,
+        input_names=["input"],
+        output_names=["output"],
+        dynamic_axes={{"input": {{0: "batch_size"}}, "output": {{0: "batch_size"}}}},
+        training=torch.onnx.TrainingMode.EVAL,
+        dynamo=False,
+    )
+"""
+    env = {**os.environ, "PYTHONPATH": pythonpath}
+    run_cmd(
+        [python_path, "-c", program],
+        cwd=da_v3_dir,
+        env=env,
+        verbose=verbose,
+    )
+
+    model_path = da_v3_dir / onnx_filename
+    new_model_path = directory / onnx_filename
+    shutil.move(model_path, new_model_path)
+    return new_model_path

src/trtutils/download/configs/depth_anything_v1.json

@@ -0,0 +1,17 @@
+{
+    "depth_anything_v1_small": {
+        "url": "huggingface_hub",
+        "name": "depth_anything_v1_small",
+        "encoder": "vits"
+    },
+    "depth_anything_v1_base": {
+        "url": "huggingface_hub",
+        "name": "depth_anything_v1_base",
+        "encoder": "vitb"
+    },
+    "depth_anything_v1_large": {
+        "url": "huggingface_hub",
+        "name": "depth_anything_v1_large",
+        "encoder": "vitl"
+    }
+}

src/trtutils/download/configs/depth_anything_v3.json

@@ -0,0 +1,12 @@
+{
+    "depth_anything_v3_mono_large": {
+        "url": "huggingface_hub",
+        "name": "depth_anything_v3_mono_large",
+        "repo_id": "depth-anything/DA3MONO-LARGE"
+    },
+    "depth_anything_v3_metric_large": {
+        "url": "huggingface_hub",
+        "name": "depth_anything_v3_metric_large",
+        "repo_id": "depth-anything/DA3METRIC-LARGE"
+    }
+}

src/trtutils/download/requirements/depth_anything_v1.txt

@@ -0,0 +1,8 @@
+huggingface-hub
+numpy<2
+onnx
+onnxscript
+opencv-python
+pillow
+torch
+torchvision

src/trtutils/download/requirements/depth_anything_v3.txt

@@ -0,0 +1,17 @@
+addict
+einops
+evo
+huggingface-hub
+imageio
+numpy<2
+omegaconf
+onnx
+onnxruntime
+onnxscript
+opencv-python
+pillow
+safetensors
+torch>=2
+torchvision
+trimesh
+xformers