Feat/fix face depth

.github/workflows/cpu-ci.yml

@@ -19,17 +19,17 @@ jobs:
     runs-on: ${{ matrix.os }}
 
     steps:
-      - name: Checkout Repository
+      - name: Checkout repository
         uses: actions/checkout@v4
         with:
           lfs: true
 
-      - name: Install Prerequisites for Linux
+      - name: Install prerequisites for Linux
         if: ${{ matrix.os == 'ubuntu-latest' || matrix.os == 'ubuntu-24.04-arm' }}
         run:
           sudo apt update && sudo apt install -y libturbojpeg exiftool ffmpeg libheif-dev poppler-utils
 
-      - name: Install Prerequisites for MacOS
+      - name: Install prerequisites for MacOS
         if: ${{ matrix.os == 'macos-13' || matrix.os == 'macos-latest' }}
         run:
           brew install libjpeg exiftool ffmpeg libheif poppler
@@ -40,18 +40,18 @@ jobs:
           python-version: ${{ matrix.python-version }}
           check-latest: true
 
-      - name: Install Dependencies
+      - name: Install dependencies
         run: |
           python -m pip install -U pip wheel "numpy>=2" "cython>=3.0.12" "setuptools>=69"
           python setup.py build_ext --inplace
           python -m pip install .
 
-      - name: Lint with Pylint
+      - name: Lint with pylint
         run: |
           python -m pip install pylint
           python -m pylint pyface --rcfile=.github/workflows/.pylintrc
 
-      - name: Run Tests with Pytest
+      - name: Run tests with pytest
         run: |
           mkdir -p tests/coverage
           python -m pip install pytest pytest-cov typeguard

.github/workflows/cuda-ci.yml

@@ -7,15 +7,55 @@ on:
       - "doc/**"
       - "**.md"
 
+env:
+  LOCAL_REGISTRY: localhost:5000
+  DOCKERFILE: docker/pr.dockerfile
+
 jobs:
+  get_runner_and_uid:
+    name: Get Runner
+    runs-on: [self-hosted, unicorn]
+    steps:
+      - name: Get UID and GID
+        id: uid_gid
+        run: |
+          echo "uid_gid=$(id -u):$(id -g)" >> "$GITHUB_OUTPUT"
+    outputs:
+      runner: ${{ runner.name }}
+      uid: ${{ steps.uid_gid.outputs.uid_gid }}
+
+  build_docker_image:
+    name: Build Docker Image
+    needs: [get_runner_and_uid]
+    runs-on: ${{ needs.get_runner_and_uid.outputs.runner }}
+
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v4
+
+      - name: Build Docker Image
+        id: docker_build
+        uses: docker/build-push-action@v3
+        with:
+          file: ${{ env.DOCKERFILE }}
+          no-cache: false
+          push: true
+          tags: ${{ env.LOCAL_REGISTRY }}/pyface-container:ci
+
+    outputs:
+      image: ${{ env.LOCAL_REGISTRY }}/pyface-container:ci
+
   ci:
     name: CI
+    needs: [get_runner_and_uid, build_docker_image]
+    runs-on: ${{ needs.get_runner_and_uid.outputs.runner }}
     strategy:
-      fail-fast: false
       matrix:
-        python-version: ["3.10", "3.12"]
-
-    runs-on: ["self-hosted", "unicorn"]
+        python-version:
+          - "3.10"
+    container:
+      image: ${{ needs.build_docker_image.outputs.image }}
+      options: --user ${{ needs.get_runner_and_uid.outputs.uid }} --gpus all
 
     steps:
       - name: Checkout repository

demo/face_depth.py

@@ -3,10 +3,10 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
-def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
+def main(img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg")):
     face_detect = pf.build_face_detection()
     face_depth = pf.build_face_depth()
 
@@ -16,7 +16,7 @@ def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
     results = face_depth([img], [face_box], return_depth=True)
     plotted = face_depth.draw_results(img, [face_box], results)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_depth1.png")
     cb.imwrite(results[0]["depth_img"], out_folder / "face_depth2.png")

demo/face_detection.py

@@ -5,11 +5,11 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     score_th: float = 0.5,
     inp_size: Tuple[int, int] = (480, 640),
 ):
@@ -22,7 +22,7 @@ def main(
     proposals_list = face_detection([img] * 7)
     plotted = face_detection.draw_proposals(img, proposals_list[0])
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_detection.png")
 

demo/face_gender.py

@@ -3,10 +3,10 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
-def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
+def main(img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg")):
     face_detect = pf.build_face_detection()
     face_gender = pf.build_gender_detection()
 
@@ -16,7 +16,7 @@ def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
     results = face_gender([img], [face_box])
     plotted = face_gender.draw_results(img, [face_box], results)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_gender.png")
 

demo/face_landmark.py

@@ -5,11 +5,11 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     score_th: float = 0.5,
     inp_size: Tuple[int, int] = (480, 640),
 ):
@@ -25,7 +25,7 @@ def main(
     result = face_landmark([img], [box])[0]
     plotted = face_landmark.draw_result(img, box, result, plot_details=True)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_landmark.jpg")
 

demo/face_normalization.py

@@ -5,15 +5,15 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     scale: float = 1,
     dst_size: Tuple[int, int] = (224, 224),
 ):
-    face_detection = pf.SCRFD()
+    face_detection = pf.build_face_detection()
     face_normaliation = pf.FaceNormalize(
         dst_size=dst_size,
         interpolation=cb.INTER.BILINEAR,
@@ -22,13 +22,9 @@ def main(
     img = cb.imread(img_path)
     proposals = face_detection([img])[0]
     norm_img = face_normaliation([img], [cb.Keypoints(proposals["lmk5pts"][0])])[0]
-    norm_img = cb.draw_keypoints(
-        norm_img,
-        face_normaliation.destination_pts,
-        scale=1,
-    )
+    norm_img = cb.draw_keypoints(norm_img, face_normaliation.destination_pts, scale=1)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(norm_img, out_folder / f"face_normalize_s={scale}.png")
 

demo/face_recognition.py

@@ -3,12 +3,12 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    src_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
-    tgt_path: str = str(cur_folder / "data" / "face_bank" / "EmmaWatson.jpg"),
+    src_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
+    tgt_path: str = str(RESOURCE_DIR / "face_bank" / "EmmaWatson.jpg"),
 ):
     face_detection = pf.build_face_detection()
     face_recognition = pf.build_face_recognition()
@@ -17,7 +17,7 @@ def main(
     tgt_img = cb.imread(tgt_path)
     lmks = [p["lmk5pts"][0] for p in face_detection([src_img, tgt_img])]
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     results = face_recognition([src_img, tgt_img], lmks)
     src_emb = results[0]["embeddings"]

demo/face_service.py

@@ -3,12 +3,12 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
-    face_bank: str = str(cur_folder / "data" / "face_bank"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
+    face_bank: str = str(RESOURCE_DIR / "face_bank"),
 ):
     face_service = pf.FaceService(
         enable_gender=True,
@@ -19,7 +19,9 @@ def main(
     )
     img = cb.imread(img_path)
     faces_on_img = face_service([img], do_1n=True)[0]
-    cb.imwrite(faces_on_img.gen_info_img(), str(cur_folder / "output" / "face_service.jpg"))
+    out_folder = cb.get_curdir(__file__) / "output"
+    out_folder.mkdir(exist_ok=True, parents=True)
+    cb.imwrite(faces_on_img.gen_info_img(), str(out_folder / "face_service.jpg"))
 
 
 if __name__ == "__main__":

docker/Dockerfile

@@ -1,5 +1,5 @@
 # syntax=docker/dockerfile:experimental
-FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu22.04 as builder
+FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04 as builder
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
     DEBIAN_FRONTEND=noninteractive \

docker/pr.dockerfile

@@ -0,0 +1,17 @@
+# syntax=docker/dockerfile:experimental
+FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04 as builder
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    DEBIAN_FRONTEND=noninteractive \
+    TZ=Asia/Taipei
+
+RUN apt-get update -y && apt-get upgrade -y && \
+    apt-get install -y --no-install-recommends \
+    tzdata wget git git-lfs libturbojpeg exiftool ffmpeg poppler-utils libpng-dev \
+    libtiff5-dev libjpeg8-dev libopenjp2-7-dev zlib1g-dev gcc \
+    libfreetype6-dev liblcms2-dev libwebp-dev tcl8.6-dev tk8.6-dev python3-tk \
+    python3-pip libharfbuzz-dev libfribidi-dev libxcb1-dev libfftw3-dev gosu \
+    libpq-dev python3-dev build-essential && \
+    ln -sf /usr/share/zoneinfo/$TZ /etc/localtime && \
+    dpkg-reconfigure -f noninteractive tzdata && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*

pyface/components/face_depth/tddfav2.py

@@ -7,7 +7,7 @@
 
 from ..enums import FacePose
 from ..utils import append_to_batch, detach_from_batch, download_model_and_return_model_fpath
-from .Sim3DR import sim3dr_cython
+from .Sim3DR import sim3dr_cython  # pylint: disable=E0611
 
 
 def rasterize(
@@ -168,9 +168,7 @@ def _prepare_bfm_npz(
     return u_base, w_shp_base, w_exp_base
 
 
-def _parse_tffda_param(
-    param: np.ndarray,
-) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+def _parse_tffda_param(param: np.ndarray) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
     """
     matrix pose form
     param: shape = (trans_dim + shape_dim + exp_dim) i.e., 62 = 12 + 40 + 10
@@ -261,16 +259,12 @@ def initialize(self) -> None:
         dummy_inputs = {k: np.zeros(input_dims[k], dtype=v["dtype"]) for k, v in self.bfm_engine.input_infos.items()}
         self.bfm_engine(**dummy_inputs)
 
-    def _transform(
-        self,
-        img: np.ndarray,
-        box: cb.Box,
-    ) -> Tuple[np.ndarray, Tuple[float, float], np.ndarray]:
+    def _transform(self, img: np.ndarray, box: cb.Box) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
         inp_w, inp_h = self.engine.input_infos["input"]["shape"][2:]
 
         expanded_box = box.square().scale(fx=0.98, fy=0.98)
+        scale = np.array([expanded_box.width / inp_w, expanded_box.height / inp_h])
         shift = expanded_box.left_top
-        scale = expanded_box.width / inp_w, expanded_box.height / inp_h
         blob = cb.imcropbox(img, expanded_box)
         blob = cb.imresize(blob, (inp_w, inp_h))
         blob = blob.transpose(2, 0, 1)[None]
@@ -280,7 +274,7 @@ def preprocess(
         self,
         imgs: List[np.ndarray],
         boxes: List[cb.Box],
-    ):
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
         if len(imgs) != len(boxes):
             raise ValueError(f"imgs and boxes should have same length, but got {len(imgs)} and {len(boxes)}")
 
@@ -294,11 +288,7 @@ def preprocess(
             shifts.append(shift)
         return blobs, scales, shifts
 
-    def _reconstruct_vertices(
-        self,
-        param: np.ndarray,
-        dense_flag: bool = False,
-    ) -> np.ndarray:
+    def _reconstruct_vertices(self, param: np.ndarray, dense_flag: bool = False) -> np.ndarray:
         def _similar_transform(pts3d, scales):
             x_s, y_s = scales[..., 1], scales[..., 0]
             z_s = (x_s + y_s) / 2
@@ -374,15 +364,22 @@ def __call__(
         blobs, scales, shifts = self.preprocess(imgs, boxes)
         preds = {k: [] for k in self.engine.output_infos.keys()}
         for batch in cb.make_batch(blobs, self.batch_size):
+            current_batch_size = len(batch)
             inputs = {
                 name: np.concatenate(append_to_batch(batch, self.batch_size)).astype(v["dtype"])  # E1123
                 for name, v in self.engine.input_infos.items()
             }
             tmp_preds = self.engine(**inputs)
             for k, v in tmp_preds.items():
-                preds[k].append(detach_from_batch(v, len(batch)))
+                preds[k].append(detach_from_batch(v, current_batch_size))
         preds = {k: np.concatenate(v, 0) for k, v in preds.items()}
-        params = np.concatenate((preds["params"], scales, shifts), -1)
+        scales = np.stack(scales, 0)  # n x 2
+        shifts = np.stack(shifts, 0)  # n x 2
+        # Ensure shapes are compatible for concatenation
+        n = preds["params"].shape[0]
+        if scales.shape[0] != n or shifts.shape[0] != n:
+            raise ValueError(f"Shape mismatch: preds['params'] has {n} rows, scales has {scales.shape[0]}, shifts has {shifts.shape[0]}")
+        params = np.concatenate((preds["params"], scales, shifts), axis=-1)
         lmk3d68pts = self._gen_3d_landmarks(params)
         pose_degrees = self._get_pose_degrees(params)