some updates to work with 2023 py/numpy/torch versions

ptgaze/common/camera.py

@@ -31,9 +31,9 @@ def project_points(self,
                        tvec: Optional[np.ndarray] = None) -> np.ndarray:
         assert points3d.shape[1] == 3
         if rvec is None:
-            rvec = np.zeros(3, dtype=np.float)
+            rvec = np.zeros(3, dtype=float)
         if tvec is None:
-            tvec = np.zeros(3, dtype=np.float)
+            tvec = np.zeros(3, dtype=float)
         points2d, _ = cv2.projectPoints(points3d, rvec, tvec,
                                         self.camera_matrix,
                                         self.dist_coefficients)

ptgaze/common/face_model.py

@@ -1,5 +1,3 @@
-import dataclasses
-
 import cv2
 import numpy as np
 from scipy.spatial.transform import Rotation
@@ -8,7 +6,6 @@
 from .face import Face
 
 
-@dataclasses.dataclass(frozen=True)
 class FaceModel:
     LANDMARKS: np.ndarray
     REYE_INDICES: np.ndarray
@@ -26,8 +23,8 @@ def estimate_head_pose(self, face: Face, camera: Camera) -> None:
         # The default values of rvec and tvec below mean that the
         # initial estimate of the head pose is not rotated and the
         # face is in front of the camera.
-        rvec = np.zeros(3, dtype=np.float)
-        tvec = np.array([0, 0, 1], dtype=np.float)
+        rvec = np.zeros(3, dtype=float)
+        tvec = np.array([0, 0, 1], dtype=float)
         _, rvec, tvec = cv2.solvePnP(self.LANDMARKS,
                                      face.landmarks,
                                      camera.camera_matrix,

ptgaze/common/face_model_68.py

@@ -1,11 +1,8 @@
-import dataclasses
-
 import numpy as np
 
 from .face_model import FaceModel
 
 
-@dataclasses.dataclass(frozen=True)
 class FaceModel68(FaceModel):
     """3D face model for Multi-PIE 68 points mark-up.
 

ptgaze/common/face_model_mediapipe.py

@@ -1,11 +1,8 @@
-import dataclasses
-
 import numpy as np
 
 from .face_model import FaceModel
 
 
-@dataclasses.dataclass(frozen=True)
 class FaceModelMediaPipe(FaceModel):
     """3D face model for MediaPipe 468 points mark-up.
 

ptgaze/common/visualizer.py

@@ -25,12 +25,12 @@ def draw_bbox(self,
                   lw: int = 1) -> None:
         assert self.image is not None
         assert bbox.shape == (2, 2)
-        bbox = np.round(bbox).astype(np.int).tolist()
+        bbox = np.round(bbox).astype(np.int64).tolist()
         cv2.rectangle(self.image, tuple(bbox[0]), tuple(bbox[1]), color, lw)
 
     @staticmethod
     def _convert_pt(point: np.ndarray) -> Tuple[int, int]:
-        return tuple(np.round(point).astype(np.int).tolist())
+        return tuple(np.round(point).astype(np.int64).tolist())
 
     def draw_points(self,
                     points: np.ndarray,
@@ -71,7 +71,7 @@ def draw_model_axes(self, face: Face, length: float, lw: int = 2) -> None:
         assert face.head_position is not None
         assert face.landmarks is not None
         # Get the axes of the model coordinate system
-        axes3d = np.eye(3, dtype=np.float) @ Rotation.from_euler(
+        axes3d = np.eye(3, dtype=float) @ Rotation.from_euler(
             'XYZ', [0, np.pi, 0]).as_matrix()
         axes3d = axes3d * length
         axes2d = self._camera.project_points(axes3d,

ptgaze/head_pose_estimation/face_landmark_estimator.py

@@ -20,15 +20,15 @@ def __init__(self, config: DictConfig):
         elif self.mode == 'face_alignment_dlib':
             self.detector = dlib.get_frontal_face_detector()
             self.predictor = face_alignment.FaceAlignment(
-                face_alignment.LandmarksType._2D,
+                face_alignment.LandmarksType.TWO_D,
                 face_detector='dlib',
                 flip_input=False,
                 device=config.device)
         elif self.mode == 'face_alignment_sfd':
             self.detector = face_alignment.detection.sfd.sfd_detector.SFDDetector(
                 device=config.device)
             self.predictor = face_alignment.FaceAlignment(
-                face_alignment.LandmarksType._2D,
+                face_alignment.LandmarksType.TWO_D,
                 flip_input=False,
                 device=config.device)
         elif self.mode == 'mediapipe':
@@ -57,10 +57,10 @@ def _detect_faces_dlib(self, image: np.ndarray) -> List[Face]:
         for bbox in bboxes:
             predictions = self.predictor(image[:, :, ::-1], bbox)
             landmarks = np.array([(pt.x, pt.y) for pt in predictions.parts()],
-                                 dtype=np.float)
+                                 dtype=float)
             bbox = np.array([[bbox.left(), bbox.top()],
                              [bbox.right(), bbox.bottom()]],
-                            dtype=np.float)
+                            dtype=float)
             detected.append(Face(bbox, landmarks))
         return detected
 
@@ -77,7 +77,7 @@ def _detect_faces_face_alignment_dlib(self,
             predictions = []
         detected = []
         for bbox, landmarks in zip(bboxes, predictions):
-            bbox = np.array(bbox, dtype=np.float).reshape(2, 2)
+            bbox = np.array(bbox, dtype=float).reshape(2, 2)
             detected.append(Face(bbox, landmarks))
         return detected
 
@@ -91,7 +91,7 @@ def _detect_faces_face_alignment_sfd(self,
             predictions = []
         detected = []
         for bbox, landmarks in zip(bboxes, predictions):
-            bbox = np.array(bbox, dtype=np.float).reshape(2, 2)
+            bbox = np.array(bbox, dtype=float).reshape(2, 2)
             detected.append(Face(bbox, landmarks))
         return detected
 

ptgaze/head_pose_estimation/head_pose_normalizer.py

@@ -65,4 +65,4 @@ def _get_scale_matrix(self, distance: float) -> np.ndarray:
             [0, 1, 0],
             [0, 0, self.normalized_distance / distance],
         ],
-                        dtype=np.float)
+                        dtype=float)

ptgaze/models/mpiifacegaze/backbones/resnet_simple.py

@@ -20,8 +20,10 @@ def __init__(self, config: DictConfig):
 
         pretrained_name = config.model.backbone.pretrained
         if pretrained_name:
-            state_dict = torch.hub.load_state_dict_from_url(
-                torchvision.models.resnet.model_urls[pretrained_name])
+            if pretrained_name != 'resnet18':
+                raise NotImplementedError()
+            url = torchvision.models.resnet.ResNet18_Weights.IMAGENET1K_V1.url
+            state_dict = torch.hub.load_state_dict_from_url(url)
             self.load_state_dict(state_dict, strict=False)
             # While the pretrained models of torchvision are trained
             # using images with RGB channel order, in this repository