Added capabilities (Capability) to the estimators.

src/facebias/estimators/__init__.py

@@ -5,13 +5,33 @@ from pathlib import Path
 from typing import Any
 
 import numpy as np
+from enum import StrEnum
+
+
+class Capability(StrEnum):
+    AGE = "age"
+    AGEGROUP = "age_group"
+    SEX = "sex"
+    SKINCOLOR = "skin_color"
+    ETHNICITY = "ethinicity"
 
 
 class BaseEstimator(ABC):
     @abstractmethod
-    def predict(self, images: dict[Path, np.ndarray]) -> dict[Path, dict[str, Any]]:
+    def predict(self, images: dict[Path, np.ndarray]) -> dict[Path, dict[Capability, Any]]:
+        """Runs the estimator on a batch of images.
+
+        The input `images` is a dictionary indexed by the image name, or ID,
+        and values are image data in the appropriate format for the estimator
+        (RGB, BGR' Lab, ...).
+        """
         pass
 
     @abstractclassmethod
-    def capabilities() -> list[str]:
+    def capabilities() -> list[Capability]:
+        """Returns the estimator capabilities
+
+        In this context, estimator capabilities mean what is infers from the
+        face image, such as `age_group`, `sex`, `skin_color`, etc.
+        """
         pass

src/facebias/estimators/fairface.py

@@ -1,6 +1,5 @@
 # -*- coding: utf-8 -*-
 
-import logging
 from collections import OrderedDict
 from pathlib import Path
 from typing import Any, Optional
@@ -9,12 +8,10 @@ import numpy as np
 import torch
 import torch.nn as nn
 import torchvision
-from facebias import load_dataset
+from facebias.estimators import BaseEstimator, Capability
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
 from torchvision import transforms
 
-from facebias.estimators import BaseEstimator
-
 
 class FairFace(BaseEstimator):
     """
@@ -41,7 +38,7 @@ class FairFace(BaseEstimator):
         )
         return self
 
-    def predict(self, ims: dict[Path, np.ndarray]) -> dict[Path, dict[str, Any]]:
+    def predict(self, ims: dict[Path, np.ndarray]) -> dict[Path, dict[Capability, Any]]:
         preds = OrderedDict()
         for p, im in ims.items():
             im = self.T(im).view(1, 3, 224, 224).to(self.device)
@@ -51,9 +48,9 @@ class FairFace(BaseEstimator):
             y = y.cpu().detach().squeeze().numpy()
 
             # Ethnicity prediction
-            # y_ethno = y[:4]
+            y_ethno = y[:4]
-            # ethno_score = np.exp(y_ethno) / np.sum(np.exp(y_ethno))
+            ethno_score = np.exp(y_ethno) / np.sum(np.exp(y_ethno))
-            # ethno_pred = np.argmax(ethno_score)
+            ethno_pred = np.argmax(ethno_score)
 
             # Age prediction
             y_age = y[9:18]
@@ -67,26 +64,37 @@ class FairFace(BaseEstimator):
             sex_label = "m" if not sex_pred else "f"
 
             preds[p] = {
-                "age_group": _to_age_label(age_pred),
+                Capability.AGEGROUP: _to_age_label(age_pred),
-                "sex": sex_label,
+                Capability.SEX: sex_label,
+                # Capability.ETHNICITY: _to_ethno_label(ethno_pred)
             }
 
         return preds
 
-    def capabilities() -> list[str]:
+    def capabilities() -> list[Capability]:
-        return ["age_group", "sex"]
+        return [Capability.AGEGROUP, Capability.SEX] #, Capability.ETHNICITY]
 
 
 def _to_age_label(age):
     return "{}-{}".format(age * 10, age * 10 + 9)
 
 
+def _to_ethno_label(val):
+    if val not in range(4):
+        return "unknown"
+    opts = ["white", "black", "asian", "indian"]
+    return opts[val]
+
+
 if __name__ == '__main__':
-    MODEL_PATH = Path("../../../models/fairface_alldata_4race_20191111.pt")
+    from facebias import load_dataset
-    model = FairFace()
-    model.load_state_dict(MODEL_PATH)
 
     DATASET_PATH = Path("../../../data/frll_neutral_front")
     dataset, _ = load_dataset(DATASET_PATH, imname_proc_fn=lambda x: x.split('_')[0])
 
+    print(FairFace.capabilities())
+    MODEL_PATH = Path("../../../models/fairface_alldata_4race_20191111.pt")
+    model = FairFace()
+    model.load_state_dict(MODEL_PATH)
+
     out = model.predict(dataset)