add metrics and postprocess

deploy/python/postprocess.py

@@ -346,6 +346,58 @@ class PersonAttribute(object):
         return batch_res
 
 
+class FaceAttribute(object):
+    def __init__(self, threshold=0.65, convert_cn=False):
+        self.threshold = threshold
+        self.convert_cn = convert_cn
+
+    def __call__(self, x, file_names=None):
+        attribute_list = [
+            ["CheekWhiskers", "刚长出的双颊胡须"], ["ArchedEyebrows", "柳叶眉"],
+            ["Attractive", "吸引人的"], ["BagsUnderEyes", "眼袋"], ["Bald", "秃头"],
+            ["Bangs", "刘海"], ["BigLips", "大嘴唇"], ["BigNose", "大鼻子"],
+            ["BlackHair", "黑发"], ["BlondHair", "金发"], ["Blurry", "模糊的"],
+            ["BrownHair", "棕发"], ["BushyEyebrows", "浓眉"], ["Chubby", "圆胖的"],
+            ["DoubleChin", "双下巴"], ["Eyeglasses", "带眼镜"], ["Goatee", "山羊胡子"],
+            ["GrayHair", "灰发或白发"], ["HeavyMakeup", "浓妆"],
+            ["HighCheekbones", "高颧骨"], ["Male", "男性"],
+            ["MouthSlightlyOpen", "微微张开嘴巴"], ["Mustache", "胡子"],
+            ["NarrowEyes", "细长的眼睛"], ["NoBeard", "无胡子"],
+            ["OvalFace", "椭圆形的脸"], ["PaleSkin", "苍白的皮肤"],
+            ["PointyNose", "尖鼻子"], ["RecedingHairline", "发际线后移"],
+            ["RosyCheeks", "红润的双颊"], ["Sideburns", "连鬓胡子"], ["Smiling", "微笑"],
+            ["StraightHair", "直发"], ["WavyHair", "卷发"],
+            ["WearingEarrings", "戴着耳环"], ["WearingHat", "戴着帽子"],
+            ["WearingLipstick", "涂了唇膏"], ["WearingNecklace", "戴着项链"],
+            ["WearingNecktie", "戴着领带"], ["Young", "年轻人"]
+        ]
+        gender_list = [["Male", "男性"], ["Female", "女性"]]
+        age_list = [["Young", "年轻人"], ["Old", "老年人"]]
+        batch_res = []
+        if self.convert_cn:
+            index = 1
+        else:
+            index = 0
+        for idx, res in enumerate(x):
+            res = res.tolist()
+            label_res = []
+            threshold_list = [self.threshold] * len(res)
+            pred_res = (np.array(res) > np.array(threshold_list)
+                        ).astype(np.int8).tolist()
+            for i, value in enumerate(pred_res):
+                if i == 20:
+                    label_res.append(gender_list[0][index]
+                                     if value == 1 else gender_list[1][index])
+                elif i == 39:
+                    label_res.append(age_list[0][index]
+                                     if value == 1 else age_list[1][index])
+                else:
+                    if value == 1:
+                        label_res.append(attribute_list[i][index])
+            batch_res.append({"attributes": label_res, "output": pred_res})
+        return batch_res
+
+
 class VehicleAttribute(object):
     def __init__(self, color_threshold=0.5, type_threshold=0.5):
         self.color_threshold = color_threshold

deploy/python/predict_cls.py

@@ -136,7 +136,8 @@ def main(config):
                 if "PersonAttribute" in config[
                         "PostProcess"] or "VehicleAttribute" in config[
                             "PostProcess"] or "TableAttribute" in config[
-                                "PostProcess"]:
+                                "PostProcess"] or "FaceAttribute" in config[
+                                    "PostProcess"]:
                     filename = batch_names[number]
                     print("{}:\t {}".format(filename, result_dict))
                 else:

ppcls/data/postprocess/attr_rec.py

@@ -84,6 +84,7 @@ class PersonAttribute(object):
         if isinstance(x, dict):
             x = x['logits']
         assert isinstance(x, paddle.Tensor)
+
         if file_names is not None:
             assert x.shape[0] == len(file_names)
         x = F.sigmoid(x).numpy()
@@ -98,6 +99,7 @@ class PersonAttribute(object):
             'Skirt&Dress'
         ]
         batch_res = []
+
         for idx, res in enumerate(x):
             res = res.tolist()
             label_res = []
@@ -171,6 +173,66 @@ class PersonAttribute(object):
         return batch_res
 
 
+class FaceAttribute(object):
+    def __init__(self, threshold=0.65, convert_cn=False):
+        self.threshold = threshold
+        self.convert_cn = convert_cn
+
+    def __call__(self, x, file_names=None):
+        if isinstance(x, dict):
+            x = x['logits']
+        assert isinstance(x, paddle.Tensor)
+
+        if file_names is not None:
+            assert x.shape[0] == len(file_names)
+        x = F.sigmoid(x).numpy()
+
+        attribute_list = [
+            ["CheekWhiskers", "刚长出的双颊胡须"], ["ArchedEyebrows", "柳叶眉"],
+            ["Attractive", "吸引人的"], ["BagsUnderEyes", "眼袋"], ["Bald", "秃头"],
+            ["Bangs", "刘海"], ["BigLips", "大嘴唇"], ["BigNose", "大鼻子"],
+            ["BlackHair", "黑发"], ["BlondHair", "金发"], ["Blurry", "模糊的"],
+            ["BrownHair", "棕发"], ["BushyEyebrows", "浓眉"], ["Chubby", "圆胖的"],
+            ["DoubleChin", "双下巴"], ["Eyeglasses", "带眼镜"], ["Goatee", "山羊胡子"],
+            ["GrayHair", "灰发或白发"], ["HeavyMakeup", "浓妆"],
+            ["HighCheekbones", "高颧骨"], ["Male", "男性"],
+            ["MouthSlightlyOpen", "微微张开嘴巴"], ["Mustache", "胡子"],
+            ["NarrowEyes", "细长的眼睛"], ["NoBeard", "无胡子"],
+            ["OvalFace", "椭圆形的脸"], ["PaleSkin", "苍白的皮肤"],
+            ["PointyNose", "尖鼻子"], ["RecedingHairline", "发际线后移"],
+            ["RosyCheeks", "红润的双颊"], ["Sideburns", "连鬓胡子"], ["Smiling", "微笑"],
+            ["StraightHair", "直发"], ["WavyHair", "卷发"],
+            ["WearingEarrings", "戴着耳环"], ["WearingHat", "戴着帽子"],
+            ["WearingLipstick", "涂了唇膏"], ["WearingNecklace", "戴着项链"],
+            ["WearingNecktie", "戴着领带"], ["Young", "年轻人"]
+        ]
+        gender_list = [["Male", "男性"], ["Female", "女性"]]
+        age_list = [["Young", "年轻人"], ["Old", "老年人"]]
+        batch_res = []
+        if self.convert_cn:
+            index = 1
+        else:
+            index = 0
+        for idx, res in enumerate(x):
+            res = res.tolist()
+            label_res = []
+            threshold_list = [self.threshold] * len(res)
+            pred_res = (np.array(res) > np.array(threshold_list)
+                        ).astype(np.int8).tolist()
+            for i, value in enumerate(pred_res):
+                if i == 20:
+                    label_res.append(gender_list[0][index]
+                                     if value == 1 else gender_list[1][index])
+                elif i == 39:
+                    label_res.append(age_list[0][index]
+                                     if value == 1 else age_list[1][index])
+                else:
+                    if value == 1:
+                        label_res.append(attribute_list[i][index])
+            batch_res.append({"attributes": label_res, "output": pred_res})
+        return batch_res
+
+
 class TableAttribute(object):
     def __init__(
             self,

ppcls/metric/__init__.py

@@ -21,7 +21,7 @@ from .metrics import DistillationTopkAcc
 from .metrics import GoogLeNetTopkAcc
 from .metrics import HammingDistance, AccuracyScore
 from .metrics import ATTRMetric
-from .metrics import TprAtFpr
+from .metrics import TprAtFpr, MultilabelMeanAccuracy
 
 
 class CombinedMetrics(AvgMetrics):

ppcls/metric/metrics.py

@@ -218,6 +218,55 @@ class TprAtFpr(nn.Layer):
         return result
 
 
+class MultilabelMeanAccuracy(nn.Layer):
+    def __init__(self, class_num=40):
+        super().__init__()
+        self.gt_all_score_list = []
+        self.gt_label_score_list = []
+        self.max_acc = 0.
+        self.class_num = class_num
+
+    def forward(self, x, label):
+        if isinstance(x, dict):
+            x = x["logits"]
+        x = F.sigmoid(x)
+        label = label[:, 0, :]
+        for i in range(len(x)):
+            self.gt_all_score_list.append(x[i].numpy())
+            self.gt_label_score_list.append(label[i].numpy())
+        return {}
+
+    def reset(self):
+        self.gt_all_score_list = []
+        self.gt_label_score_list = []
+        self.max_acc = 0.
+
+    @property
+    def avg(self):
+        return self.max_acc
+
+    @property
+    def avg_info(self):
+        max_acc = 0.
+        result = ""
+        gt_all_score_list = np.array(self.gt_all_score_list)
+        gt_label_score_list = np.array(self.gt_label_score_list)
+        for i in range(10):
+            threshold = 0.4 + i * 0.05
+            pred_label = (gt_all_score_list > threshold).astype(int)
+            TP = np.sum(
+                (gt_label_score_list == 1) * (pred_label == 1)).astype(float)
+            TN = np.sum(
+                (gt_label_score_list == 0) * (pred_label == 0)).astype(float)
+            acc = (TP + TN) / len(gt_all_score_list)
+            if max_acc <= acc:
+                max_acc = acc
+                result = "threshold: {}, mean_acc: {}".format(
+                    threshold, max_acc / self.class_num)
+        self.max_acc = max_acc / self.class_num
+        return result
+
+
 class Recallk(nn.Layer):
     def __init__(self, topk=(1, 5), descending=True):
         super().__init__()