add evaluation framework and imagenet evaluation (#69)

README.md

@@ -4,6 +4,17 @@ MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applicatio
 
 MobileNetV2: Inverted Residuals and Linear Bottlenecks
 
+Results of accuracy evaluation with [tools/eval](../../tools/eval).
+
+| Models | Top-1 Accuracy | Top-5 Accuracy |
+| ------ | -------------- | -------------- |
+| MobileNet V1 | 67.64 | 87.97 |
+| MobileNet V1 quant | 55.53 | 78.74 |
+| MobileNet V2 | 69.44 | 89.23 |
+| MobileNet V2 quant | 68.37 | 88.56 |
+
+\*: 'quant' stands for 'quantized'.
+
 ## Demo
 
 Run the following command to try the demo:
@@ -24,3 +35,4 @@ All files in this directory are licensed under [Apache 2.0 License](./LICENSE).
 - MobileNet V2: https://arxiv.org/abs/1801.04381
 - MobileNet V1 weight and scripts for training: https://github.com/wjc852456/pytorch-mobilenet-v1
 - MobileNet V2 weight: https://github.com/onnx/models/tree/main/vision/classification/mobilenet
+

mobilenet_v1.py

@@ -2,9 +2,11 @@ import numpy as np
 import cv2 as cv
 
 class MobileNetV1:
-    def __init__(self, modelPath, labelPath, backendId=0, targetId=0):
+    def __init__(self, modelPath, labelPath=None, topK=1, backendId=0, targetId=0):
         self.model_path = modelPath
         self.label_path = labelPath
+        assert topK >= 1
+        self.top_k = topK
         self.backend_id = backendId
         self.target_id = targetId
 
@@ -23,9 +25,10 @@ class MobileNetV1:
 
     def _load_labels(self):
         labels = []
-        with open(self.label_path, 'r') as f:
-            for line in f:
-                labels.append(line.strip())
+        if self.label_path is not None:
+            with open(self.label_path, 'r') as f:
+                for line in f:
+                    labels.append(line.strip())
         return labels
 
     @property
@@ -61,9 +64,18 @@ class MobileNetV1:
         return results
 
     def _postprocess(self, output_blob):
-        predicted_labels = []
+        batched_class_id_list = []
         for o in output_blob:
-            class_id = np.argmax(o)
-            predicted_labels.append(self.labels[class_id])
-        return predicted_labels
+            class_id_list = o.argsort()[::-1][:self.top_k]
+            batched_class_id_list.append(class_id_list)
+        if len(self.labels) > 0:
+            batched_predicted_labels = []
+            for class_id_list in batched_class_id_list:
+                predicted_labels = []
+                for class_id in class_id_list:
+                    predicted_labels.append(self._labels[class_id])
+                batched_predicted_labels.append(predicted_labels)
+            return batched_predicted_labels
+        else:
+            return batched_class_id_list
 

mobilenet_v2.py

@@ -2,9 +2,11 @@ import numpy as np
 import cv2 as cv
 
 class MobileNetV2:
-    def __init__(self, modelPath, labelPath, backendId=0, targetId=0):
+    def __init__(self, modelPath, labelPath=None, topK=1, backendId=0, targetId=0):
         self.model_path = modelPath
         self.label_path = labelPath
+        assert topK >= 1
+        self.top_k = topK
         self.backend_id = backendId
         self.target_id = targetId
 
@@ -23,9 +25,10 @@ class MobileNetV2:
 
     def _load_labels(self):
         labels = []
-        with open(self.label_path, 'r') as f:
-            for line in f:
-                labels.append(line.strip())
+        if self.label_path is not None:
+            with open(self.label_path, 'r') as f:
+                for line in f:
+                    labels.append(line.strip())
         return labels
 
     @property
@@ -61,9 +64,18 @@ class MobileNetV2:
         return results
 
     def _postprocess(self, output_blob):
-        predicted_labels = []
+        batched_class_id_list = []
         for o in output_blob:
-            class_id = np.argmax(o)
-            predicted_labels.append(self.labels[class_id])
-        return predicted_labels
+            class_id_list = o.argsort()[::-1][:self.top_k]
+            batched_class_id_list.append(class_id_list)
+        if len(self.labels) > 0:
+            batched_predicted_labels = []
+            for class_id_list in batched_class_id_list:
+                predicted_labels = []
+                for class_id in class_id_list:
+                    predicted_labels.append(self._labels[class_id])
+                batched_predicted_labels.append(predicted_labels)
+            return batched_predicted_labels
+        else:
+            return batched_class_id_list