fix mobilenet mean and measure top 1 error

chainer · Dec 6, 2019 · 012cbc8 · 012cbc8
1 parent 85ee099
commit 012cbc8
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 6 deletions.
diff --git a/chainercv/links/model/mobilenet/mobilenet_v2.py b/chainercv/links/model/mobilenet/mobilenet_v2.py
@@ -33,7 +33,7 @@ def _depth_multiplied_output_channels(base_out_channels,
                            min_depth)
 
 
-_zero_mean = np.asarray([0] * 3, dtype=np.float)[:, np.newaxis, np.newaxis]
+_mean = np.asarray([128] * 3, dtype=np.float)[:, np.newaxis, np.newaxis]
 _tf_mobilenetv2_scale = np.asarray(
     [1 / 128.0] * 3, dtype=np.float)[:, np.newaxis, np.newaxis]
 
@@ -96,8 +96,7 @@ class MobileNetV2(PickableSequentialChain):
             the default values are used.
             If a supported pretrained model is used,
             the mean value used to train the pretrained model is used.
-            Otherwise, the mean value calculated from ILSVRC 2012 dataset
-            is used.
+            Otherwise, the mean value used by TF's implementation is used.
         initialW (callable): Initializer for the weights.
         initial_bias (callable): Initializer for the biases.
 
@@ -115,7 +114,7 @@ class MobileNetV2(PickableSequentialChain):
             'param': {
                 'n_class':
                 1001,  # first element is background
-                'mean': _zero_mean,
+                'mean': _mean,
                 'scale': _tf_mobilenetv2_scale,
             },
             'overwritable': None,
@@ -144,7 +143,7 @@ def __init__(self,
                 'scale': scale
             }, pretrained_model, self._models[arch], {
                 'n_class': 1001,
-                'mean': _zero_mean,
+                'mean': _mean,
                 'scale': _tf_mobilenetv2_scale
             })
         self.mean = param['mean']

diff --git a/examples/classification/README.md b/examples/classification/README.md
@@ -8,6 +8,7 @@ Single crop error rates of the models with the weights converted from Caffe weig
 
 | Model | Top 1 | Original Top 1 |
 |:-:|:-:|:-:|
+| MobileNetV2 | 29.9 % | 28.0 % [6] |
 | VGG16 | 29.0 % | 28.5 % [1] |
 | ResNet50 (`arch=he`) | 24.8 % | 24.7 % [2] |
 | ResNet101 (`arch=he`) | 23.6 % | 23.6 % [2] |
@@ -22,6 +23,7 @@ Ten crop error rate.
 
 | Model | Top 1 | Original Top 1 |
 |:-:|:-:|:-:|
+| MobileNetV2 | 27.4 % |  |
 | VGG16 | 27.1 % |   |
 | ResNet50 (`arch=he`) | 23.0 % | 22.9 % [2] |
 | ResNet101 (`arch=he`) | 21.8 % | 21.8 % [2] |
@@ -37,7 +39,7 @@ The results can be reproduced by the following command.
 These scores are obtained using OpenCV backend. If Pillow is used, scores would differ.
 
 ```
-$ python eval_imagenet.py <path_to_val_dataset> [--model vgg16|resnet50|resnet101|resnet152|se-resnet50|se-resnet101|se-resnet152] [--pretrained-model <model_path>] [--batchsize <batchsize>] [--gpu <gpu>] [--crop center|10]
+$ python eval_imagenet.py <path_to_val_dataset> [--model mobilenetv2|vgg16|resnet50|resnet101|resnet152|se-resnet50|se-resnet101|se-resnet152] [--pretrained-model <model_path>] [--batchsize <batchsize>] [--gpu <gpu>] [--crop center|10]
 ```
 
 ### Trained model
@@ -104,3 +106,4 @@ The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) dataset has 1000
 3. Jie Hu, Li Shen, Gang Sun. "Squeeze-and-Excitation Networks" CVPR 2018
 4. https://github.com/hujie-frank/SENet
 5. Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, Kaiming He. "Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour" https://arxiv.org/abs/1706.02677
+6. Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen. "MobileNetV2: Inverted Residuals and Linear Bottlenecks" https://arxiv.org/abs/1801.04381