Summary of the most successful Imagenet prediction standardized architectures


Table below shows a summary of the most successful Imagenet prediction standardized architectures sorted by its performance for the Imagenet classification task. The more successful network is InceptionResNetV2 but at the cost of having more than 55 million of parameters. Xception has a similar accuracy using less than a half of parameters. MobileNetV2, an architecture desinged to be used in mobile devices, with nly 3.4 million of parameters has an accuracy similar to the achived with older architectures as VGG16 and VGG19 that used more than 100 million of parameters.

Model Input size Imagenet Top-1 Accuracy Parameters Depth Residual Publication Source
InceptionResNetV2 (299,299,3) 0,804 55.9M 572 Yes 2016-02 Google
InceptionV4 (299,299,3) 0,802 55.9M   No 2016-02 Google
Xception (299,299,3) 0,790 22.9M 126 Yes 2016-10 Google
InceptionV3 (299,299,3) 0,788 23.8M 159 No 2015-12 Google
DenseNet201 (224,224,3) 0,770 20.2M 201 Dense 2016-08 Facebook
ResNet50 (224,224,3) 0,759 25.6M 168 Yes 2015-12 Microsoft
DenseNet169 (224,224,3) 0,759 14.3M 169 Dense 2016-08 Facebook
MobileNetV2 (1.4) (224,224,3) 0,747 6.9M   Yes 2018-01 Google
DenseNet121 (224,224,3) 0,745 8M 121 Dense 2016-08 Facebook
VGG19 (224,224,3) 0,727 143.7M 26 No 2014-09 Oxford
MobileNetV2 (224,224,3) 0,720 3.4M   Yes 2018-01 Google
VGG16 (224,224,3) 0,715 138.3M 23 No 2014-09 Oxford
MobileNetV1 (224,224,3) 0,706 4.2M   Yes 2017-06 Google
SqueezeNet (227,227,3) 0,575 1.2M   Yes 2016-02 DeepScale
AlexNet (227,227,3) 0,571 62M 8 No 2012-00 BVLC

This table can be used for comparing against the different architectures. It can help in the decision process of choosing between the different pretrained networks for transfer learning tasks.

Back to blog