Dual Path Networks

NeurIPS 2017 Yunpeng ChenJianan LiHuaxin XiaoXiaojie JinShuicheng YanJiashi Feng

In this work, we present a simple, highly efficient and modularized Dual Path Network (DPN) for image classification which presents a new topology of connection paths internally. By revealing the equivalence of the state-of-the-art Residual Network (ResNet) and Densely Convolutional Network (DenseNet) within the HORNN framework, we find that ResNet enables feature re-usage while DenseNet enables new features exploration which are both important for learning good representations... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Image Classification ImageNet DPN-98 (224x224) Top 1 Accuracy 79.95% # 63
Top 5 Accuracy 94.85% # 43
Image Classification ImageNet DPN-131 (224x224) Top 1 Accuracy 80.07% # 61
Top 5 Accuracy 94.88% # 42
Number of params 80M # 17
Image Classification ImageNet DPN-92 (320x320) Top 1 Accuracy 80.66% # 56
Top 5 Accuracy 95.34% # 36
Image Classification ImageNet DPN-92 (320x320, Mean-Max Pooling) Top 1 Accuracy 80.96% # 53
Top 5 Accuracy 95.47% # 34
Image Classification ImageNet DPN-98 (320x320) Top 1 Accuracy 81.06% # 52
Top 5 Accuracy 95.56% # 32
Image Classification ImageNet DPN-98 (320x320, Mean-Max Pooling) Top 1 Accuracy 81.28% # 48
Top 5 Accuracy 95.6% # 31
Image Classification ImageNet DPN-131 (320x320) Top 1 Accuracy 81.38% # 46
Top 5 Accuracy 95.77% # 28
Number of params 80M # 17
Image Classification ImageNet DPN-68 (224x224) Top 1 Accuracy 76.43% # 110
Top 5 Accuracy 93.07% # 77
Image Classification ImageNet DPN-68 (320x320) Top 1 Accuracy 77.85% # 95
Top 5 Accuracy 94.10% # 62
Image Classification ImageNet DPN-68 (320x320, Mean-Max Pooling) Top 1 Accuracy 78.49% # 86
Top 5 Accuracy 94.48% # 54
Image Classification ImageNet DPN-92 (224x224) Top 1 Accuracy 79.27% # 70
Top 5 Accuracy 94.63% # 48

Methods used in the Paper