Identity Mappings in Deep Residual Networks

Kaiming HeXiangyu ZhangShaoqing RenJian Sun

   Papers with code   Abstract  PDF

Deep residual networks have emerged as a family of extremely deep architectures showing compelling accuracy and nice convergence behaviors. In this paper, we analyze the propagation formulations behind the residual building blocks, which suggest that the forward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connections and after-addition activation... (read more)

Benchmarked Models

RANK
MODEL
REPO
CODE RESULT
PAPER RESULT
ε-REPRODUCED
BUILD
1
PreResNet-152b
79.9%
--
2
PreResNet-269b
79.1%
--
3
PreResNet-152
79.1%
--
4
PreResNet-101b
79.0%
--
5
PreResNet-200b
78.7%
--
6
PreResNet-101
78.3%
--
7
PreResNet-50
77.6%
--
8
PreResNet-50b
77.5%
--
9
PreResNet-BC-38b
77.1%
--
10
PreResNet-34
75.1%
--
11
PreResNet-BC-26b
74.5%
--
12
PreResNet-26
73.7%
--
13
PreResNet-18
71.6%
--
14
PreResNet-18 x0.75
69.8%
--
15
PreResNet-16
69.5%
--
16
PreResNet=BC-14b
68.7%
--
17
PreResNet-14
67.4%
--
18
PreResNet-12
66.1%
--
19
PreResNet-18 x0.5
66.0%
--
20
PrepResNet-10
64.9%
--
21
PreResNet-18 x0.25
59.9%
--