Best performing ImageNet model from Wide Residual Networks BMVC 2016 paper https://arxiv.org/abs/1605.07146
The model is slower than ResNet-101 and faster than ResNet-152, with better accuracy:
| Model | top-1 err, % | top-5 err, % | #params | time/batch 16 |
|---|---|---|---|---|
| ResNet-50 | 24.01 | 7.02 | 25.6M | 49 |
| ResNet-101 | 22.44 | 6.21 | 44.5M | 82 |
| ResNet-152 | 22.16 | 6.16 | 60.2M | 115 |
| WRN-50-2-bottleneck | 21.9 | 6.03 | 68.9M | 93 |
| pre-ResNet-200 | 21.66 | 5.79 | 64.7M | 154 |
Download (263MB): https://yadi.sk/d/-8AWymOPyVZns
PyTorch and Tensorflow pretrained weights and model definitions:
https://github.com/szagoruyko/functional-zoo/blob/master/wide-resnet-50-2-export.ipynb
Convergence plot:
If you find this model useful please cite this paper:
@INPROCEEDINGS{Zagoruyko2016WRN,
author = {Sergey Zagoruyko and Nikos Komodakis},
title = {Wide Residual Networks},
booktitle = {BMVC},
year = {2016},
}nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> output]
(1): cudnn.SpatialConvolution(3 -> 64, 7x7, 2,2, 3,3) without bias
(2): nn.SpatialBatchNormalization (4D) (64)
(3): cudnn.ReLU
(4): nn.SpatialMaxPooling(3x3, 2,2, 1,1)
(5): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(64 -> 128, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (128)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(128 -> 128, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (128)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(128 -> 256, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (256)
| }
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> output]
(1): cudnn.SpatialConvolution(64 -> 256, 1x1) without bias
(2): nn.SpatialBatchNormalization (4D) (256)
}
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(256 -> 128, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (128)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(128 -> 128, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (128)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(128 -> 256, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (256)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(256 -> 128, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (128)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(128 -> 128, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (128)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(128 -> 256, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (256)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
}
(6): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> output]
(1): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(256 -> 256, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (256)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(256 -> 256, 3x3, 2,2, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (256)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(256 -> 512, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (512)
| }
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> output]
(1): cudnn.SpatialConvolution(256 -> 512, 1x1, 2,2) without bias
(2): nn.SpatialBatchNormalization (4D) (512)
}
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(512 -> 256, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (256)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(256 -> 256, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (256)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(256 -> 512, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (512)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(512 -> 256, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (256)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(256 -> 256, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (256)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(256 -> 512, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (512)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(4): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(512 -> 256, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (256)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(256 -> 256, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (256)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(256 -> 512, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (512)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
}
(7): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
(1): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(512 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 2,2, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> output]
(1): cudnn.SpatialConvolution(512 -> 1024, 1x1, 2,2) without bias
(2): nn.SpatialBatchNormalization (4D) (1024)
}
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(4): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(5): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(6): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 512, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (512)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(512 -> 512, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (512)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(512 -> 1024, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (1024)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
}
(8): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(1024 -> 1024, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (1024)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(1024 -> 1024, 3x3, 2,2, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (1024)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(1024 -> 2048, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (2048)
| }
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> output]
(1): cudnn.SpatialConvolution(1024 -> 2048, 1x1, 2,2) without bias
(2): nn.SpatialBatchNormalization (4D) (2048)
}
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(2048 -> 1024, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (1024)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(1024 -> 1024, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (1024)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(1024 -> 2048, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (2048)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ConcatTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): cudnn.SpatialConvolution(2048 -> 1024, 1x1) without bias
| (2): nn.SpatialBatchNormalization (4D) (1024)
| (3): cudnn.ReLU
| (4): cudnn.SpatialConvolution(1024 -> 1024, 3x3, 1,1, 1,1) without bias
| (5): nn.SpatialBatchNormalization (4D) (1024)
| (6): cudnn.ReLU
| (7): cudnn.SpatialConvolution(1024 -> 2048, 1x1) without bias
| (8): nn.SpatialBatchNormalization (4D) (2048)
| }
`-> (2): nn.Identity
... -> output
}
(2): nn.CAddTable
(3): cudnn.ReLU
}
}
(9): cudnn.SpatialAveragePooling(7x7, 1,1)
(10): nn.View(2048)
(11): nn.Linear(2048 -> 1000)
}