├── models ├── __init__.py ├── lenet.py ├── vgg.py ├── mobilenet.py ├── mobilenetv2.py ├── googlenet.py ├── resnext.py ├── dpn.py ├── densenet.py ├── shufflenet.py ├── senet.py ├── preact_resnet.py ├── dla_simple.py ├── pnasnet.py ├── resnet.py ├── dla.py ├── regnet.py ├── shufflenetv2.py └── efficientnet.py ├── LICENSE ├── README.md ├── utils.py └── main.py /models/__init__.py: -------------------------------------------------------------------------------- 1 | from .vgg import * 2 | from .dpn import * 3 | from .lenet import * 4 | from .senet import * 5 | from .pnasnet import * 6 | from .densenet import * 7 | from .googlenet import * 8 | from .shufflenet import * 9 | from .shufflenetv2 import * 10 | from .resnet import * 11 | from .resnext import * 12 | from .preact_resnet import * 13 | from .mobilenet import * 14 | from .mobilenetv2 import * 15 | from .efficientnet import * 16 | from .regnet import * 17 | from .dla_simple import * 18 | from .dla import * 19 | -------------------------------------------------------------------------------- /models/lenet.py: -------------------------------------------------------------------------------- 1 | '''LeNet in PyTorch.''' 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | class LeNet(nn.Module): 6 | def __init__(self): 7 | super(LeNet, self).__init__() 8 | self.conv1 = nn.Conv2d(3, 6, 5) 9 | self.conv2 = nn.Conv2d(6, 16, 5) 10 | self.fc1 = nn.Linear(16*5*5, 120) 11 | self.fc2 = nn.Linear(120, 84) 12 | self.fc3 = nn.Linear(84, 10) 13 | 14 | def forward(self, x): 15 | out = F.relu(self.conv1(x)) 16 | out = F.max_pool2d(out, 2) 17 | out = F.relu(self.conv2(out)) 18 | out = F.max_pool2d(out, 2) 19 | out = out.view(out.size(0), -1) 20 | out = F.relu(self.fc1(out)) 21 | out = F.relu(self.fc2(out)) 22 | out = self.fc3(out) 23 | return out 24 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2017 liukuang 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Train CIFAR10 with PyTorch 2 | 3 | I'm playing with [PyTorch](http://pytorch.org/) on the CIFAR10 dataset. 4 | 5 | ## Prerequisites 6 | - Python 3.6+ 7 | - PyTorch 1.0+ 8 | 9 | ## Training 10 | ``` 11 | # Start training with: 12 | python main.py 13 | 14 | # You can manually resume the training with: 15 | python main.py --resume --lr=0.01 16 | ``` 17 | 18 | ## Accuracy 19 | | Model | Acc. | 20 | | ----------------- | ----------- | 21 | | [VGG16](https://arxiv.org/abs/1409.1556) | 92.64% | 22 | | [ResNet18](https://arxiv.org/abs/1512.03385) | 93.02% | 23 | | [ResNet50](https://arxiv.org/abs/1512.03385) | 93.62% | 24 | | [ResNet101](https://arxiv.org/abs/1512.03385) | 93.75% | 25 | | [RegNetX_200MF](https://arxiv.org/abs/2003.13678) | 94.24% | 26 | | [RegNetY_400MF](https://arxiv.org/abs/2003.13678) | 94.29% | 27 | | [MobileNetV2](https://arxiv.org/abs/1801.04381) | 94.43% | 28 | | [ResNeXt29(32x4d)](https://arxiv.org/abs/1611.05431) | 94.73% | 29 | | [ResNeXt29(2x64d)](https://arxiv.org/abs/1611.05431) | 94.82% | 30 | | [SimpleDLA](https://arxiv.org/abs/1707.064) | 94.89% | 31 | | [DenseNet121](https://arxiv.org/abs/1608.06993) | 95.04% | 32 | | [PreActResNet18](https://arxiv.org/abs/1603.05027) | 95.11% | 33 | | [DPN92](https://arxiv.org/abs/1707.01629) | 95.16% | 34 | | [DLA](https://arxiv.org/pdf/1707.06484.pdf) | 95.47% | 35 | 36 | -------------------------------------------------------------------------------- /models/vgg.py: -------------------------------------------------------------------------------- 1 | '''VGG11/13/16/19 in Pytorch.''' 2 | import torch 3 | import torch.nn as nn 4 | 5 | 6 | cfg = { 7 | 'VGG11': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 8 | 'VGG13': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 9 | 'VGG16': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 10 | 'VGG19': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], 11 | } 12 | 13 | 14 | class VGG(nn.Module): 15 | def __init__(self, vgg_name): 16 | super(VGG, self).__init__() 17 | self.features = self._make_layers(cfg[vgg_name]) 18 | self.classifier = nn.Linear(512, 10) 19 | 20 | def forward(self, x): 21 | out = self.features(x) 22 | out = out.view(out.size(0), -1) 23 | out = self.classifier(out) 24 | return out 25 | 26 | def _make_layers(self, cfg): 27 | layers = [] 28 | in_channels = 3 29 | for x in cfg: 30 | if x == 'M': 31 | layers += [nn.MaxPool2d(kernel_size=2, stride=2)] 32 | else: 33 | layers += [nn.Conv2d(in_channels, x, kernel_size=3, padding=1), 34 | nn.BatchNorm2d(x), 35 | nn.ReLU(inplace=True)] 36 | in_channels = x 37 | layers += [nn.AvgPool2d(kernel_size=1, stride=1)] 38 | return nn.Sequential(*layers) 39 | 40 | 41 | def test(): 42 | net = VGG('VGG11') 43 | x = torch.randn(2,3,32,32) 44 | y = net(x) 45 | print(y.size()) 46 | 47 | # test() 48 | -------------------------------------------------------------------------------- /models/mobilenet.py: -------------------------------------------------------------------------------- 1 | '''MobileNet in PyTorch. 2 | 3 | See the paper "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" 4 | for more details. 5 | ''' 6 | import torch 7 | import torch.nn as nn 8 | import torch.nn.functional as F 9 | 10 | 11 | class Block(nn.Module): 12 | '''Depthwise conv + Pointwise conv''' 13 | def __init__(self, in_planes, out_planes, stride=1): 14 | super(Block, self).__init__() 15 | self.conv1 = nn.Conv2d(in_planes, in_planes, kernel_size=3, stride=stride, padding=1, groups=in_planes, bias=False) 16 | self.bn1 = nn.BatchNorm2d(in_planes) 17 | self.conv2 = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) 18 | self.bn2 = nn.BatchNorm2d(out_planes) 19 | 20 | def forward(self, x): 21 | out = F.relu(self.bn1(self.conv1(x))) 22 | out = F.relu(self.bn2(self.conv2(out))) 23 | return out 24 | 25 | 26 | class MobileNet(nn.Module): 27 | # (128,2) means conv planes=128, conv stride=2, by default conv stride=1 28 | cfg = [64, (128,2), 128, (256,2), 256, (512,2), 512, 512, 512, 512, 512, (1024,2), 1024] 29 | 30 | def __init__(self, num_classes=10): 31 | super(MobileNet, self).__init__() 32 | self.conv1 = nn.Conv2d(3, 32, kernel_size=3, stride=1, padding=1, bias=False) 33 | self.bn1 = nn.BatchNorm2d(32) 34 | self.layers = self._make_layers(in_planes=32) 35 | self.linear = nn.Linear(1024, num_classes) 36 | 37 | def _make_layers(self, in_planes): 38 | layers = [] 39 | for x in self.cfg: 40 | out_planes = x if isinstance(x, int) else x[0] 41 | stride = 1 if isinstance(x, int) else x[1] 42 | layers.append(Block(in_planes, out_planes, stride)) 43 | in_planes = out_planes 44 | return nn.Sequential(*layers) 45 | 46 | def forward(self, x): 47 | out = F.relu(self.bn1(self.conv1(x))) 48 | out = self.layers(out) 49 | out = F.avg_pool2d(out, 2) 50 | out = out.view(out.size(0), -1) 51 | out = self.linear(out) 52 | return out 53 | 54 | 55 | def test(): 56 | net = MobileNet() 57 | x = torch.randn(1,3,32,32) 58 | y = net(x) 59 | print(y.size()) 60 | 61 | # test() 62 | -------------------------------------------------------------------------------- /models/mobilenetv2.py: -------------------------------------------------------------------------------- 1 | '''MobileNetV2 in PyTorch. 2 | 3 | See the paper "Inverted Residuals and Linear Bottlenecks: 4 | Mobile Networks for Classification, Detection and Segmentation" for more details. 5 | ''' 6 | import torch 7 | import torch.nn as nn 8 | import torch.nn.functional as F 9 | 10 | 11 | class Block(nn.Module): 12 | '''expand + depthwise + pointwise''' 13 | def __init__(self, in_planes, out_planes, expansion, stride): 14 | super(Block, self).__init__() 15 | self.stride = stride 16 | 17 | planes = expansion * in_planes 18 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, stride=1, padding=0, bias=False) 19 | self.bn1 = nn.BatchNorm2d(planes) 20 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, groups=planes, bias=False) 21 | self.bn2 = nn.BatchNorm2d(planes) 22 | self.conv3 = nn.Conv2d(planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) 23 | self.bn3 = nn.BatchNorm2d(out_planes) 24 | 25 | self.shortcut = nn.Sequential() 26 | if stride == 1 and in_planes != out_planes: 27 | self.shortcut = nn.Sequential( 28 | nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False), 29 | nn.BatchNorm2d(out_planes), 30 | ) 31 | 32 | def forward(self, x): 33 | out = F.relu(self.bn1(self.conv1(x))) 34 | out = F.relu(self.bn2(self.conv2(out))) 35 | out = self.bn3(self.conv3(out)) 36 | out = out + self.shortcut(x) if self.stride==1 else out 37 | return out 38 | 39 | 40 | class MobileNetV2(nn.Module): 41 | # (expansion, out_planes, num_blocks, stride) 42 | cfg = [(1, 16, 1, 1), 43 | (6, 24, 2, 1), # NOTE: change stride 2 -> 1 for CIFAR10 44 | (6, 32, 3, 2), 45 | (6, 64, 4, 2), 46 | (6, 96, 3, 1), 47 | (6, 160, 3, 2), 48 | (6, 320, 1, 1)] 49 | 50 | def __init__(self, num_classes=10): 51 | super(MobileNetV2, self).__init__() 52 | # NOTE: change conv1 stride 2 -> 1 for CIFAR10 53 | self.conv1 = nn.Conv2d(3, 32, kernel_size=3, stride=1, padding=1, bias=False) 54 | self.bn1 = nn.BatchNorm2d(32) 55 | self.layers = self._make_layers(in_planes=32) 56 | self.conv2 = nn.Conv2d(320, 1280, kernel_size=1, stride=1, padding=0, bias=False) 57 | self.bn2 = nn.BatchNorm2d(1280) 58 | self.linear = nn.Linear(1280, num_classes) 59 | 60 | def _make_layers(self, in_planes): 61 | layers = [] 62 | for expansion, out_planes, num_blocks, stride in self.cfg: 63 | strides = [stride] + [1]*(num_blocks-1) 64 | for stride in strides: 65 | layers.append(Block(in_planes, out_planes, expansion, stride)) 66 | in_planes = out_planes 67 | return nn.Sequential(*layers) 68 | 69 | def forward(self, x): 70 | out = F.relu(self.bn1(self.conv1(x))) 71 | out = self.layers(out) 72 | out = F.relu(self.bn2(self.conv2(out))) 73 | # NOTE: change pooling kernel_size 7 -> 4 for CIFAR10 74 | out = F.avg_pool2d(out, 4) 75 | out = out.view(out.size(0), -1) 76 | out = self.linear(out) 77 | return out 78 | 79 | 80 | def test(): 81 | net = MobileNetV2() 82 | x = torch.randn(2,3,32,32) 83 | y = net(x) 84 | print(y.size()) 85 | 86 | # test() 87 | -------------------------------------------------------------------------------- /models/googlenet.py: -------------------------------------------------------------------------------- 1 | '''GoogLeNet with PyTorch.''' 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | 6 | 7 | class Inception(nn.Module): 8 | def __init__(self, in_planes, n1x1, n3x3red, n3x3, n5x5red, n5x5, pool_planes): 9 | super(Inception, self).__init__() 10 | # 1x1 conv branch 11 | self.b1 = nn.Sequential( 12 | nn.Conv2d(in_planes, n1x1, kernel_size=1), 13 | nn.BatchNorm2d(n1x1), 14 | nn.ReLU(True), 15 | ) 16 | 17 | # 1x1 conv -> 3x3 conv branch 18 | self.b2 = nn.Sequential( 19 | nn.Conv2d(in_planes, n3x3red, kernel_size=1), 20 | nn.BatchNorm2d(n3x3red), 21 | nn.ReLU(True), 22 | nn.Conv2d(n3x3red, n3x3, kernel_size=3, padding=1), 23 | nn.BatchNorm2d(n3x3), 24 | nn.ReLU(True), 25 | ) 26 | 27 | # 1x1 conv -> 5x5 conv branch 28 | self.b3 = nn.Sequential( 29 | nn.Conv2d(in_planes, n5x5red, kernel_size=1), 30 | nn.BatchNorm2d(n5x5red), 31 | nn.ReLU(True), 32 | nn.Conv2d(n5x5red, n5x5, kernel_size=3, padding=1), 33 | nn.BatchNorm2d(n5x5), 34 | nn.ReLU(True), 35 | nn.Conv2d(n5x5, n5x5, kernel_size=3, padding=1), 36 | nn.BatchNorm2d(n5x5), 37 | nn.ReLU(True), 38 | ) 39 | 40 | # 3x3 pool -> 1x1 conv branch 41 | self.b4 = nn.Sequential( 42 | nn.MaxPool2d(3, stride=1, padding=1), 43 | nn.Conv2d(in_planes, pool_planes, kernel_size=1), 44 | nn.BatchNorm2d(pool_planes), 45 | nn.ReLU(True), 46 | ) 47 | 48 | def forward(self, x): 49 | y1 = self.b1(x) 50 | y2 = self.b2(x) 51 | y3 = self.b3(x) 52 | y4 = self.b4(x) 53 | return torch.cat([y1,y2,y3,y4], 1) 54 | 55 | 56 | class GoogLeNet(nn.Module): 57 | def __init__(self): 58 | super(GoogLeNet, self).__init__() 59 | self.pre_layers = nn.Sequential( 60 | nn.Conv2d(3, 192, kernel_size=3, padding=1), 61 | nn.BatchNorm2d(192), 62 | nn.ReLU(True), 63 | ) 64 | 65 | self.a3 = Inception(192, 64, 96, 128, 16, 32, 32) 66 | self.b3 = Inception(256, 128, 128, 192, 32, 96, 64) 67 | 68 | self.maxpool = nn.MaxPool2d(3, stride=2, padding=1) 69 | 70 | self.a4 = Inception(480, 192, 96, 208, 16, 48, 64) 71 | self.b4 = Inception(512, 160, 112, 224, 24, 64, 64) 72 | self.c4 = Inception(512, 128, 128, 256, 24, 64, 64) 73 | self.d4 = Inception(512, 112, 144, 288, 32, 64, 64) 74 | self.e4 = Inception(528, 256, 160, 320, 32, 128, 128) 75 | 76 | self.a5 = Inception(832, 256, 160, 320, 32, 128, 128) 77 | self.b5 = Inception(832, 384, 192, 384, 48, 128, 128) 78 | 79 | self.avgpool = nn.AvgPool2d(8, stride=1) 80 | self.linear = nn.Linear(1024, 10) 81 | 82 | def forward(self, x): 83 | out = self.pre_layers(x) 84 | out = self.a3(out) 85 | out = self.b3(out) 86 | out = self.maxpool(out) 87 | out = self.a4(out) 88 | out = self.b4(out) 89 | out = self.c4(out) 90 | out = self.d4(out) 91 | out = self.e4(out) 92 | out = self.maxpool(out) 93 | out = self.a5(out) 94 | out = self.b5(out) 95 | out = self.avgpool(out) 96 | out = out.view(out.size(0), -1) 97 | out = self.linear(out) 98 | return out 99 | 100 | 101 | def test(): 102 | net = GoogLeNet() 103 | x = torch.randn(1,3,32,32) 104 | y = net(x) 105 | print(y.size()) 106 | 107 | # test() 108 | -------------------------------------------------------------------------------- /models/resnext.py: -------------------------------------------------------------------------------- 1 | '''ResNeXt in PyTorch. 2 | 3 | See the paper "Aggregated Residual Transformations for Deep Neural Networks" for more details. 4 | ''' 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | 10 | class Block(nn.Module): 11 | '''Grouped convolution block.''' 12 | expansion = 2 13 | 14 | def __init__(self, in_planes, cardinality=32, bottleneck_width=4, stride=1): 15 | super(Block, self).__init__() 16 | group_width = cardinality * bottleneck_width 17 | self.conv1 = nn.Conv2d(in_planes, group_width, kernel_size=1, bias=False) 18 | self.bn1 = nn.BatchNorm2d(group_width) 19 | self.conv2 = nn.Conv2d(group_width, group_width, kernel_size=3, stride=stride, padding=1, groups=cardinality, bias=False) 20 | self.bn2 = nn.BatchNorm2d(group_width) 21 | self.conv3 = nn.Conv2d(group_width, self.expansion*group_width, kernel_size=1, bias=False) 22 | self.bn3 = nn.BatchNorm2d(self.expansion*group_width) 23 | 24 | self.shortcut = nn.Sequential() 25 | if stride != 1 or in_planes != self.expansion*group_width: 26 | self.shortcut = nn.Sequential( 27 | nn.Conv2d(in_planes, self.expansion*group_width, kernel_size=1, stride=stride, bias=False), 28 | nn.BatchNorm2d(self.expansion*group_width) 29 | ) 30 | 31 | def forward(self, x): 32 | out = F.relu(self.bn1(self.conv1(x))) 33 | out = F.relu(self.bn2(self.conv2(out))) 34 | out = self.bn3(self.conv3(out)) 35 | out += self.shortcut(x) 36 | out = F.relu(out) 37 | return out 38 | 39 | 40 | class ResNeXt(nn.Module): 41 | def __init__(self, num_blocks, cardinality, bottleneck_width, num_classes=10): 42 | super(ResNeXt, self).__init__() 43 | self.cardinality = cardinality 44 | self.bottleneck_width = bottleneck_width 45 | self.in_planes = 64 46 | 47 | self.conv1 = nn.Conv2d(3, 64, kernel_size=1, bias=False) 48 | self.bn1 = nn.BatchNorm2d(64) 49 | self.layer1 = self._make_layer(num_blocks[0], 1) 50 | self.layer2 = self._make_layer(num_blocks[1], 2) 51 | self.layer3 = self._make_layer(num_blocks[2], 2) 52 | # self.layer4 = self._make_layer(num_blocks[3], 2) 53 | self.linear = nn.Linear(cardinality*bottleneck_width*8, num_classes) 54 | 55 | def _make_layer(self, num_blocks, stride): 56 | strides = [stride] + [1]*(num_blocks-1) 57 | layers = [] 58 | for stride in strides: 59 | layers.append(Block(self.in_planes, self.cardinality, self.bottleneck_width, stride)) 60 | self.in_planes = Block.expansion * self.cardinality * self.bottleneck_width 61 | # Increase bottleneck_width by 2 after each stage. 62 | self.bottleneck_width *= 2 63 | return nn.Sequential(*layers) 64 | 65 | def forward(self, x): 66 | out = F.relu(self.bn1(self.conv1(x))) 67 | out = self.layer1(out) 68 | out = self.layer2(out) 69 | out = self.layer3(out) 70 | # out = self.layer4(out) 71 | out = F.avg_pool2d(out, 8) 72 | out = out.view(out.size(0), -1) 73 | out = self.linear(out) 74 | return out 75 | 76 | 77 | def ResNeXt29_2x64d(): 78 | return ResNeXt(num_blocks=[3,3,3], cardinality=2, bottleneck_width=64) 79 | 80 | def ResNeXt29_4x64d(): 81 | return ResNeXt(num_blocks=[3,3,3], cardinality=4, bottleneck_width=64) 82 | 83 | def ResNeXt29_8x64d(): 84 | return ResNeXt(num_blocks=[3,3,3], cardinality=8, bottleneck_width=64) 85 | 86 | def ResNeXt29_32x4d(): 87 | return ResNeXt(num_blocks=[3,3,3], cardinality=32, bottleneck_width=4) 88 | 89 | def test_resnext(): 90 | net = ResNeXt29_2x64d() 91 | x = torch.randn(1,3,32,32) 92 | y = net(x) 93 | print(y.size()) 94 | 95 | # test_resnext() 96 | -------------------------------------------------------------------------------- /models/dpn.py: -------------------------------------------------------------------------------- 1 | '''Dual Path Networks in PyTorch.''' 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | 6 | 7 | class Bottleneck(nn.Module): 8 | def __init__(self, last_planes, in_planes, out_planes, dense_depth, stride, first_layer): 9 | super(Bottleneck, self).__init__() 10 | self.out_planes = out_planes 11 | self.dense_depth = dense_depth 12 | 13 | self.conv1 = nn.Conv2d(last_planes, in_planes, kernel_size=1, bias=False) 14 | self.bn1 = nn.BatchNorm2d(in_planes) 15 | self.conv2 = nn.Conv2d(in_planes, in_planes, kernel_size=3, stride=stride, padding=1, groups=32, bias=False) 16 | self.bn2 = nn.BatchNorm2d(in_planes) 17 | self.conv3 = nn.Conv2d(in_planes, out_planes+dense_depth, kernel_size=1, bias=False) 18 | self.bn3 = nn.BatchNorm2d(out_planes+dense_depth) 19 | 20 | self.shortcut = nn.Sequential() 21 | if first_layer: 22 | self.shortcut = nn.Sequential( 23 | nn.Conv2d(last_planes, out_planes+dense_depth, kernel_size=1, stride=stride, bias=False), 24 | nn.BatchNorm2d(out_planes+dense_depth) 25 | ) 26 | 27 | def forward(self, x): 28 | out = F.relu(self.bn1(self.conv1(x))) 29 | out = F.relu(self.bn2(self.conv2(out))) 30 | out = self.bn3(self.conv3(out)) 31 | x = self.shortcut(x) 32 | d = self.out_planes 33 | out = torch.cat([x[:,:d,:,:]+out[:,:d,:,:], x[:,d:,:,:], out[:,d:,:,:]], 1) 34 | out = F.relu(out) 35 | return out 36 | 37 | 38 | class DPN(nn.Module): 39 | def __init__(self, cfg): 40 | super(DPN, self).__init__() 41 | in_planes, out_planes = cfg['in_planes'], cfg['out_planes'] 42 | num_blocks, dense_depth = cfg['num_blocks'], cfg['dense_depth'] 43 | 44 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) 45 | self.bn1 = nn.BatchNorm2d(64) 46 | self.last_planes = 64 47 | self.layer1 = self._make_layer(in_planes[0], out_planes[0], num_blocks[0], dense_depth[0], stride=1) 48 | self.layer2 = self._make_layer(in_planes[1], out_planes[1], num_blocks[1], dense_depth[1], stride=2) 49 | self.layer3 = self._make_layer(in_planes[2], out_planes[2], num_blocks[2], dense_depth[2], stride=2) 50 | self.layer4 = self._make_layer(in_planes[3], out_planes[3], num_blocks[3], dense_depth[3], stride=2) 51 | self.linear = nn.Linear(out_planes[3]+(num_blocks[3]+1)*dense_depth[3], 10) 52 | 53 | def _make_layer(self, in_planes, out_planes, num_blocks, dense_depth, stride): 54 | strides = [stride] + [1]*(num_blocks-1) 55 | layers = [] 56 | for i,stride in enumerate(strides): 57 | layers.append(Bottleneck(self.last_planes, in_planes, out_planes, dense_depth, stride, i==0)) 58 | self.last_planes = out_planes + (i+2) * dense_depth 59 | return nn.Sequential(*layers) 60 | 61 | def forward(self, x): 62 | out = F.relu(self.bn1(self.conv1(x))) 63 | out = self.layer1(out) 64 | out = self.layer2(out) 65 | out = self.layer3(out) 66 | out = self.layer4(out) 67 | out = F.avg_pool2d(out, 4) 68 | out = out.view(out.size(0), -1) 69 | out = self.linear(out) 70 | return out 71 | 72 | 73 | def DPN26(): 74 | cfg = { 75 | 'in_planes': (96,192,384,768), 76 | 'out_planes': (256,512,1024,2048), 77 | 'num_blocks': (2,2,2,2), 78 | 'dense_depth': (16,32,24,128) 79 | } 80 | return DPN(cfg) 81 | 82 | def DPN92(): 83 | cfg = { 84 | 'in_planes': (96,192,384,768), 85 | 'out_planes': (256,512,1024,2048), 86 | 'num_blocks': (3,4,20,3), 87 | 'dense_depth': (16,32,24,128) 88 | } 89 | return DPN(cfg) 90 | 91 | 92 | def test(): 93 | net = DPN92() 94 | x = torch.randn(1,3,32,32) 95 | y = net(x) 96 | print(y) 97 | 98 | # test() 99 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | '''Some helper functions for PyTorch, including: 2 | - get_mean_and_std: calculate the mean and std value of dataset. 3 | - msr_init: net parameter initialization. 4 | - progress_bar: progress bar mimic xlua.progress. 5 | ''' 6 | import os 7 | import sys 8 | import time 9 | import math 10 | 11 | import torch.nn as nn 12 | import torch.nn.init as init 13 | 14 | 15 | def get_mean_and_std(dataset): 16 | '''Compute the mean and std value of dataset.''' 17 | dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2) 18 | mean = torch.zeros(3) 19 | std = torch.zeros(3) 20 | print('==> Computing mean and std..') 21 | for inputs, targets in dataloader: 22 | for i in range(3): 23 | mean[i] += inputs[:,i,:,:].mean() 24 | std[i] += inputs[:,i,:,:].std() 25 | mean.div_(len(dataset)) 26 | std.div_(len(dataset)) 27 | return mean, std 28 | 29 | def init_params(net): 30 | '''Init layer parameters.''' 31 | for m in net.modules(): 32 | if isinstance(m, nn.Conv2d): 33 | init.kaiming_normal(m.weight, mode='fan_out') 34 | if m.bias: 35 | init.constant(m.bias, 0) 36 | elif isinstance(m, nn.BatchNorm2d): 37 | init.constant(m.weight, 1) 38 | init.constant(m.bias, 0) 39 | elif isinstance(m, nn.Linear): 40 | init.normal(m.weight, std=1e-3) 41 | if m.bias: 42 | init.constant(m.bias, 0) 43 | 44 | 45 | _, term_width = os.popen('stty size', 'r').read().split() 46 | term_width = int(term_width) 47 | 48 | TOTAL_BAR_LENGTH = 65. 49 | last_time = time.time() 50 | begin_time = last_time 51 | def progress_bar(current, total, msg=None): 52 | global last_time, begin_time 53 | if current == 0: 54 | begin_time = time.time() # Reset for new bar. 55 | 56 | cur_len = int(TOTAL_BAR_LENGTH*current/total) 57 | rest_len = int(TOTAL_BAR_LENGTH - cur_len) - 1 58 | 59 | sys.stdout.write(' [') 60 | for i in range(cur_len): 61 | sys.stdout.write('=') 62 | sys.stdout.write('>') 63 | for i in range(rest_len): 64 | sys.stdout.write('.') 65 | sys.stdout.write(']') 66 | 67 | cur_time = time.time() 68 | step_time = cur_time - last_time 69 | last_time = cur_time 70 | tot_time = cur_time - begin_time 71 | 72 | L = [] 73 | L.append(' Step: %s' % format_time(step_time)) 74 | L.append(' | Tot: %s' % format_time(tot_time)) 75 | if msg: 76 | L.append(' | ' + msg) 77 | 78 | msg = ''.join(L) 79 | sys.stdout.write(msg) 80 | for i in range(term_width-int(TOTAL_BAR_LENGTH)-len(msg)-3): 81 | sys.stdout.write(' ') 82 | 83 | # Go back to the center of the bar. 84 | for i in range(term_width-int(TOTAL_BAR_LENGTH/2)+2): 85 | sys.stdout.write('\b') 86 | sys.stdout.write(' %d/%d ' % (current+1, total)) 87 | 88 | if current < total-1: 89 | sys.stdout.write('\r') 90 | else: 91 | sys.stdout.write('\n') 92 | sys.stdout.flush() 93 | 94 | def format_time(seconds): 95 | days = int(seconds / 3600/24) 96 | seconds = seconds - days*3600*24 97 | hours = int(seconds / 3600) 98 | seconds = seconds - hours*3600 99 | minutes = int(seconds / 60) 100 | seconds = seconds - minutes*60 101 | secondsf = int(seconds) 102 | seconds = seconds - secondsf 103 | millis = int(seconds*1000) 104 | 105 | f = '' 106 | i = 1 107 | if days > 0: 108 | f += str(days) + 'D' 109 | i += 1 110 | if hours > 0 and i <= 2: 111 | f += str(hours) + 'h' 112 | i += 1 113 | if minutes > 0 and i <= 2: 114 | f += str(minutes) + 'm' 115 | i += 1 116 | if secondsf > 0 and i <= 2: 117 | f += str(secondsf) + 's' 118 | i += 1 119 | if millis > 0 and i <= 2: 120 | f += str(millis) + 'ms' 121 | i += 1 122 | if f == '': 123 | f = '0ms' 124 | return f 125 | -------------------------------------------------------------------------------- /models/densenet.py: -------------------------------------------------------------------------------- 1 | '''DenseNet in PyTorch.''' 2 | import math 3 | 4 | import torch 5 | import torch.nn as nn 6 | import torch.nn.functional as F 7 | 8 | 9 | class Bottleneck(nn.Module): 10 | def __init__(self, in_planes, growth_rate): 11 | super(Bottleneck, self).__init__() 12 | self.bn1 = nn.BatchNorm2d(in_planes) 13 | self.conv1 = nn.Conv2d(in_planes, 4*growth_rate, kernel_size=1, bias=False) 14 | self.bn2 = nn.BatchNorm2d(4*growth_rate) 15 | self.conv2 = nn.Conv2d(4*growth_rate, growth_rate, kernel_size=3, padding=1, bias=False) 16 | 17 | def forward(self, x): 18 | out = self.conv1(F.relu(self.bn1(x))) 19 | out = self.conv2(F.relu(self.bn2(out))) 20 | out = torch.cat([out,x], 1) 21 | return out 22 | 23 | 24 | class Transition(nn.Module): 25 | def __init__(self, in_planes, out_planes): 26 | super(Transition, self).__init__() 27 | self.bn = nn.BatchNorm2d(in_planes) 28 | self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=1, bias=False) 29 | 30 | def forward(self, x): 31 | out = self.conv(F.relu(self.bn(x))) 32 | out = F.avg_pool2d(out, 2) 33 | return out 34 | 35 | 36 | class DenseNet(nn.Module): 37 | def __init__(self, block, nblocks, growth_rate=12, reduction=0.5, num_classes=10): 38 | super(DenseNet, self).__init__() 39 | self.growth_rate = growth_rate 40 | 41 | num_planes = 2*growth_rate 42 | self.conv1 = nn.Conv2d(3, num_planes, kernel_size=3, padding=1, bias=False) 43 | 44 | self.dense1 = self._make_dense_layers(block, num_planes, nblocks[0]) 45 | num_planes += nblocks[0]*growth_rate 46 | out_planes = int(math.floor(num_planes*reduction)) 47 | self.trans1 = Transition(num_planes, out_planes) 48 | num_planes = out_planes 49 | 50 | self.dense2 = self._make_dense_layers(block, num_planes, nblocks[1]) 51 | num_planes += nblocks[1]*growth_rate 52 | out_planes = int(math.floor(num_planes*reduction)) 53 | self.trans2 = Transition(num_planes, out_planes) 54 | num_planes = out_planes 55 | 56 | self.dense3 = self._make_dense_layers(block, num_planes, nblocks[2]) 57 | num_planes += nblocks[2]*growth_rate 58 | out_planes = int(math.floor(num_planes*reduction)) 59 | self.trans3 = Transition(num_planes, out_planes) 60 | num_planes = out_planes 61 | 62 | self.dense4 = self._make_dense_layers(block, num_planes, nblocks[3]) 63 | num_planes += nblocks[3]*growth_rate 64 | 65 | self.bn = nn.BatchNorm2d(num_planes) 66 | self.linear = nn.Linear(num_planes, num_classes) 67 | 68 | def _make_dense_layers(self, block, in_planes, nblock): 69 | layers = [] 70 | for i in range(nblock): 71 | layers.append(block(in_planes, self.growth_rate)) 72 | in_planes += self.growth_rate 73 | return nn.Sequential(*layers) 74 | 75 | def forward(self, x): 76 | out = self.conv1(x) 77 | out = self.trans1(self.dense1(out)) 78 | out = self.trans2(self.dense2(out)) 79 | out = self.trans3(self.dense3(out)) 80 | out = self.dense4(out) 81 | out = F.avg_pool2d(F.relu(self.bn(out)), 4) 82 | out = out.view(out.size(0), -1) 83 | out = self.linear(out) 84 | return out 85 | 86 | def DenseNet121(): 87 | return DenseNet(Bottleneck, [6,12,24,16], growth_rate=32) 88 | 89 | def DenseNet169(): 90 | return DenseNet(Bottleneck, [6,12,32,32], growth_rate=32) 91 | 92 | def DenseNet201(): 93 | return DenseNet(Bottleneck, [6,12,48,32], growth_rate=32) 94 | 95 | def DenseNet161(): 96 | return DenseNet(Bottleneck, [6,12,36,24], growth_rate=48) 97 | 98 | def densenet_cifar(): 99 | return DenseNet(Bottleneck, [6,12,24,16], growth_rate=12) 100 | 101 | def test(): 102 | net = densenet_cifar() 103 | x = torch.randn(1,3,32,32) 104 | y = net(x) 105 | print(y) 106 | 107 | # test() 108 | -------------------------------------------------------------------------------- /models/shufflenet.py: -------------------------------------------------------------------------------- 1 | '''ShuffleNet in PyTorch. 2 | 3 | See the paper "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices" for more details. 4 | ''' 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | 10 | class ShuffleBlock(nn.Module): 11 | def __init__(self, groups): 12 | super(ShuffleBlock, self).__init__() 13 | self.groups = groups 14 | 15 | def forward(self, x): 16 | '''Channel shuffle: [N,C,H,W] -> [N,g,C/g,H,W] -> [N,C/g,g,H,w] -> [N,C,H,W]''' 17 | N,C,H,W = x.size() 18 | g = self.groups 19 | return x.view(N,g,C//g,H,W).permute(0,2,1,3,4).reshape(N,C,H,W) 20 | 21 | 22 | class Bottleneck(nn.Module): 23 | def __init__(self, in_planes, out_planes, stride, groups): 24 | super(Bottleneck, self).__init__() 25 | self.stride = stride 26 | 27 | mid_planes = out_planes/4 28 | g = 1 if in_planes==24 else groups 29 | self.conv1 = nn.Conv2d(in_planes, mid_planes, kernel_size=1, groups=g, bias=False) 30 | self.bn1 = nn.BatchNorm2d(mid_planes) 31 | self.shuffle1 = ShuffleBlock(groups=g) 32 | self.conv2 = nn.Conv2d(mid_planes, mid_planes, kernel_size=3, stride=stride, padding=1, groups=mid_planes, bias=False) 33 | self.bn2 = nn.BatchNorm2d(mid_planes) 34 | self.conv3 = nn.Conv2d(mid_planes, out_planes, kernel_size=1, groups=groups, bias=False) 35 | self.bn3 = nn.BatchNorm2d(out_planes) 36 | 37 | self.shortcut = nn.Sequential() 38 | if stride == 2: 39 | self.shortcut = nn.Sequential(nn.AvgPool2d(3, stride=2, padding=1)) 40 | 41 | def forward(self, x): 42 | out = F.relu(self.bn1(self.conv1(x))) 43 | out = self.shuffle1(out) 44 | out = F.relu(self.bn2(self.conv2(out))) 45 | out = self.bn3(self.conv3(out)) 46 | res = self.shortcut(x) 47 | out = F.relu(torch.cat([out,res], 1)) if self.stride==2 else F.relu(out+res) 48 | return out 49 | 50 | 51 | class ShuffleNet(nn.Module): 52 | def __init__(self, cfg): 53 | super(ShuffleNet, self).__init__() 54 | out_planes = cfg['out_planes'] 55 | num_blocks = cfg['num_blocks'] 56 | groups = cfg['groups'] 57 | 58 | self.conv1 = nn.Conv2d(3, 24, kernel_size=1, bias=False) 59 | self.bn1 = nn.BatchNorm2d(24) 60 | self.in_planes = 24 61 | self.layer1 = self._make_layer(out_planes[0], num_blocks[0], groups) 62 | self.layer2 = self._make_layer(out_planes[1], num_blocks[1], groups) 63 | self.layer3 = self._make_layer(out_planes[2], num_blocks[2], groups) 64 | self.linear = nn.Linear(out_planes[2], 10) 65 | 66 | def _make_layer(self, out_planes, num_blocks, groups): 67 | layers = [] 68 | for i in range(num_blocks): 69 | stride = 2 if i == 0 else 1 70 | cat_planes = self.in_planes if i == 0 else 0 71 | layers.append(Bottleneck(self.in_planes, out_planes-cat_planes, stride=stride, groups=groups)) 72 | self.in_planes = out_planes 73 | return nn.Sequential(*layers) 74 | 75 | def forward(self, x): 76 | out = F.relu(self.bn1(self.conv1(x))) 77 | out = self.layer1(out) 78 | out = self.layer2(out) 79 | out = self.layer3(out) 80 | out = F.avg_pool2d(out, 4) 81 | out = out.view(out.size(0), -1) 82 | out = self.linear(out) 83 | return out 84 | 85 | 86 | def ShuffleNetG2(): 87 | cfg = { 88 | 'out_planes': [200,400,800], 89 | 'num_blocks': [4,8,4], 90 | 'groups': 2 91 | } 92 | return ShuffleNet(cfg) 93 | 94 | def ShuffleNetG3(): 95 | cfg = { 96 | 'out_planes': [240,480,960], 97 | 'num_blocks': [4,8,4], 98 | 'groups': 3 99 | } 100 | return ShuffleNet(cfg) 101 | 102 | 103 | def test(): 104 | net = ShuffleNetG2() 105 | x = torch.randn(1,3,32,32) 106 | y = net(x) 107 | print(y) 108 | 109 | # test() 110 | -------------------------------------------------------------------------------- /models/senet.py: -------------------------------------------------------------------------------- 1 | '''SENet in PyTorch. 2 | 3 | SENet is the winner of ImageNet-2017. The paper is not released yet. 4 | ''' 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | 10 | class BasicBlock(nn.Module): 11 | def __init__(self, in_planes, planes, stride=1): 12 | super(BasicBlock, self).__init__() 13 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 14 | self.bn1 = nn.BatchNorm2d(planes) 15 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) 16 | self.bn2 = nn.BatchNorm2d(planes) 17 | 18 | self.shortcut = nn.Sequential() 19 | if stride != 1 or in_planes != planes: 20 | self.shortcut = nn.Sequential( 21 | nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=False), 22 | nn.BatchNorm2d(planes) 23 | ) 24 | 25 | # SE layers 26 | self.fc1 = nn.Conv2d(planes, planes//16, kernel_size=1) # Use nn.Conv2d instead of nn.Linear 27 | self.fc2 = nn.Conv2d(planes//16, planes, kernel_size=1) 28 | 29 | def forward(self, x): 30 | out = F.relu(self.bn1(self.conv1(x))) 31 | out = self.bn2(self.conv2(out)) 32 | 33 | # Squeeze 34 | w = F.avg_pool2d(out, out.size(2)) 35 | w = F.relu(self.fc1(w)) 36 | w = F.sigmoid(self.fc2(w)) 37 | # Excitation 38 | out = out * w # New broadcasting feature from v0.2! 39 | 40 | out += self.shortcut(x) 41 | out = F.relu(out) 42 | return out 43 | 44 | 45 | class PreActBlock(nn.Module): 46 | def __init__(self, in_planes, planes, stride=1): 47 | super(PreActBlock, self).__init__() 48 | self.bn1 = nn.BatchNorm2d(in_planes) 49 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 50 | self.bn2 = nn.BatchNorm2d(planes) 51 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) 52 | 53 | if stride != 1 or in_planes != planes: 54 | self.shortcut = nn.Sequential( 55 | nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=False) 56 | ) 57 | 58 | # SE layers 59 | self.fc1 = nn.Conv2d(planes, planes//16, kernel_size=1) 60 | self.fc2 = nn.Conv2d(planes//16, planes, kernel_size=1) 61 | 62 | def forward(self, x): 63 | out = F.relu(self.bn1(x)) 64 | shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x 65 | out = self.conv1(out) 66 | out = self.conv2(F.relu(self.bn2(out))) 67 | 68 | # Squeeze 69 | w = F.avg_pool2d(out, out.size(2)) 70 | w = F.relu(self.fc1(w)) 71 | w = F.sigmoid(self.fc2(w)) 72 | # Excitation 73 | out = out * w 74 | 75 | out += shortcut 76 | return out 77 | 78 | 79 | class SENet(nn.Module): 80 | def __init__(self, block, num_blocks, num_classes=10): 81 | super(SENet, self).__init__() 82 | self.in_planes = 64 83 | 84 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) 85 | self.bn1 = nn.BatchNorm2d(64) 86 | self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) 87 | self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) 88 | self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) 89 | self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) 90 | self.linear = nn.Linear(512, num_classes) 91 | 92 | def _make_layer(self, block, planes, num_blocks, stride): 93 | strides = [stride] + [1]*(num_blocks-1) 94 | layers = [] 95 | for stride in strides: 96 | layers.append(block(self.in_planes, planes, stride)) 97 | self.in_planes = planes 98 | return nn.Sequential(*layers) 99 | 100 | def forward(self, x): 101 | out = F.relu(self.bn1(self.conv1(x))) 102 | out = self.layer1(out) 103 | out = self.layer2(out) 104 | out = self.layer3(out) 105 | out = self.layer4(out) 106 | out = F.avg_pool2d(out, 4) 107 | out = out.view(out.size(0), -1) 108 | out = self.linear(out) 109 | return out 110 | 111 | 112 | def SENet18(): 113 | return SENet(PreActBlock, [2,2,2,2]) 114 | 115 | 116 | def test(): 117 | net = SENet18() 118 | y = net(torch.randn(1,3,32,32)) 119 | print(y.size()) 120 | 121 | # test() 122 | -------------------------------------------------------------------------------- /models/preact_resnet.py: -------------------------------------------------------------------------------- 1 | '''Pre-activation ResNet in PyTorch. 2 | 3 | Reference: 4 | [1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun 5 | Identity Mappings in Deep Residual Networks. arXiv:1603.05027 6 | ''' 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | 11 | 12 | class PreActBlock(nn.Module): 13 | '''Pre-activation version of the BasicBlock.''' 14 | expansion = 1 15 | 16 | def __init__(self, in_planes, planes, stride=1): 17 | super(PreActBlock, self).__init__() 18 | self.bn1 = nn.BatchNorm2d(in_planes) 19 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 20 | self.bn2 = nn.BatchNorm2d(planes) 21 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) 22 | 23 | if stride != 1 or in_planes != self.expansion*planes: 24 | self.shortcut = nn.Sequential( 25 | nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False) 26 | ) 27 | 28 | def forward(self, x): 29 | out = F.relu(self.bn1(x)) 30 | shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x 31 | out = self.conv1(out) 32 | out = self.conv2(F.relu(self.bn2(out))) 33 | out += shortcut 34 | return out 35 | 36 | 37 | class PreActBottleneck(nn.Module): 38 | '''Pre-activation version of the original Bottleneck module.''' 39 | expansion = 4 40 | 41 | def __init__(self, in_planes, planes, stride=1): 42 | super(PreActBottleneck, self).__init__() 43 | self.bn1 = nn.BatchNorm2d(in_planes) 44 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) 45 | self.bn2 = nn.BatchNorm2d(planes) 46 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 47 | self.bn3 = nn.BatchNorm2d(planes) 48 | self.conv3 = nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=False) 49 | 50 | if stride != 1 or in_planes != self.expansion*planes: 51 | self.shortcut = nn.Sequential( 52 | nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False) 53 | ) 54 | 55 | def forward(self, x): 56 | out = F.relu(self.bn1(x)) 57 | shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x 58 | out = self.conv1(out) 59 | out = self.conv2(F.relu(self.bn2(out))) 60 | out = self.conv3(F.relu(self.bn3(out))) 61 | out += shortcut 62 | return out 63 | 64 | 65 | class PreActResNet(nn.Module): 66 | def __init__(self, block, num_blocks, num_classes=10): 67 | super(PreActResNet, self).__init__() 68 | self.in_planes = 64 69 | 70 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) 71 | self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) 72 | self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) 73 | self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) 74 | self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) 75 | self.linear = nn.Linear(512*block.expansion, num_classes) 76 | 77 | def _make_layer(self, block, planes, num_blocks, stride): 78 | strides = [stride] + [1]*(num_blocks-1) 79 | layers = [] 80 | for stride in strides: 81 | layers.append(block(self.in_planes, planes, stride)) 82 | self.in_planes = planes * block.expansion 83 | return nn.Sequential(*layers) 84 | 85 | def forward(self, x): 86 | out = self.conv1(x) 87 | out = self.layer1(out) 88 | out = self.layer2(out) 89 | out = self.layer3(out) 90 | out = self.layer4(out) 91 | out = F.avg_pool2d(out, 4) 92 | out = out.view(out.size(0), -1) 93 | out = self.linear(out) 94 | return out 95 | 96 | 97 | def PreActResNet18(): 98 | return PreActResNet(PreActBlock, [2,2,2,2]) 99 | 100 | def PreActResNet34(): 101 | return PreActResNet(PreActBlock, [3,4,6,3]) 102 | 103 | def PreActResNet50(): 104 | return PreActResNet(PreActBottleneck, [3,4,6,3]) 105 | 106 | def PreActResNet101(): 107 | return PreActResNet(PreActBottleneck, [3,4,23,3]) 108 | 109 | def PreActResNet152(): 110 | return PreActResNet(PreActBottleneck, [3,8,36,3]) 111 | 112 | 113 | def test(): 114 | net = PreActResNet18() 115 | y = net((torch.randn(1,3,32,32))) 116 | print(y.size()) 117 | 118 | # test() 119 | -------------------------------------------------------------------------------- /models/dla_simple.py: -------------------------------------------------------------------------------- 1 | '''Simplified version of DLA in PyTorch. 2 | 3 | Note this implementation is not identical to the original paper version. 4 | But it seems works fine. 5 | 6 | See dla.py for the original paper version. 7 | 8 | Reference: 9 | Deep Layer Aggregation. https://arxiv.org/abs/1707.06484 10 | ''' 11 | import torch 12 | import torch.nn as nn 13 | import torch.nn.functional as F 14 | 15 | 16 | class BasicBlock(nn.Module): 17 | expansion = 1 18 | 19 | def __init__(self, in_planes, planes, stride=1): 20 | super(BasicBlock, self).__init__() 21 | self.conv1 = nn.Conv2d( 22 | in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 23 | self.bn1 = nn.BatchNorm2d(planes) 24 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, 25 | stride=1, padding=1, bias=False) 26 | self.bn2 = nn.BatchNorm2d(planes) 27 | 28 | self.shortcut = nn.Sequential() 29 | if stride != 1 or in_planes != self.expansion*planes: 30 | self.shortcut = nn.Sequential( 31 | nn.Conv2d(in_planes, self.expansion*planes, 32 | kernel_size=1, stride=stride, bias=False), 33 | nn.BatchNorm2d(self.expansion*planes) 34 | ) 35 | 36 | def forward(self, x): 37 | out = F.relu(self.bn1(self.conv1(x))) 38 | out = self.bn2(self.conv2(out)) 39 | out += self.shortcut(x) 40 | out = F.relu(out) 41 | return out 42 | 43 | 44 | class Root(nn.Module): 45 | def __init__(self, in_channels, out_channels, kernel_size=1): 46 | super(Root, self).__init__() 47 | self.conv = nn.Conv2d( 48 | in_channels, out_channels, kernel_size, 49 | stride=1, padding=(kernel_size - 1) // 2, bias=False) 50 | self.bn = nn.BatchNorm2d(out_channels) 51 | 52 | def forward(self, xs): 53 | x = torch.cat(xs, 1) 54 | out = F.relu(self.bn(self.conv(x))) 55 | return out 56 | 57 | 58 | class Tree(nn.Module): 59 | def __init__(self, block, in_channels, out_channels, level=1, stride=1): 60 | super(Tree, self).__init__() 61 | self.root = Root(2*out_channels, out_channels) 62 | if level == 1: 63 | self.left_tree = block(in_channels, out_channels, stride=stride) 64 | self.right_tree = block(out_channels, out_channels, stride=1) 65 | else: 66 | self.left_tree = Tree(block, in_channels, 67 | out_channels, level=level-1, stride=stride) 68 | self.right_tree = Tree(block, out_channels, 69 | out_channels, level=level-1, stride=1) 70 | 71 | def forward(self, x): 72 | out1 = self.left_tree(x) 73 | out2 = self.right_tree(out1) 74 | out = self.root([out1, out2]) 75 | return out 76 | 77 | 78 | class SimpleDLA(nn.Module): 79 | def __init__(self, block=BasicBlock, num_classes=10): 80 | super(SimpleDLA, self).__init__() 81 | self.base = nn.Sequential( 82 | nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False), 83 | nn.BatchNorm2d(16), 84 | nn.ReLU(True) 85 | ) 86 | 87 | self.layer1 = nn.Sequential( 88 | nn.Conv2d(16, 16, kernel_size=3, stride=1, padding=1, bias=False), 89 | nn.BatchNorm2d(16), 90 | nn.ReLU(True) 91 | ) 92 | 93 | self.layer2 = nn.Sequential( 94 | nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1, bias=False), 95 | nn.BatchNorm2d(32), 96 | nn.ReLU(True) 97 | ) 98 | 99 | self.layer3 = Tree(block, 32, 64, level=1, stride=1) 100 | self.layer4 = Tree(block, 64, 128, level=2, stride=2) 101 | self.layer5 = Tree(block, 128, 256, level=2, stride=2) 102 | self.layer6 = Tree(block, 256, 512, level=1, stride=2) 103 | self.linear = nn.Linear(512, num_classes) 104 | 105 | def forward(self, x): 106 | out = self.base(x) 107 | out = self.layer1(out) 108 | out = self.layer2(out) 109 | out = self.layer3(out) 110 | out = self.layer4(out) 111 | out = self.layer5(out) 112 | out = self.layer6(out) 113 | out = F.avg_pool2d(out, 4) 114 | out = out.view(out.size(0), -1) 115 | out = self.linear(out) 116 | return out 117 | 118 | 119 | def test(): 120 | net = SimpleDLA() 121 | print(net) 122 | x = torch.randn(1, 3, 32, 32) 123 | y = net(x) 124 | print(y.size()) 125 | 126 | 127 | if __name__ == '__main__': 128 | test() 129 | -------------------------------------------------------------------------------- /models/pnasnet.py: -------------------------------------------------------------------------------- 1 | '''PNASNet in PyTorch. 2 | 3 | Paper: Progressive Neural Architecture Search 4 | ''' 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | 10 | class SepConv(nn.Module): 11 | '''Separable Convolution.''' 12 | def __init__(self, in_planes, out_planes, kernel_size, stride): 13 | super(SepConv, self).__init__() 14 | self.conv1 = nn.Conv2d(in_planes, out_planes, 15 | kernel_size, stride, 16 | padding=(kernel_size-1)//2, 17 | bias=False, groups=in_planes) 18 | self.bn1 = nn.BatchNorm2d(out_planes) 19 | 20 | def forward(self, x): 21 | return self.bn1(self.conv1(x)) 22 | 23 | 24 | class CellA(nn.Module): 25 | def __init__(self, in_planes, out_planes, stride=1): 26 | super(CellA, self).__init__() 27 | self.stride = stride 28 | self.sep_conv1 = SepConv(in_planes, out_planes, kernel_size=7, stride=stride) 29 | if stride==2: 30 | self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) 31 | self.bn1 = nn.BatchNorm2d(out_planes) 32 | 33 | def forward(self, x): 34 | y1 = self.sep_conv1(x) 35 | y2 = F.max_pool2d(x, kernel_size=3, stride=self.stride, padding=1) 36 | if self.stride==2: 37 | y2 = self.bn1(self.conv1(y2)) 38 | return F.relu(y1+y2) 39 | 40 | class CellB(nn.Module): 41 | def __init__(self, in_planes, out_planes, stride=1): 42 | super(CellB, self).__init__() 43 | self.stride = stride 44 | # Left branch 45 | self.sep_conv1 = SepConv(in_planes, out_planes, kernel_size=7, stride=stride) 46 | self.sep_conv2 = SepConv(in_planes, out_planes, kernel_size=3, stride=stride) 47 | # Right branch 48 | self.sep_conv3 = SepConv(in_planes, out_planes, kernel_size=5, stride=stride) 49 | if stride==2: 50 | self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) 51 | self.bn1 = nn.BatchNorm2d(out_planes) 52 | # Reduce channels 53 | self.conv2 = nn.Conv2d(2*out_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False) 54 | self.bn2 = nn.BatchNorm2d(out_planes) 55 | 56 | def forward(self, x): 57 | # Left branch 58 | y1 = self.sep_conv1(x) 59 | y2 = self.sep_conv2(x) 60 | # Right branch 61 | y3 = F.max_pool2d(x, kernel_size=3, stride=self.stride, padding=1) 62 | if self.stride==2: 63 | y3 = self.bn1(self.conv1(y3)) 64 | y4 = self.sep_conv3(x) 65 | # Concat & reduce channels 66 | b1 = F.relu(y1+y2) 67 | b2 = F.relu(y3+y4) 68 | y = torch.cat([b1,b2], 1) 69 | return F.relu(self.bn2(self.conv2(y))) 70 | 71 | class PNASNet(nn.Module): 72 | def __init__(self, cell_type, num_cells, num_planes): 73 | super(PNASNet, self).__init__() 74 | self.in_planes = num_planes 75 | self.cell_type = cell_type 76 | 77 | self.conv1 = nn.Conv2d(3, num_planes, kernel_size=3, stride=1, padding=1, bias=False) 78 | self.bn1 = nn.BatchNorm2d(num_planes) 79 | 80 | self.layer1 = self._make_layer(num_planes, num_cells=6) 81 | self.layer2 = self._downsample(num_planes*2) 82 | self.layer3 = self._make_layer(num_planes*2, num_cells=6) 83 | self.layer4 = self._downsample(num_planes*4) 84 | self.layer5 = self._make_layer(num_planes*4, num_cells=6) 85 | 86 | self.linear = nn.Linear(num_planes*4, 10) 87 | 88 | def _make_layer(self, planes, num_cells): 89 | layers = [] 90 | for _ in range(num_cells): 91 | layers.append(self.cell_type(self.in_planes, planes, stride=1)) 92 | self.in_planes = planes 93 | return nn.Sequential(*layers) 94 | 95 | def _downsample(self, planes): 96 | layer = self.cell_type(self.in_planes, planes, stride=2) 97 | self.in_planes = planes 98 | return layer 99 | 100 | def forward(self, x): 101 | out = F.relu(self.bn1(self.conv1(x))) 102 | out = self.layer1(out) 103 | out = self.layer2(out) 104 | out = self.layer3(out) 105 | out = self.layer4(out) 106 | out = self.layer5(out) 107 | out = F.avg_pool2d(out, 8) 108 | out = self.linear(out.view(out.size(0), -1)) 109 | return out 110 | 111 | 112 | def PNASNetA(): 113 | return PNASNet(CellA, num_cells=6, num_planes=44) 114 | 115 | def PNASNetB(): 116 | return PNASNet(CellB, num_cells=6, num_planes=32) 117 | 118 | 119 | def test(): 120 | net = PNASNetB() 121 | x = torch.randn(1,3,32,32) 122 | y = net(x) 123 | print(y) 124 | 125 | # test() 126 | -------------------------------------------------------------------------------- /models/resnet.py: -------------------------------------------------------------------------------- 1 | '''ResNet in PyTorch. 2 | 3 | For Pre-activation ResNet, see 'preact_resnet.py'. 4 | 5 | Reference: 6 | [1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun 7 | Deep Residual Learning for Image Recognition. arXiv:1512.03385 8 | ''' 9 | import torch 10 | import torch.nn as nn 11 | import torch.nn.functional as F 12 | 13 | 14 | class BasicBlock(nn.Module): 15 | expansion = 1 16 | 17 | def __init__(self, in_planes, planes, stride=1): 18 | super(BasicBlock, self).__init__() 19 | self.conv1 = nn.Conv2d( 20 | in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 21 | self.bn1 = nn.BatchNorm2d(planes) 22 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, 23 | stride=1, padding=1, bias=False) 24 | self.bn2 = nn.BatchNorm2d(planes) 25 | 26 | self.shortcut = nn.Sequential() 27 | if stride != 1 or in_planes != self.expansion*planes: 28 | self.shortcut = nn.Sequential( 29 | nn.Conv2d(in_planes, self.expansion*planes, 30 | kernel_size=1, stride=stride, bias=False), 31 | nn.BatchNorm2d(self.expansion*planes) 32 | ) 33 | 34 | def forward(self, x): 35 | out = F.relu(self.bn1(self.conv1(x))) 36 | out = self.bn2(self.conv2(out)) 37 | out += self.shortcut(x) 38 | out = F.relu(out) 39 | return out 40 | 41 | 42 | class Bottleneck(nn.Module): 43 | expansion = 4 44 | 45 | def __init__(self, in_planes, planes, stride=1): 46 | super(Bottleneck, self).__init__() 47 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) 48 | self.bn1 = nn.BatchNorm2d(planes) 49 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, 50 | stride=stride, padding=1, bias=False) 51 | self.bn2 = nn.BatchNorm2d(planes) 52 | self.conv3 = nn.Conv2d(planes, self.expansion * 53 | planes, kernel_size=1, bias=False) 54 | self.bn3 = nn.BatchNorm2d(self.expansion*planes) 55 | 56 | self.shortcut = nn.Sequential() 57 | if stride != 1 or in_planes != self.expansion*planes: 58 | self.shortcut = nn.Sequential( 59 | nn.Conv2d(in_planes, self.expansion*planes, 60 | kernel_size=1, stride=stride, bias=False), 61 | nn.BatchNorm2d(self.expansion*planes) 62 | ) 63 | 64 | def forward(self, x): 65 | out = F.relu(self.bn1(self.conv1(x))) 66 | out = F.relu(self.bn2(self.conv2(out))) 67 | out = self.bn3(self.conv3(out)) 68 | out += self.shortcut(x) 69 | out = F.relu(out) 70 | return out 71 | 72 | 73 | class ResNet(nn.Module): 74 | def __init__(self, block, num_blocks, num_classes=10): 75 | super(ResNet, self).__init__() 76 | self.in_planes = 64 77 | 78 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, 79 | stride=1, padding=1, bias=False) 80 | self.bn1 = nn.BatchNorm2d(64) 81 | self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) 82 | self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) 83 | self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) 84 | self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) 85 | self.linear = nn.Linear(512*block.expansion, num_classes) 86 | 87 | def _make_layer(self, block, planes, num_blocks, stride): 88 | strides = [stride] + [1]*(num_blocks-1) 89 | layers = [] 90 | for stride in strides: 91 | layers.append(block(self.in_planes, planes, stride)) 92 | self.in_planes = planes * block.expansion 93 | return nn.Sequential(*layers) 94 | 95 | def forward(self, x): 96 | out = F.relu(self.bn1(self.conv1(x))) 97 | out = self.layer1(out) 98 | out = self.layer2(out) 99 | out = self.layer3(out) 100 | out = self.layer4(out) 101 | out = F.avg_pool2d(out, 4) 102 | out = out.view(out.size(0), -1) 103 | out = self.linear(out) 104 | return out 105 | 106 | 107 | def ResNet18(): 108 | return ResNet(BasicBlock, [2, 2, 2, 2]) 109 | 110 | 111 | def ResNet34(): 112 | return ResNet(BasicBlock, [3, 4, 6, 3]) 113 | 114 | 115 | def ResNet50(): 116 | return ResNet(Bottleneck, [3, 4, 6, 3]) 117 | 118 | 119 | def ResNet101(): 120 | return ResNet(Bottleneck, [3, 4, 23, 3]) 121 | 122 | 123 | def ResNet152(): 124 | return ResNet(Bottleneck, [3, 8, 36, 3]) 125 | 126 | 127 | def test(): 128 | net = ResNet18() 129 | y = net(torch.randn(1, 3, 32, 32)) 130 | print(y.size()) 131 | 132 | # test() 133 | -------------------------------------------------------------------------------- /models/dla.py: -------------------------------------------------------------------------------- 1 | '''DLA in PyTorch. 2 | 3 | Reference: 4 | Deep Layer Aggregation. https://arxiv.org/abs/1707.06484 5 | ''' 6 | import torch 7 | import torch.nn as nn 8 | import torch.nn.functional as F 9 | 10 | 11 | class BasicBlock(nn.Module): 12 | expansion = 1 13 | 14 | def __init__(self, in_planes, planes, stride=1): 15 | super(BasicBlock, self).__init__() 16 | self.conv1 = nn.Conv2d( 17 | in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 18 | self.bn1 = nn.BatchNorm2d(planes) 19 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, 20 | stride=1, padding=1, bias=False) 21 | self.bn2 = nn.BatchNorm2d(planes) 22 | 23 | self.shortcut = nn.Sequential() 24 | if stride != 1 or in_planes != self.expansion*planes: 25 | self.shortcut = nn.Sequential( 26 | nn.Conv2d(in_planes, self.expansion*planes, 27 | kernel_size=1, stride=stride, bias=False), 28 | nn.BatchNorm2d(self.expansion*planes) 29 | ) 30 | 31 | def forward(self, x): 32 | out = F.relu(self.bn1(self.conv1(x))) 33 | out = self.bn2(self.conv2(out)) 34 | out += self.shortcut(x) 35 | out = F.relu(out) 36 | return out 37 | 38 | 39 | class Root(nn.Module): 40 | def __init__(self, in_channels, out_channels, kernel_size=1): 41 | super(Root, self).__init__() 42 | self.conv = nn.Conv2d( 43 | in_channels, out_channels, kernel_size, 44 | stride=1, padding=(kernel_size - 1) // 2, bias=False) 45 | self.bn = nn.BatchNorm2d(out_channels) 46 | 47 | def forward(self, xs): 48 | x = torch.cat(xs, 1) 49 | out = F.relu(self.bn(self.conv(x))) 50 | return out 51 | 52 | 53 | class Tree(nn.Module): 54 | def __init__(self, block, in_channels, out_channels, level=1, stride=1): 55 | super(Tree, self).__init__() 56 | self.level = level 57 | if level == 1: 58 | self.root = Root(2*out_channels, out_channels) 59 | self.left_node = block(in_channels, out_channels, stride=stride) 60 | self.right_node = block(out_channels, out_channels, stride=1) 61 | else: 62 | self.root = Root((level+2)*out_channels, out_channels) 63 | for i in reversed(range(1, level)): 64 | subtree = Tree(block, in_channels, out_channels, 65 | level=i, stride=stride) 66 | self.__setattr__('level_%d' % i, subtree) 67 | self.prev_root = block(in_channels, out_channels, stride=stride) 68 | self.left_node = block(out_channels, out_channels, stride=1) 69 | self.right_node = block(out_channels, out_channels, stride=1) 70 | 71 | def forward(self, x): 72 | xs = [self.prev_root(x)] if self.level > 1 else [] 73 | for i in reversed(range(1, self.level)): 74 | level_i = self.__getattr__('level_%d' % i) 75 | x = level_i(x) 76 | xs.append(x) 77 | x = self.left_node(x) 78 | xs.append(x) 79 | x = self.right_node(x) 80 | xs.append(x) 81 | out = self.root(xs) 82 | return out 83 | 84 | 85 | class DLA(nn.Module): 86 | def __init__(self, block=BasicBlock, num_classes=10): 87 | super(DLA, self).__init__() 88 | self.base = nn.Sequential( 89 | nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False), 90 | nn.BatchNorm2d(16), 91 | nn.ReLU(True) 92 | ) 93 | 94 | self.layer1 = nn.Sequential( 95 | nn.Conv2d(16, 16, kernel_size=3, stride=1, padding=1, bias=False), 96 | nn.BatchNorm2d(16), 97 | nn.ReLU(True) 98 | ) 99 | 100 | self.layer2 = nn.Sequential( 101 | nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1, bias=False), 102 | nn.BatchNorm2d(32), 103 | nn.ReLU(True) 104 | ) 105 | 106 | self.layer3 = Tree(block, 32, 64, level=1, stride=1) 107 | self.layer4 = Tree(block, 64, 128, level=2, stride=2) 108 | self.layer5 = Tree(block, 128, 256, level=2, stride=2) 109 | self.layer6 = Tree(block, 256, 512, level=1, stride=2) 110 | self.linear = nn.Linear(512, num_classes) 111 | 112 | def forward(self, x): 113 | out = self.base(x) 114 | out = self.layer1(out) 115 | out = self.layer2(out) 116 | out = self.layer3(out) 117 | out = self.layer4(out) 118 | out = self.layer5(out) 119 | out = self.layer6(out) 120 | out = F.avg_pool2d(out, 4) 121 | out = out.view(out.size(0), -1) 122 | out = self.linear(out) 123 | return out 124 | 125 | 126 | def test(): 127 | net = DLA() 128 | print(net) 129 | x = torch.randn(1, 3, 32, 32) 130 | y = net(x) 131 | print(y.size()) 132 | 133 | 134 | if __name__ == '__main__': 135 | test() 136 | -------------------------------------------------------------------------------- /models/regnet.py: -------------------------------------------------------------------------------- 1 | '''RegNet in PyTorch. 2 | 3 | Paper: "Designing Network Design Spaces". 4 | 5 | Reference: https://github.com/keras-team/keras-applications/blob/master/keras_applications/efficientnet.py 6 | ''' 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | 11 | 12 | class SE(nn.Module): 13 | '''Squeeze-and-Excitation block.''' 14 | 15 | def __init__(self, in_planes, se_planes): 16 | super(SE, self).__init__() 17 | self.se1 = nn.Conv2d(in_planes, se_planes, kernel_size=1, bias=True) 18 | self.se2 = nn.Conv2d(se_planes, in_planes, kernel_size=1, bias=True) 19 | 20 | def forward(self, x): 21 | out = F.adaptive_avg_pool2d(x, (1, 1)) 22 | out = F.relu(self.se1(out)) 23 | out = self.se2(out).sigmoid() 24 | out = x * out 25 | return out 26 | 27 | 28 | class Block(nn.Module): 29 | def __init__(self, w_in, w_out, stride, group_width, bottleneck_ratio, se_ratio): 30 | super(Block, self).__init__() 31 | # 1x1 32 | w_b = int(round(w_out * bottleneck_ratio)) 33 | self.conv1 = nn.Conv2d(w_in, w_b, kernel_size=1, bias=False) 34 | self.bn1 = nn.BatchNorm2d(w_b) 35 | # 3x3 36 | num_groups = w_b // group_width 37 | self.conv2 = nn.Conv2d(w_b, w_b, kernel_size=3, 38 | stride=stride, padding=1, groups=num_groups, bias=False) 39 | self.bn2 = nn.BatchNorm2d(w_b) 40 | # se 41 | self.with_se = se_ratio > 0 42 | if self.with_se: 43 | w_se = int(round(w_in * se_ratio)) 44 | self.se = SE(w_b, w_se) 45 | # 1x1 46 | self.conv3 = nn.Conv2d(w_b, w_out, kernel_size=1, bias=False) 47 | self.bn3 = nn.BatchNorm2d(w_out) 48 | 49 | self.shortcut = nn.Sequential() 50 | if stride != 1 or w_in != w_out: 51 | self.shortcut = nn.Sequential( 52 | nn.Conv2d(w_in, w_out, 53 | kernel_size=1, stride=stride, bias=False), 54 | nn.BatchNorm2d(w_out) 55 | ) 56 | 57 | def forward(self, x): 58 | out = F.relu(self.bn1(self.conv1(x))) 59 | out = F.relu(self.bn2(self.conv2(out))) 60 | if self.with_se: 61 | out = self.se(out) 62 | out = self.bn3(self.conv3(out)) 63 | out += self.shortcut(x) 64 | out = F.relu(out) 65 | return out 66 | 67 | 68 | class RegNet(nn.Module): 69 | def __init__(self, cfg, num_classes=10): 70 | super(RegNet, self).__init__() 71 | self.cfg = cfg 72 | self.in_planes = 64 73 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, 74 | stride=1, padding=1, bias=False) 75 | self.bn1 = nn.BatchNorm2d(64) 76 | self.layer1 = self._make_layer(0) 77 | self.layer2 = self._make_layer(1) 78 | self.layer3 = self._make_layer(2) 79 | self.layer4 = self._make_layer(3) 80 | self.linear = nn.Linear(self.cfg['widths'][-1], num_classes) 81 | 82 | def _make_layer(self, idx): 83 | depth = self.cfg['depths'][idx] 84 | width = self.cfg['widths'][idx] 85 | stride = self.cfg['strides'][idx] 86 | group_width = self.cfg['group_width'] 87 | bottleneck_ratio = self.cfg['bottleneck_ratio'] 88 | se_ratio = self.cfg['se_ratio'] 89 | 90 | layers = [] 91 | for i in range(depth): 92 | s = stride if i == 0 else 1 93 | layers.append(Block(self.in_planes, width, 94 | s, group_width, bottleneck_ratio, se_ratio)) 95 | self.in_planes = width 96 | return nn.Sequential(*layers) 97 | 98 | def forward(self, x): 99 | out = F.relu(self.bn1(self.conv1(x))) 100 | out = self.layer1(out) 101 | out = self.layer2(out) 102 | out = self.layer3(out) 103 | out = self.layer4(out) 104 | out = F.adaptive_avg_pool2d(out, (1, 1)) 105 | out = out.view(out.size(0), -1) 106 | out = self.linear(out) 107 | return out 108 | 109 | 110 | def RegNetX_200MF(): 111 | cfg = { 112 | 'depths': [1, 1, 4, 7], 113 | 'widths': [24, 56, 152, 368], 114 | 'strides': [1, 1, 2, 2], 115 | 'group_width': 8, 116 | 'bottleneck_ratio': 1, 117 | 'se_ratio': 0, 118 | } 119 | return RegNet(cfg) 120 | 121 | 122 | def RegNetX_400MF(): 123 | cfg = { 124 | 'depths': [1, 2, 7, 12], 125 | 'widths': [32, 64, 160, 384], 126 | 'strides': [1, 1, 2, 2], 127 | 'group_width': 16, 128 | 'bottleneck_ratio': 1, 129 | 'se_ratio': 0, 130 | } 131 | return RegNet(cfg) 132 | 133 | 134 | def RegNetY_400MF(): 135 | cfg = { 136 | 'depths': [1, 2, 7, 12], 137 | 'widths': [32, 64, 160, 384], 138 | 'strides': [1, 1, 2, 2], 139 | 'group_width': 16, 140 | 'bottleneck_ratio': 1, 141 | 'se_ratio': 0.25, 142 | } 143 | return RegNet(cfg) 144 | 145 | 146 | def test(): 147 | net = RegNetX_200MF() 148 | print(net) 149 | x = torch.randn(2, 3, 32, 32) 150 | y = net(x) 151 | print(y.shape) 152 | 153 | 154 | if __name__ == '__main__': 155 | test() 156 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | '''Train CIFAR10 with PyTorch.''' 2 | import torch 3 | import torch.nn as nn 4 | import torch.optim as optim 5 | import torch.nn.functional as F 6 | import torch.backends.cudnn as cudnn 7 | 8 | import torchvision 9 | import torchvision.transforms as transforms 10 | 11 | import os 12 | import argparse 13 | 14 | from models import * 15 | from utils import progress_bar 16 | 17 | 18 | parser = argparse.ArgumentParser(description='PyTorch CIFAR10 Training') 19 | parser.add_argument('--lr', default=0.1, type=float, help='learning rate') 20 | parser.add_argument('--resume', '-r', action='store_true', 21 | help='resume from checkpoint') 22 | args = parser.parse_args() 23 | 24 | device = 'cuda' if torch.cuda.is_available() else 'cpu' 25 | best_acc = 0 # best test accuracy 26 | start_epoch = 0 # start from epoch 0 or last checkpoint epoch 27 | 28 | # Data 29 | print('==> Preparing data..') 30 | transform_train = transforms.Compose([ 31 | transforms.RandomCrop(32, padding=4), 32 | transforms.RandomHorizontalFlip(), 33 | transforms.ToTensor(), 34 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), 35 | ]) 36 | 37 | transform_test = transforms.Compose([ 38 | transforms.ToTensor(), 39 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), 40 | ]) 41 | 42 | trainset = torchvision.datasets.CIFAR10( 43 | root='./data', train=True, download=True, transform=transform_train) 44 | trainloader = torch.utils.data.DataLoader( 45 | trainset, batch_size=128, shuffle=True, num_workers=2) 46 | 47 | testset = torchvision.datasets.CIFAR10( 48 | root='./data', train=False, download=True, transform=transform_test) 49 | testloader = torch.utils.data.DataLoader( 50 | testset, batch_size=100, shuffle=False, num_workers=2) 51 | 52 | classes = ('plane', 'car', 'bird', 'cat', 'deer', 53 | 'dog', 'frog', 'horse', 'ship', 'truck') 54 | 55 | # Model 56 | print('==> Building model..') 57 | # net = VGG('VGG19') 58 | # net = ResNet18() 59 | # net = PreActResNet18() 60 | # net = GoogLeNet() 61 | # net = DenseNet121() 62 | # net = ResNeXt29_2x64d() 63 | # net = MobileNet() 64 | # net = MobileNetV2() 65 | # net = DPN92() 66 | # net = ShuffleNetG2() 67 | # net = SENet18() 68 | # net = ShuffleNetV2(1) 69 | # net = EfficientNetB0() 70 | # net = RegNetX_200MF() 71 | net = SimpleDLA() 72 | net = net.to(device) 73 | if device == 'cuda': 74 | net = torch.nn.DataParallel(net) 75 | cudnn.benchmark = True 76 | 77 | if args.resume: 78 | # Load checkpoint. 79 | print('==> Resuming from checkpoint..') 80 | assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!' 81 | checkpoint = torch.load('./checkpoint/ckpt.pth') 82 | net.load_state_dict(checkpoint['net']) 83 | best_acc = checkpoint['acc'] 84 | start_epoch = checkpoint['epoch'] 85 | 86 | criterion = nn.CrossEntropyLoss() 87 | optimizer = optim.SGD(net.parameters(), lr=args.lr, 88 | momentum=0.9, weight_decay=5e-4) 89 | scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=200) 90 | 91 | 92 | # Training 93 | def train(epoch): 94 | print('\nEpoch: %d' % epoch) 95 | net.train() 96 | train_loss = 0 97 | correct = 0 98 | total = 0 99 | for batch_idx, (inputs, targets) in enumerate(trainloader): 100 | inputs, targets = inputs.to(device), targets.to(device) 101 | optimizer.zero_grad() 102 | outputs = net(inputs) 103 | loss = criterion(outputs, targets) 104 | loss.backward() 105 | optimizer.step() 106 | 107 | train_loss += loss.item() 108 | _, predicted = outputs.max(1) 109 | total += targets.size(0) 110 | correct += predicted.eq(targets).sum().item() 111 | 112 | progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' 113 | % (train_loss/(batch_idx+1), 100.*correct/total, correct, total)) 114 | 115 | 116 | def test(epoch): 117 | global best_acc 118 | net.eval() 119 | test_loss = 0 120 | correct = 0 121 | total = 0 122 | with torch.no_grad(): 123 | for batch_idx, (inputs, targets) in enumerate(testloader): 124 | inputs, targets = inputs.to(device), targets.to(device) 125 | outputs = net(inputs) 126 | loss = criterion(outputs, targets) 127 | 128 | test_loss += loss.item() 129 | _, predicted = outputs.max(1) 130 | total += targets.size(0) 131 | correct += predicted.eq(targets).sum().item() 132 | 133 | progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' 134 | % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) 135 | 136 | # Save checkpoint. 137 | acc = 100.*correct/total 138 | if acc > best_acc: 139 | print('Saving..') 140 | state = { 141 | 'net': net.state_dict(), 142 | 'acc': acc, 143 | 'epoch': epoch, 144 | } 145 | if not os.path.isdir('checkpoint'): 146 | os.mkdir('checkpoint') 147 | torch.save(state, './checkpoint/ckpt.pth') 148 | best_acc = acc 149 | 150 | 151 | for epoch in range(start_epoch, start_epoch+200): 152 | train(epoch) 153 | test(epoch) 154 | scheduler.step() 155 | -------------------------------------------------------------------------------- /models/shufflenetv2.py: -------------------------------------------------------------------------------- 1 | '''ShuffleNetV2 in PyTorch. 2 | 3 | See the paper "ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design" for more details. 4 | ''' 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | 9 | 10 | class ShuffleBlock(nn.Module): 11 | def __init__(self, groups=2): 12 | super(ShuffleBlock, self).__init__() 13 | self.groups = groups 14 | 15 | def forward(self, x): 16 | '''Channel shuffle: [N,C,H,W] -> [N,g,C/g,H,W] -> [N,C/g,g,H,w] -> [N,C,H,W]''' 17 | N, C, H, W = x.size() 18 | g = self.groups 19 | return x.view(N, g, C//g, H, W).permute(0, 2, 1, 3, 4).reshape(N, C, H, W) 20 | 21 | 22 | class SplitBlock(nn.Module): 23 | def __init__(self, ratio): 24 | super(SplitBlock, self).__init__() 25 | self.ratio = ratio 26 | 27 | def forward(self, x): 28 | c = int(x.size(1) * self.ratio) 29 | return x[:, :c, :, :], x[:, c:, :, :] 30 | 31 | 32 | class BasicBlock(nn.Module): 33 | def __init__(self, in_channels, split_ratio=0.5): 34 | super(BasicBlock, self).__init__() 35 | self.split = SplitBlock(split_ratio) 36 | in_channels = int(in_channels * split_ratio) 37 | self.conv1 = nn.Conv2d(in_channels, in_channels, 38 | kernel_size=1, bias=False) 39 | self.bn1 = nn.BatchNorm2d(in_channels) 40 | self.conv2 = nn.Conv2d(in_channels, in_channels, 41 | kernel_size=3, stride=1, padding=1, groups=in_channels, bias=False) 42 | self.bn2 = nn.BatchNorm2d(in_channels) 43 | self.conv3 = nn.Conv2d(in_channels, in_channels, 44 | kernel_size=1, bias=False) 45 | self.bn3 = nn.BatchNorm2d(in_channels) 46 | self.shuffle = ShuffleBlock() 47 | 48 | def forward(self, x): 49 | x1, x2 = self.split(x) 50 | out = F.relu(self.bn1(self.conv1(x2))) 51 | out = self.bn2(self.conv2(out)) 52 | out = F.relu(self.bn3(self.conv3(out))) 53 | out = torch.cat([x1, out], 1) 54 | out = self.shuffle(out) 55 | return out 56 | 57 | 58 | class DownBlock(nn.Module): 59 | def __init__(self, in_channels, out_channels): 60 | super(DownBlock, self).__init__() 61 | mid_channels = out_channels // 2 62 | # left 63 | self.conv1 = nn.Conv2d(in_channels, in_channels, 64 | kernel_size=3, stride=2, padding=1, groups=in_channels, bias=False) 65 | self.bn1 = nn.BatchNorm2d(in_channels) 66 | self.conv2 = nn.Conv2d(in_channels, mid_channels, 67 | kernel_size=1, bias=False) 68 | self.bn2 = nn.BatchNorm2d(mid_channels) 69 | # right 70 | self.conv3 = nn.Conv2d(in_channels, mid_channels, 71 | kernel_size=1, bias=False) 72 | self.bn3 = nn.BatchNorm2d(mid_channels) 73 | self.conv4 = nn.Conv2d(mid_channels, mid_channels, 74 | kernel_size=3, stride=2, padding=1, groups=mid_channels, bias=False) 75 | self.bn4 = nn.BatchNorm2d(mid_channels) 76 | self.conv5 = nn.Conv2d(mid_channels, mid_channels, 77 | kernel_size=1, bias=False) 78 | self.bn5 = nn.BatchNorm2d(mid_channels) 79 | 80 | self.shuffle = ShuffleBlock() 81 | 82 | def forward(self, x): 83 | # left 84 | out1 = self.bn1(self.conv1(x)) 85 | out1 = F.relu(self.bn2(self.conv2(out1))) 86 | # right 87 | out2 = F.relu(self.bn3(self.conv3(x))) 88 | out2 = self.bn4(self.conv4(out2)) 89 | out2 = F.relu(self.bn5(self.conv5(out2))) 90 | # concat 91 | out = torch.cat([out1, out2], 1) 92 | out = self.shuffle(out) 93 | return out 94 | 95 | 96 | class ShuffleNetV2(nn.Module): 97 | def __init__(self, net_size): 98 | super(ShuffleNetV2, self).__init__() 99 | out_channels = configs[net_size]['out_channels'] 100 | num_blocks = configs[net_size]['num_blocks'] 101 | 102 | self.conv1 = nn.Conv2d(3, 24, kernel_size=3, 103 | stride=1, padding=1, bias=False) 104 | self.bn1 = nn.BatchNorm2d(24) 105 | self.in_channels = 24 106 | self.layer1 = self._make_layer(out_channels[0], num_blocks[0]) 107 | self.layer2 = self._make_layer(out_channels[1], num_blocks[1]) 108 | self.layer3 = self._make_layer(out_channels[2], num_blocks[2]) 109 | self.conv2 = nn.Conv2d(out_channels[2], out_channels[3], 110 | kernel_size=1, stride=1, padding=0, bias=False) 111 | self.bn2 = nn.BatchNorm2d(out_channels[3]) 112 | self.linear = nn.Linear(out_channels[3], 10) 113 | 114 | def _make_layer(self, out_channels, num_blocks): 115 | layers = [DownBlock(self.in_channels, out_channels)] 116 | for i in range(num_blocks): 117 | layers.append(BasicBlock(out_channels)) 118 | self.in_channels = out_channels 119 | return nn.Sequential(*layers) 120 | 121 | def forward(self, x): 122 | out = F.relu(self.bn1(self.conv1(x))) 123 | # out = F.max_pool2d(out, 3, stride=2, padding=1) 124 | out = self.layer1(out) 125 | out = self.layer2(out) 126 | out = self.layer3(out) 127 | out = F.relu(self.bn2(self.conv2(out))) 128 | out = F.avg_pool2d(out, 4) 129 | out = out.view(out.size(0), -1) 130 | out = self.linear(out) 131 | return out 132 | 133 | 134 | configs = { 135 | 0.5: { 136 | 'out_channels': (48, 96, 192, 1024), 137 | 'num_blocks': (3, 7, 3) 138 | }, 139 | 140 | 1: { 141 | 'out_channels': (116, 232, 464, 1024), 142 | 'num_blocks': (3, 7, 3) 143 | }, 144 | 1.5: { 145 | 'out_channels': (176, 352, 704, 1024), 146 | 'num_blocks': (3, 7, 3) 147 | }, 148 | 2: { 149 | 'out_channels': (224, 488, 976, 2048), 150 | 'num_blocks': (3, 7, 3) 151 | } 152 | } 153 | 154 | 155 | def test(): 156 | net = ShuffleNetV2(net_size=0.5) 157 | x = torch.randn(3, 3, 32, 32) 158 | y = net(x) 159 | print(y.shape) 160 | 161 | 162 | # test() 163 | -------------------------------------------------------------------------------- /models/efficientnet.py: -------------------------------------------------------------------------------- 1 | '''EfficientNet in PyTorch. 2 | 3 | Paper: "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks". 4 | 5 | Reference: https://github.com/keras-team/keras-applications/blob/master/keras_applications/efficientnet.py 6 | ''' 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | 11 | 12 | def swish(x): 13 | return x * x.sigmoid() 14 | 15 | 16 | def drop_connect(x, drop_ratio): 17 | keep_ratio = 1.0 - drop_ratio 18 | mask = torch.empty([x.shape[0], 1, 1, 1], dtype=x.dtype, device=x.device) 19 | mask.bernoulli_(keep_ratio) 20 | x.div_(keep_ratio) 21 | x.mul_(mask) 22 | return x 23 | 24 | 25 | class SE(nn.Module): 26 | '''Squeeze-and-Excitation block with Swish.''' 27 | 28 | def __init__(self, in_channels, se_channels): 29 | super(SE, self).__init__() 30 | self.se1 = nn.Conv2d(in_channels, se_channels, 31 | kernel_size=1, bias=True) 32 | self.se2 = nn.Conv2d(se_channels, in_channels, 33 | kernel_size=1, bias=True) 34 | 35 | def forward(self, x): 36 | out = F.adaptive_avg_pool2d(x, (1, 1)) 37 | out = swish(self.se1(out)) 38 | out = self.se2(out).sigmoid() 39 | out = x * out 40 | return out 41 | 42 | 43 | class Block(nn.Module): 44 | '''expansion + depthwise + pointwise + squeeze-excitation''' 45 | 46 | def __init__(self, 47 | in_channels, 48 | out_channels, 49 | kernel_size, 50 | stride, 51 | expand_ratio=1, 52 | se_ratio=0., 53 | drop_rate=0.): 54 | super(Block, self).__init__() 55 | self.stride = stride 56 | self.drop_rate = drop_rate 57 | self.expand_ratio = expand_ratio 58 | 59 | # Expansion 60 | channels = expand_ratio * in_channels 61 | self.conv1 = nn.Conv2d(in_channels, 62 | channels, 63 | kernel_size=1, 64 | stride=1, 65 | padding=0, 66 | bias=False) 67 | self.bn1 = nn.BatchNorm2d(channels) 68 | 69 | # Depthwise conv 70 | self.conv2 = nn.Conv2d(channels, 71 | channels, 72 | kernel_size=kernel_size, 73 | stride=stride, 74 | padding=(1 if kernel_size == 3 else 2), 75 | groups=channels, 76 | bias=False) 77 | self.bn2 = nn.BatchNorm2d(channels) 78 | 79 | # SE layers 80 | se_channels = int(in_channels * se_ratio) 81 | self.se = SE(channels, se_channels) 82 | 83 | # Output 84 | self.conv3 = nn.Conv2d(channels, 85 | out_channels, 86 | kernel_size=1, 87 | stride=1, 88 | padding=0, 89 | bias=False) 90 | self.bn3 = nn.BatchNorm2d(out_channels) 91 | 92 | # Skip connection if in and out shapes are the same (MV-V2 style) 93 | self.has_skip = (stride == 1) and (in_channels == out_channels) 94 | 95 | def forward(self, x): 96 | out = x if self.expand_ratio == 1 else swish(self.bn1(self.conv1(x))) 97 | out = swish(self.bn2(self.conv2(out))) 98 | out = self.se(out) 99 | out = self.bn3(self.conv3(out)) 100 | if self.has_skip: 101 | if self.training and self.drop_rate > 0: 102 | out = drop_connect(out, self.drop_rate) 103 | out = out + x 104 | return out 105 | 106 | 107 | class EfficientNet(nn.Module): 108 | def __init__(self, cfg, num_classes=10): 109 | super(EfficientNet, self).__init__() 110 | self.cfg = cfg 111 | self.conv1 = nn.Conv2d(3, 112 | 32, 113 | kernel_size=3, 114 | stride=1, 115 | padding=1, 116 | bias=False) 117 | self.bn1 = nn.BatchNorm2d(32) 118 | self.layers = self._make_layers(in_channels=32) 119 | self.linear = nn.Linear(cfg['out_channels'][-1], num_classes) 120 | 121 | def _make_layers(self, in_channels): 122 | layers = [] 123 | cfg = [self.cfg[k] for k in ['expansion', 'out_channels', 'num_blocks', 'kernel_size', 124 | 'stride']] 125 | b = 0 126 | blocks = sum(self.cfg['num_blocks']) 127 | for expansion, out_channels, num_blocks, kernel_size, stride in zip(*cfg): 128 | strides = [stride] + [1] * (num_blocks - 1) 129 | for stride in strides: 130 | drop_rate = self.cfg['drop_connect_rate'] * b / blocks 131 | layers.append( 132 | Block(in_channels, 133 | out_channels, 134 | kernel_size, 135 | stride, 136 | expansion, 137 | se_ratio=0.25, 138 | drop_rate=drop_rate)) 139 | in_channels = out_channels 140 | return nn.Sequential(*layers) 141 | 142 | def forward(self, x): 143 | out = swish(self.bn1(self.conv1(x))) 144 | out = self.layers(out) 145 | out = F.adaptive_avg_pool2d(out, 1) 146 | out = out.view(out.size(0), -1) 147 | dropout_rate = self.cfg['dropout_rate'] 148 | if self.training and dropout_rate > 0: 149 | out = F.dropout(out, p=dropout_rate) 150 | out = self.linear(out) 151 | return out 152 | 153 | 154 | def EfficientNetB0(): 155 | cfg = { 156 | 'num_blocks': [1, 2, 2, 3, 3, 4, 1], 157 | 'expansion': [1, 6, 6, 6, 6, 6, 6], 158 | 'out_channels': [16, 24, 40, 80, 112, 192, 320], 159 | 'kernel_size': [3, 3, 5, 3, 5, 5, 3], 160 | 'stride': [1, 2, 2, 2, 1, 2, 1], 161 | 'dropout_rate': 0.2, 162 | 'drop_connect_rate': 0.2, 163 | } 164 | return EfficientNet(cfg) 165 | 166 | 167 | def test(): 168 | net = EfficientNetB0() 169 | x = torch.randn(2, 3, 32, 32) 170 | y = net(x) 171 | print(y.shape) 172 | 173 | 174 | if __name__ == '__main__': 175 | test() 176 | --------------------------------------------------------------------------------