MCPcopy Create free account
hub / github.com/TorchSSL/TorchSSL / __init__

Method __init__

models/nets/resnet50.py:130–193  ·  view source on GitHub ↗
(
            self,
            block: Type[Union[BasicBlock, Bottleneck]] = Bottleneck,
            layers: List[int] = [3, 4, 6, 3],
            n_class: int = 1000,
            zero_init_residual: bool = False,
            groups: int = 1,
            width_per_group: int = 64,
            replace_stride_with_dilation: Optional[List[bool]] = None,
            norm_layer: Optional[Callable[..., nn.Module]] = None,
            is_remix=False
    )

Source from the content-addressed store, hash-verified

128class ResNet50(nn.Module):
129
130 def __init__(
131 self,
132 block: Type[Union[BasicBlock, Bottleneck]] = Bottleneck,
133 layers: List[int] = [3, 4, 6, 3],
134 n_class: int = 1000,
135 zero_init_residual: bool = False,
136 groups: int = 1,
137 width_per_group: int = 64,
138 replace_stride_with_dilation: Optional[List[bool]] = None,
139 norm_layer: Optional[Callable[..., nn.Module]] = None,
140 is_remix=False
141 ) -> None:
142 super(ResNet50, self).__init__()
143 if norm_layer is None:
144 norm_layer = nn.BatchNorm2d
145 self._norm_layer = norm_layer
146
147 self.inplanes = 64
148 self.dilation = 1
149 if replace_stride_with_dilation is None:
150 # each element in the tuple indicates if we should replace
151 # the 2x2 stride with a dilated convolution instead
152 replace_stride_with_dilation = [False, False, False]
153 if len(replace_stride_with_dilation) != 3:
154 raise ValueError("replace_stride_with_dilation should be None "
155 "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
156 self.groups = groups
157 self.base_width = width_per_group
158 self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
159 bias=False)
160 self.bn1 = norm_layer(self.inplanes)
161 self.relu = nn.ReLU(inplace=True)
162 self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
163 self.layer1 = self._make_layer(block, 64, layers[0])
164 self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
165 dilate=replace_stride_with_dilation[0])
166 self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
167 dilate=replace_stride_with_dilation[1])
168 self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
169 dilate=replace_stride_with_dilation[2])
170 self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
171 self.fc = nn.Linear(512 * block.expansion, n_class)
172
173 # rot_classifier for Remix Match
174 self.is_remix = is_remix
175 if is_remix:
176 self.rot_classifier = nn.Linear(2048, 4)
177
178 for m in self.modules():
179 if isinstance(m, nn.Conv2d):
180 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
181 elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
182 nn.init.constant_(m.weight, 1)
183 nn.init.constant_(m.bias, 0)
184
185 # Zero-initialize the last BN in each residual branch,
186 # so that the residual branch starts with zeros, and each residual block behaves like an identity.
187 # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677

Callers 2

__init__Method · 0.45
__init__Method · 0.45

Calls 1

_make_layerMethod · 0.95

Tested by

no test coverage detected