Bases: BaseBackbone
ResNet18 backbone with the number of features as an arguments. For nf=20, the ResNet has 1/3 of the features
of the original. Code adapted from:
1. https://github.com/facebookresearch/GradientEpisodicMemory
2. https://worksheets.codalab.org/rest/bundles/0xaf60b5ed6a4a44c89d296aae9bc6a0f1/contents/blob/models.py
Source code in sequel/backbones/pytorch/resnet.py
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91 | class ResNet(BaseBackbone):
"""ResNet18 backbone with the number of features as an arguments. For `nf=20`, the ResNet has 1/3 of the features
of the original. Code adapted from:
1. https://github.com/facebookresearch/GradientEpisodicMemory
2. https://worksheets.codalab.org/rest/bundles/0xaf60b5ed6a4a44c89d296aae9bc6a0f1/contents/blob/models.py
"""
def __init__(self, block, num_blocks, num_classes, nf=20, *args, **kwargs):
super().__init__(*args, **kwargs)
self.encoder = ResNetEncoder(block, num_blocks, num_classes, nf)
self.classifier = nn.Linear(nf * 8 * block.expansion, num_classes)
def forward(self, inp: torch.Tensor, head_ids: Optional[Iterable] = None):
out = self.encoder(inp)
out = self.classifier(out)
if self.multihead:
out = self.select_output_head(out, head_ids)
return out
|