| 8 | |
| 9 | |
| 10 | class MYNET(nn.Module): |
| 11 | |
| 12 | def __init__(self, args, mode=None): |
| 13 | super().__init__() |
| 14 | |
| 15 | self.mode = mode |
| 16 | self.args = args |
| 17 | # self.num_features = 512 |
| 18 | if self.args.dataset in ['cifar100','manyshotcifar']: |
| 19 | self.encoder = resnet20() |
| 20 | self.num_features = 64 |
| 21 | if self.args.dataset in ['mini_imagenet','manyshotmini','imagenet100','imagenet1000', 'mini_imagenet_withpath']: |
| 22 | self.encoder = resnet18(False, args) # pretrained=False |
| 23 | self.num_features = 512 |
| 24 | if self.args.dataset in ['cub200','manyshotcub']: |
| 25 | self.encoder = resnet18(True, args) # pretrained=True follow TOPIC, models for cub is imagenet pre-trained. https://github.com/xyutao/fscil/issues/11#issuecomment-687548790 |
| 26 | self.num_features = 512 |
| 27 | self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) |
| 28 | |
| 29 | self.fc = nn.Linear(self.num_features, self.args.num_classes, bias=False) |
| 30 | |
| 31 | def forward_metric(self, x): |
| 32 | x = self.encode(x) |
| 33 | if 'cos' in self.mode: |
| 34 | x = F.linear(F.normalize(x, p=2, dim=-1), F.normalize(self.fc.weight, p=2, dim=-1)) |
| 35 | x = self.args.temperature * x |
| 36 | |
| 37 | elif 'dot' in self.mode: |
| 38 | x = self.fc(x) |
| 39 | x = self.args.temperature * x |
| 40 | return x |
| 41 | |
| 42 | def encode(self, x): |
| 43 | x = self.encoder(x) |
| 44 | x = F.adaptive_avg_pool2d(x, 1) |
| 45 | x = x.squeeze(-1).squeeze(-1) |
| 46 | return x |
| 47 | |
| 48 | def forward(self, input): |
| 49 | if self.mode != 'encoder': |
| 50 | input = self.forward_metric(input) |
| 51 | return input |
| 52 | elif self.mode == 'encoder': |
| 53 | input = self.encode(input) |
| 54 | return input |
| 55 | else: |
| 56 | raise ValueError('Unknown mode') |
| 57 | |
| 58 | def update_fc(self,dataloader,class_list,session): |
| 59 | for batch in dataloader: |
| 60 | data, label = [_.cuda() for _ in batch] |
| 61 | data=self.encode(data).detach() |
| 62 | |
| 63 | if self.args.not_data_init: |
| 64 | new_fc = nn.Parameter( |
| 65 | torch.rand(len(class_list), self.num_features, device="cuda"), |
| 66 | requires_grad=True) |
| 67 | nn.init.kaiming_uniform_(new_fc, a=math.sqrt(5)) |