| import torch | |
| from torch import nn | |
| from warnings import filterwarnings | |
| from torchvision.transforms import ToTensor, Resize, Normalize, Compose | |
| filterwarnings("ignore") | |
| DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") | |
| KERNEL_SIZE = (3,3) | |
| class VGGFACE(nn.Module): | |
| def __init__(self, *args, **kwargs) -> None: | |
| super().__init__(*args, **kwargs) | |
| self.conv1_1 = nn.Conv2d(3, 64, KERNEL_SIZE, 1, 1) | |
| self.conv1_2 = nn.Conv2d(64, 64, KERNEL_SIZE, 1, 1) | |
| self.conv2_1 = nn.Conv2d(64, 128, KERNEL_SIZE, 1, 1) | |
| self.conv2_2 = nn.Conv2d(128, 128, KERNEL_SIZE, 1, 1) | |
| self.conv3_1 = nn.Conv2d(128, 256, KERNEL_SIZE, 1, 1) | |
| self.conv3_2 = nn.Conv2d(256, 256, KERNEL_SIZE, 1, 1) | |
| self.conv3_3 = nn.Conv2d(256, 256, KERNEL_SIZE, 1, 1) | |
| self.conv4_1 = nn.Conv2d(256, 512, KERNEL_SIZE, 1, 1) | |
| self.conv4_2 = nn.Conv2d(512, 512, KERNEL_SIZE, 1, 1) | |
| self.conv4_3 = nn.Conv2d(512, 512, KERNEL_SIZE, 1, 1) | |
| self.conv5_1 = nn.Conv2d(512, 512, KERNEL_SIZE, 1, 1) | |
| self.conv5_2 = nn.Conv2d(512, 512, KERNEL_SIZE, 1, 1) | |
| self.conv5_3 = nn.Conv2d(512, 512, KERNEL_SIZE, 1, 1) | |
| self.fc6 = nn.Linear(49*512, 4096) | |
| self.fc7 = nn.Linear(4096, 4096) | |
| self.fc8 = nn.Linear(4096, 2622) | |
| self.relu = nn.ReLU() | |
| self.maxpool = nn.MaxPool2d(2) | |
| self._features = [ | |
| self.conv1_1, self.relu, | |
| self.conv1_2, self.relu, | |
| self.maxpool, | |
| self.conv2_1, self.relu, | |
| self.conv2_2, self.relu, | |
| self.maxpool, | |
| self.conv3_1, self.relu, | |
| self.conv3_2, self.relu, | |
| self.conv3_3, self.relu, | |
| self.maxpool, | |
| self.conv4_1, self.relu, | |
| self.conv4_2, self.relu, | |
| self.conv4_3, self.relu, | |
| self.maxpool, | |
| self.conv5_1, self.relu, | |
| self.conv5_2, self.relu, | |
| self.conv5_3, self.relu, | |
| self.maxpool, | |
| nn.Flatten(start_dim=0) | |
| ] | |
| self._classifier = [ | |
| self.fc6, self.relu, | |
| self.fc7, self.relu, | |
| self.fc8 | |
| ] | |
| self._embedder = [ | |
| self.conv1_1, self.relu, | |
| self.conv1_2, self.relu, | |
| self.maxpool, | |
| self.conv2_1, self.relu, | |
| self.conv2_2, self.relu, | |
| self.maxpool, | |
| self.conv3_1, self.relu, | |
| self.conv3_2, self.relu, | |
| self.conv3_3, self.relu, | |
| self.maxpool, | |
| self.conv4_1, self.relu, | |
| self.conv4_2, self.relu, | |
| self.conv4_3, self.relu, | |
| self.maxpool, | |
| self.conv5_1, self.relu, | |
| self.conv5_2, self.relu, | |
| self.conv5_3, self.relu, | |
| self.maxpool, | |
| nn.Flatten(start_dim=0), | |
| self.fc6, | |
| ] | |
| self.transform = Compose([ToTensor() ,Resize((224, 224)), Normalize(mean=(93.59396362304688/255, 104.76238250732422/255, 129.186279296875/255), std=(1, 1, 1))]) | |
| def features(self, x): | |
| x = self.transform(x) | |
| x = x.to(DEVICE) | |
| for layer in self._features: | |
| x = layer(x) | |
| return x | |
| def classifier(self, x): | |
| for layer in self._classifier: | |
| x = layer(x) | |
| return x | |
| def embedder(self, x): | |
| x = self.transform(x) | |
| x = x.to(DEVICE) | |
| for layer in self._embedder: | |
| x = layer(x) | |
| return x | |
| def forward(self, x:torch.Tensor): | |
| x = self.features(x) | |
| return self.classifier(x) | |
| def embeddings(self, x:torch.Tensor): | |
| return self.embedder(x).cpu().flatten().detach().numpy() | |
| __call__ = embeddings | |
| MODEL_FACE = VGGFACE() | |
| MODEL_FACE.load_state_dict(torch.load("models/vgg_face_dag.pth"), strict=True) | |
| MODEL_FACE.to(DEVICE) | |
| if __name__ == "__main__": | |
| print(MODEL_FACE.state_dict().keys()) |