Spaces:
Running
Running
import torch | |
import torch.nn as nn | |
import torch.nn.functional as F | |
import torch | |
import torch.nn as nn | |
import torchvision.models as models | |
from torchvision.models import ResNet50_Weights | |
import numpy as np | |
from torchvision import transforms | |
class CustomResNet(nn.Module): | |
def __init__(self): | |
super(CustomResNet, self).__init__() | |
if torch.cuda.is_available(): | |
self.device = 'cuda' | |
elif torch.backends.mps.is_available(): | |
self.device = 'mps' | |
else: | |
self.device = 'cpu' | |
self.transform = transforms.Compose([ | |
transforms.Resize((224, 224)), | |
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) | |
]) | |
self.resnet = models.resnet50(weights=ResNet50_Weights.DEFAULT) | |
# Remove the final fully connected layer | |
self.resnet = nn.Sequential(*list(self.resnet.children())[:-1]) | |
# Define MLP | |
self.fc1 = nn.Linear(2 * (512 * 4), 256) | |
self.fc2 = nn.Linear(256, 1) | |
self.gradients = None | |
def activations_hook(self, grad): | |
self.gradients = grad | |
def forward(self, x1, x2, Location): | |
N = x1.shape[0] | |
x1 = self.transform(x1.to(torch.float).to(self.device)) | |
x2 = self.transform(x2.to(torch.float).to(self.device)) | |
Location = Location.to(torch.float).to(self.device) | |
# Process both images through the same ResNet | |
f1 = self.resnet[:8](x1) | |
h = f1.register_hook(self.activations_hook) | |
f1 = self.resnet[8:](f1) | |
f2 = self.resnet(x2) | |
# Flatten the features | |
f1 = f1.view(f1.size(0), -1) | |
f2 = f2.view(f2.size(0), -1) | |
f_ad = f1*Location[:,0].reshape(N,1) + f2*Location[:,1].reshape(N,1) | |
f_context = f1*(1-Location[:,0].reshape(N,1)) + f2*(1-Location[:,1].reshape(N,1)) | |
# Concatenate the features | |
combined = torch.cat((f_ad, f_context), dim=1) | |
# Pass through MLP | |
x = torch.relu(self.fc1(combined)) | |
x = self.fc2(x) | |
return x | |
# method for the gradient extraction | |
def get_activations_gradient(self): | |
return self.gradients | |
# method for the activation exctraction | |
def get_activations(self, x): | |
x = self.transform(x.to(torch.float).to(self.device)) | |
return self.resnet[:8](x) |