Probing experiments

[odunkel]

Hello,

thanks for the very interesting work 'Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing'.
Do you also plan to release the code, trained classifiers etc. for performing the probing analysis?

Thanks for your support in advance.

[Bingyan-Liu]

Thank you for your attention. Unfortunately, for the Probing experiments, we validated on the test data directly after each training session without saving the trained classifier. The classifier is a two-layer MLP as follows, which can be easily trained using torch once the data is prepared.

class Net(torch.nn.Module):
def init(self, input_size: int = 64 * 64, num_classes: int = 10):
super(Net,self).init()
self.input_size = input_size
self.hidden_size = max(input_size // 8, 2048)
self.mlp = torch.nn.Sequential(
torch.nn.Flatten(),
torch.nn.Dropout(0.1),
torch.nn.Linear(in_features=self.input_size, out_features=self.hidden_size),
torch.nn.Tanh(),
# torch.nn.ReLU(),
torch.nn.Linear(in_features=self.hidden_size, out_features=num_classes)
)
def forward(self, x):
output = self.mlp(x)
return output