Thank you for sharing your excellent work and code. It's been very insightful, but I've encountered a couple of issues regarding the implementation that I hope you can help clarify:
Necessity of the Classifier in Figure 2: Is the Classifier shown in Figure 2 of your paper essential for the model to function effectively? Specifically, if I work exclusively with a tampered dataset, could the Classifier be omitted? I presume that the system might still perform adequately without it, but I would appreciate your insight on this matter.
Training with Only Tampered Images: In my experimental setup, I use only tampered images to train the model, fine-tuning it based on your casiav2_ep40_lr2.5e-4.pth. However, the loss behavior is puzzling; it decreases to 2.0794 and then plateaus, showing no further decrease. I have set LAMBDA_CLS to 0.0 to negate the Classifier's impact. What could be the possible reasons for this? I have tested it on two different datasets and observed consistent results.
Hi,
Thank you for sharing your excellent work and code. It's been very insightful, but I've encountered a couple of issues regarding the implementation that I hope you can help clarify:
Necessity of the Classifier in Figure 2: Is the Classifier shown in Figure 2 of your paper essential for the model to function effectively? Specifically, if I work exclusively with a tampered dataset, could the Classifier be omitted? I presume that the system might still perform adequately without it, but I would appreciate your insight on this matter.
Training with Only Tampered Images: In my experimental setup, I use only tampered images to train the model, fine-tuning it based on your
casiav2_ep40_lr2.5e-4.pth. However, the loss behavior is puzzling; it decreases to 2.0794 and then plateaus, showing no further decrease. I have setLAMBDA_CLSto 0.0 to negate the Classifier's impact. What could be the possible reasons for this? I have tested it on two different datasets and observed consistent results.