Using the downloaded weights to predict images results in noise.

[Fngg]

Hello, thank you for your outstanding work! When I use the weights provided in this code repository, specifically the 200-epoch and 2000-epoch Simvp_gSTA or Simvp_Incepu weights trained on the mmnist dataset, to make predictions on the mmnist dataset, I find that the predicted images contain many noise points, as shown in the image below (the model weights used are the 2000-epoch Simvp_gSTA weights, which were downloaded from this repository). What could be the reason for this? Looking forward to your reply.