How was the accelerator_path checkpoint trained?

[j-rooroo]

Dear Authors,

Thanks for your excellent work and for releasing the code.

In train.py, accelerator_path is loaded after expanding conv_in to 8 channels and before expanding conv_out to 5 channels.

So I understand that accelerator_path should be a SAR-conditioned U-Net checkpoint with 8 input channels and 4 output channels, before adding the confidence channel.

Could you clarify how this checkpoint was trained?

Was it trained as a first stage with standard diffusion MSE loss on the same SAR/EO training split? Is there a script/config or checkpoint for reproducing this stage?

Best regards,

[ZhiPWei]

Dear Authors,

Thanks for your excellent work and for releasing the code.

In train.py, accelerator_path is loaded after expanding conv_in to 8 channels and before expanding conv_out to 5 channels.

So I understand that accelerator_path should be a SAR-conditioned U-Net checkpoint with 8 input channels and 4 output channels, before adding the confidence channel.

Could you clarify how this checkpoint was trained?

Was it trained as a first stage with standard diffusion MSE loss on the same SAR/EO training split? Is there a script/config or checkpoint for reproducing this stage?

Best regards,

Hello，I meet the same problem. How do you solve it?