About Trainable parameters in ReferenceNet

[burnquiet]

In train stage 1 , code is:

for name, param in reference_unet.named_parameters():
    if "up_blocks.3" in name:
        param.requires_grad_(False)
    else:
        param.requires_grad_(True)

but up_blocks.3.resnets need grad (cuz "norm_hidden_states" is the normed resnet's output), should it be ? :

for name, param in reference_unet.named_parameters():
    if "up_blocks.3.attentions" in name:
        param.requires_grad_(False)
    else:
        param.requires_grad_(True)