Save weight test issue

[sove45]

I used train3d.py for training, saved the weights after training, and then loaded them. We found that regardless of which parameters I set as trainable, the network converged based on the training loss results. However, the results from the saved model during testing were always the same.

[sove45]

No matter what model I input, the saved model’s weights are exactly the same as the model right after initialization

[libalala]

args中的sam_ckpt也要设置为你保存下来的权重路径，光设置pretrain好像不行