Hi, thanks for the great work on this repo!
I’ve been looking into the training pipeline of DCTDP, especially in TrainTask.train() and save_ckpt() implementations. From what I’ve read, the model uses a noise_model (e.g., NoisyActivation) during training, and it's wrapped with DistributedDataParallel. However, I couldn’t find any code that saves the noise_model's weights or its optimizer (noise_opt) in the checkpoint (e.g., within save_ckpt() or elsewhere).
Is the noise_model intentionally excluded from checkpoint saving?
Or is it an oversight and should be added (e.g., using noise_model.module.state_dict() and noise_opt.state_dict())?
Let me know if I missed something. Thanks in advance!
Hi, thanks for the great work on this repo!
I’ve been looking into the training pipeline of DCTDP, especially in TrainTask.train() and save_ckpt() implementations. From what I’ve read, the model uses a noise_model (e.g., NoisyActivation) during training, and it's wrapped with DistributedDataParallel. However, I couldn’t find any code that saves the noise_model's weights or its optimizer (noise_opt) in the checkpoint (e.g., within save_ckpt() or elsewhere).
Is the noise_model intentionally excluded from checkpoint saving?
Or is it an oversight and should be added (e.g., using noise_model.module.state_dict() and noise_opt.state_dict())?
Let me know if I missed something. Thanks in advance!