fdbtrs / pocketnet Goto Github PK

`backbone_teacher = iresnet100(num_features=cfg.embedding_size).to(local_rank)
try:
backbone_teacher_pth = os.path.join(cfg.teacher_pth, str((epoch + 1)*11372) + "backbone.pth")
backbone_teacher.load_state_dict(torch.load(backbone_teacher_pth, map_location=torch.device(local_rank)))

        if rank == 0:
            logging.info("backbone teacher loaded for epoch {} successfully!".format(epoch))
    except (FileNotFoundError, KeyError, IndexError, RuntimeError):
        logging.info("load teacher backbone for epoch {} init, failed!".format(epoch))
        break`

I can not find the teacher model.

would you mind releasing the training log of PocketNetS-128 (no KD)? I'm curious about how loss decrease when training without KD, comparing with that with KD.

Hello, in the code of architecture search, ie DART/searchs/search.py, I found that the difference between the training data and the validation data is whether there is random cropping, is that right?
49 # get dataset and meta info
50 input_size, input_channels, n_classes, train_data = dataset.get_train_dataset(cfg.root, cfg.dataset)
51 val_data = dataset.get_dataset_without_crop(cfg.root, cfg.dataset)
I am also studying the application of NAS in recognition tasks recently, but I am confused about how to set training data and validation data because I found that the conventional setting seems to be unable to be applied to the NAS. Can you tell me your understanding of this? Thanks a lot!!!

Firstly, I really admire all your works (PocketNet, QuantFace, MixFaceNets).

I have a concern, the inference time is extremely slow (even with GPU support), any suggestions on how to make it faster?

Thanks!

First, thanks for your impressive work. I have a small question about the MSEloss formula in your paper. why there is a '1 -' in the formula?