How to specify the weight of the downloaded teacher model without access to the Internet? about torchdistill

[yoshitomo-matsubara]

Hi @wongyufei,

You can specify the file path of checkpoint by editing ckpt in the yaml file.
Note that the object loaded from ckpt should be a dict that contains a key of model with model weights. (if ckpt file does not exist and pretrained: True, pretrained weights in torchvision will be used in case of image classification.)
e.g., you can save model weights compatible with torchdistill in the following way:

import torch
from torchvision.models import mobilenet_v2
ckpt = dict()
classifier = mobilenet_v2(pretrained=True)
ckpt['model'] = classifier.state_dict()
torch.save(ckpt, './test_ckpt.pt')

and you can replace ckpt entry with ckpt: './test_ckpt.pt' under student_model/teacher_model of the yaml file