Instantiating VGG16 with state dictionary without downloading weights

vision
1

I do have a pth file for vgg16 as vgg16-397923af.pth. I like to instantiate a new vgg16 model with exact same weights in vgg16-397923af.pth file. vgg16 = models.vgg16() always tries to download weights first, however I don’t want to use any downloads just a reference to the pth file. Is this possible?

2

That’s not the case as the default argument will randomly initialize the parameters:

model = models.vgg16()
# no output
model = models.vgg16(pretrained=True)
# Downloading: "https://download.pytorch.org/models/vgg16-397923af.pth" to ...

and will store it in ~/.cache/torch/hub/checkpoints by default.

3

Thanks ptrblck,
when model = models.vgg16() it download weights, but when model = models.vgg16(pretrained=False) it randomly initializes the parameters.

pip list | grep torch
torch 2.0.1
torchaudio 2.0.2
torchvision 0.15.2

4

It’s not doing it for me as mentioned in my previous post and also in the docs:

weights (VGG16_Weights, optional) – The pretrained weights to use. See VGG16_Weights below for more details, and possible values. By default, no pre-trained weights are used.

so you would need to come up with a way to reproduce the issue.

5

Well you are correct. Looks like I was absent minded when posting this. It is great to have you here ptrblck, rapidly responding to everything since from the beginning.

6

Good to hear, it’s indeed not the case as I was struggling to find any way this would be possible. :slight_smile: