Pretrained MobilenetV2 depth 0.35 support

I’ve been working with MobilenetV2 model and noticed that there is a pretrained version for it in

model_urls = {
    'mobilenet_v2': '',

and it works great!

Now I want to train 0.35 depth, I was able to create model passing width_mult=0.35, as **kwargs . It works.
But when I try to create it with pretrained=True it crashes on size mismatch from checkpoint and model.

I found some pre-trained models with needed depth multiplier, but I’d like so guidance how to load them to instance of torchvision model.

If you guide me - would be happy to implement loading of this model it torch code, but first of all I need to find matching 0.35 weights.

Thank you for any recommendations.

Hi, searching for the same task. Did you find a solution to load the weights on 0.35 depth?

Not really, I use pytorch at work for research and tf for production, so I took tf model 0.35 depth. If I will decide to train it from scratch, I will update

1 Like