Vision: pretrained models for batch normalized version of vgg

[bodokaiser]

According to here we should be able to reuse the retrained weight from e.g. vgg16 and apply them to vgg16bn we will just need to write a small conversion (as order in features differs due to additional nn.BatchNorm2d layers).

Do you agree? Are PR welcome on this?

[colesbury]

I don't think you can take the weights from the model without batch norm and stick them in the model with batch normalization. The activation statistics will almost certainly be wrong at each batch-normalized layer if someone tries fine-tuning the model or otherwise continue training.

[bodokaiser]

Could we then also offer pretained weights for the batch normalized versions?

Depending on how you obtain them anyway I could also look if I get a caffemodel converted to pytorch state dict.

[wangg12]

+1 for the request.