Add ResNet, AlexNet, and VGG model definitions and model zoo

[colesbury]

You can create a ResNet-18 model pre-trained on ImageNet with:

import torchvision.models
resnet = torchvision.models.resnet18(pretrained=True)

[fmassa]

About the padding in alexnet, if I remember properly the original network does not contain padding.
I saw a padded version of alexnet in Fast R-CNN paper, but padding was added even to the pooling layers.
Is it something we want to add here?

[fmassa]

I'm wondering, it might be better at some point to move the model_zoo.py to a different place.
The only part specific to vision is the model names and paths, and whenever we want to support models from other domains, we will need such a file.

[soumith]

good idea. torch.utils?

[colesbury]

torch.zoo?

[colesbury]

I also think the model zoo is a work-in-progress. There's still some unanswered questions and things I'd like to have:

1. We don't have a good way of specifying how inputs should be transformed. For example, the ImageNet trained models expect images that have zero mean and unit std.
2. It would be nice for users to be able to distribute models that aren't included directly as part of PyTorch. For example, you might want to do zoo.load("colesbury/resnet.pth"). This could be tied to GitHub, for example
3. For (2) to work well, we'd probably need a good way to distribute model definitions (Python code) as well as the weights

[apaszke]

I love the idea of using github names! I think github supports git lfs, so we only need to standardize the structure of the repo and it should be good.

We could just clone the repo in some local directory where it would be cached, and then load it from there.

[colesbury]

Updated to use torch.utils.model_zoo

[normster]

Hi @colesbury, do you remember any more details on how these models (specifically resnet18) were trained? i.e. which year of ImageNet, final train/val accuracy, batch size.

Thank you!

[colesbury]

The ResNet models were trained on ILSVRC2012. They were trained with the default options using https://github.com/pytorch/examples/tree/master/imagenet. (batch size 256, 90 epochs, lr=0.1 decay by 10 every 30 epochs). I think I used 4 GPUs, but it might have been 8.

Accuracy on the validation set is in the table here:
https://pytorch.org/docs/stable/torchvision/models.html

You can verify the accuracy on the validation set by running the example imagenet script with:

python main.py -a resnet18 --evaluate --pretrained [path to imagenet]

You can compute the accuracy on the training set by swapping your train and val directories and running the above command.