load_state_dict via wandb artifact -> missing keys

[rowandempster]

Reproduction steps:

model = AVA_model(opt['model'])

ckpt = api.artifact('wato-action-classification/acar/models:v18').get_path('ckpt_11.pth.tar').download('/home/docker/catkin_ws/src/ACAR-Net/pretrained')

model.load_state_dict(torch.load(ckpt)['state_dict'])

yields Missing Keys: backbone.module.fast_conv1.weight.

Root cause: torch.load(ckpt)['state_dict'] contains keys in the format module.backbone.module.fast_conv1.weight (note the extra module. prepended).

Proposed resolution: Figure out why the state dict is being saved with the module. prepended and remove it.

[azhx]

For the purpose of multi-gpu training/inference, Acar always loads the model with DistributedDataParallel activated. I think this extra module. is a necessary abstraction for parallelization.

More info here: https://discuss.pytorch.org/t/missing-keys-unexpected-keys-in-state-dict-when-loading-self-trained-model/22379/2

I think the best solution for you is to simply remove the extra module. but you can also initialize the model like this

net = AVA_model(opt.model)
net.cuda()   
net = DistributedDataParallel(net, device_ids=[local_rank], broadcast_buffers=False)
#local_rank <= 1