Change download location of create_model

[sachinruk]

Is there a way to change the download location of a model when using create_model. If I use checkpoint_path it expects the model to already be there.

[rwightman]

@sachinruk in the latest timm version you can do this

rn50_cfg = timm.models.get_pretrained_cfg('resnet50')
rn50_cfg['url'] = 'https://mycustom/checkpoint.pth'

model = timm.create_model('resnet50', pretrained_cfg=rn50_cfg, pretrained=True)

as per #1323, you can also load local checkpoints as if they are pretrained (change the head, use features_only, etc)

rn50_cfg = timm.models.get_pretrained_cfg('resnet50')
rn50_cfg.pop('url')
rn50_cfg['file'] = '/blah/checkpoint.pth'

model = timm.create_model('resnet50', pretrained_cfg=rn50_cfg, pretrained=True, num_classes=10)

I'll working a way to make these two options even easier by detecting url or file paths as an overload argument, but for now must be done by modifing and passing whoe pretrained_cfg in...

[rwightman]

or do you mean the location on your disk where it goes? I just follow pytorch behaviour which uses the cache directory for hub checkpoints, see https://github.com/pytorch/pytorch/blob/master/torch/hub.py#L325-L341

so you could use TORCH_HOME env var to move the download cache dir

I don't currently have any args to override the download destination dir per use though

[sachinruk]

The latter. Ta! Will set the env variable.

Because I'm running this on k8s pod our platform team has got specific paths where we can write models to, so for larger models the job kept crashing.