Migrating from Metalhead 0.9

Metalhead 0.10 replaces the hand-written models with ports of timm models, built with create_model. The old API still works in 0.10 and will be removed in 0.11. The deprecated API covers:

• the model constructors (ResNet(50), ViT(:base), ...) and their lowercase builders (Metalhead.resnet, ...);
• backbone, classifier and Metalhead.loadpretrain!;
• the Metalhead.Layers module;
• pretrain = true, and the weights it loads from the FluxML organisation on the Hub.

The model constructors, backbone, classifier and loadpretrain! warn once per session, naming their replacement. The lowercase builders and Metalhead.Layers, which the constructors use, don't warn.

What changes

The input layout (width × height × channels × batch) is unchanged. Models are now returned in test mode.

The keyword arguments are timm's, so they differ per family. The docstrings of the old constructors don't apply to the new models: look up the timm model instead. The models also have a different structure (the PyTorch one), so code that indexes into a model, like model.layers[1][2], needs updating.

Pretrained weights

The pretrained weights of 0.9 came from torchvision. timm hosts the same weights, as <model>.tv_in1k (torchvision's IMAGENET1K_V1) and <model>.tv2_in1k (IMAGENET1K_V2). The table gives the repository with the version that pretrain = true loaded, whose weights create_model loads. For ResNet-50, the outputs of the two models are bitwise identical.

timm also has newer checkpoints of most of these architectures, usually more accurate. See the model list.

Architectures

These architectures are not ported yet: ConvNeXt(:xlarge), ViT(:huge), ViT(:giant) and ViT(:gigantic).