Add a feature registry for models

[lorenzoh]

Implements #246.

PR Checklist

• [x] Tests are added
• [x] Documentation (this is an internal change for now, docs will be added in follow-up when functionality is made available and used in domain package)

Usage examples

From #269:

using FastAI: models
# loading this adds the models to registry
using FastVision

# Load original model, 1000 output classes, no weights (`ResNet(18)`):
load(models()["metalhead/resnet18"]);

# Load original model, 1000 output classes, with weights (`ResNet(18), pretrain=true`):
load(models()["metalhead/resnet18"], pretrained = true);

# Load only backbone, without weights:
load(models()["metalhead/resnet18"], variant = "backbone");

# Load only backbone, with weights:
load(models()["metalhead/resnet18"], pretrained = true, variant = "backbone");

# Load model for task, adapting layers as necessary:
task = ImageClassificationSingle((256, 256), 1:5, C = Gray{N0f8}) # input with 1 color channel, 5 classes
load(models()["metalhead/resnet18"], input = task.blocks.x, output = task.blocks.y)
# Also works with pretrained weights
load(models()["metalhead/resnet18"], pretrained = true, input = task.blocks.x, output = task.blocks.y)

# Correct variants are selected automatically given the blocks:
load(models()["metalhead/resnet18"], output = FastAI.ConvFeatures)  # uses backbone variant

# Support for multiple checkpoints, selectable by name:
load(models()["metalhead/resnet18"], checkpoint = "imagenet1k")

Docs

The proposed interface is well-described by the registry description, pasted below:

A FeatureRegistry for models. Allows you to find and load models for various learning tasks using a unified interface. Call models() to see a table view of available models:

using FastAI
models()

Which models are available depends on the loaded packages. For example, FastVision.jl adds vision models from Metalhead to the registry. Index the registry with a model ID to get more information about that model:

using FastAI: models
using FastVision  # loading the package extends the list of available models

models()["metalhead/resnet18"]

If you've selected a model, call load to then instantiate a model:

model = load("metalhead/resnet18")

By default, load loads a default version of the model without any pretrained weights.

load(model) also accepts keyword arguments that allow you to specify variants of the model and weight checkpoints that should be loaded.

Loading a checkpoint of pretrained weights:

• load(entry; pretrained = true): Use any pretrained weights, if they are available.
• load(entry; checkpoint = "checkpoint-name"): Use the weights with given name. See entry.checkpoints for available checkpoints (if any).
• load(entry; pretrained = false): Don't use pretrained weights

Loading a model variant for a specific task:

• load(entry; input = ImageTensor, output = OneHotLabel): Load a model variant matching an input and output block.
• load(entry; variant = "backbone"): Load a model variant by name. Seeentry.variants` for available variants.

[github-actions[bot]]

A documentation preview has been successfully built, view it here: Documentation preview PR-267

[lorenzoh]

@darsnack @theabhirath would love to get feedback on the API. Anything unclear or an important feature missing, let me know!