ferrotorch-hub

Pretrained model registry, download, and caching for ferrotorch.

What it provides

• list_models -- browse all registered pretrained model architectures
• get_model_info -- query metadata (name, description, parameter count, weights format)
• load_pretrained -- load pretrained state dicts from the local cache
• download_weights -- fetch model weights from a URL and cache locally
• HubCache -- manage the on-disk cache directory (~/.cache/ferrotorch/hub)

Mirrors the workflow of torch.hub and torchvision.models with pretrained weight support.

Quick start

use ferrotorch_hub::{list_models, get_model_info, load_pretrained};

// Browse available models.
for model in list_models() {
    println!("{}: {} ({} params)", model.name, model.description, model.num_parameters);
}

// Load pretrained weights (requires cached weights on disk).
let state_dict = load_pretrained::<f32>("resnet50").unwrap();

Note: a handful of entries in the registry (e.g., the placeholder unet, maskrcnn_resnet50_fpn, and deeplabv3_resnet50 rows) ship with an all-zero weights_sha256 because no authoritative public mirror has been pinned yet. load_pretrained and download_weights deliberately fail-fast on those entries with InvalidArgument rather than silently skipping verification — see registry.rs for the per-entry rationale.

Part of ferrotorch

This crate is one component of the ferrotorch workspace. See the workspace README for full documentation.

License

MIT OR Apache-2.0