---
license: apache-2.0
---

This an example of a single Pytorch .pt2 model archive that packages [STAC Machine Learning Model](https://github.com/stac-extensions/mlm?tab=readme-ov-file#machine-learning-model-extension-specification)_ (MLM) metadata.

MLM metadata is stored as YAML in the pt2's `extra` properties. 

See https://docs.pytorch.org/docs/2.9/export/pt2_archive.html for more info on the pt2 archive spec.

See https://github.com/stac-extensions/mlm/blob/main/examples/torch/mlm-metadata.yaml for an example of MLM YAML metadata.

This model was exported with this script below. Download the original checkpoitn here: https://huggingface.co/torchgeo/ftw/blob/main/commercial/3-class/sentinel2_unet_effb3-5d591cbb.pth

```python
from pathlib import Path
import torch
import torchvision.transforms.v2 as T
from stac_model.torch.export import save
import segmentation_models_pytorch as smp


path = "sentinel2_unet_effb3-ed36f465.pth"
ckpt = torch.load(path, map_location="cpu", weights_only=False)
hparams = ckpt["hyper_parameters"]
state_dict = {k.replace("model.", ""): v for k, v in ckpt["state_dict"].items()}
del state_dict["criterion.weight"]
model = smp.Unet(
    encoder_name=hparams["backbone"],
    encoder_weights=None,
    in_channels=hparams["in_channels"],
    classes=hparams["num_classes"],
)
model.load_state_dict(state_dict, strict=True)

transforms = torch.nn.Sequential(
    torch.nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False),
    T.Normalize(mean=[0.0], std=[3000.0])
)
metadata_path = "mlm.yaml"
save(
    output_file=Path("model.pt2"),
    input_shape=[-1, hparams["in_channels"], -1, -1],
    model=model,
    transforms=transforms,
    metadata=metadata_path,
    device="cpu",
    dtype=torch.float32,
    aoti_compile_and_package=False
)

```