Pretrained weights

Pretrained encoder provenance

The first pretrained segmentation baseline uses SMP's U-Net with a ResNet18 encoder. Its encoder starts from the publisher's timm/resnet18.a1_in1k checkpoint, trained on ImageNet-1k. The U-Net decoder and one-channel segmentation head are new, randomly initialized layers. These encoder weights are not a pretrained PV detector and do not establish PV segmentation accuracy.

Exact weight source

Field	Pinned value
Model publisher	PyTorch Image Models (timm)
Model repository	timm/resnet18.a1_in1k
Revision	491b427b45c94c7fb0e78b5474cc919aff584bbf
Artifact	model.safetensors
Size	46,807,446 bytes
SHA-256	80c49dee3da4822c009c5a7fe591e9223c5a2cfcf95a4067ca4dfb5a7b89c612
Weight license declaration	apache-2.0 in the publisher's pinned model card

The revision-specific file metadata provides the artifact size and LFS SHA-256. The adapter checks both against the actual downloaded bytes. It uses this explicit publisher declaration for the weights, separately from the software licenses below.

Prepare a local cache

Install the baseline dependencies, then explicitly permit the pinned download:

uv sync --locked --extra baseline
uv run --locked --extra baseline python - <<'PY'
import json
from pvextent.models.resnet_unet import prepare_encoder_weights

provenance = prepare_encoder_weights(
    "data/pretrained/resnet18.a1_in1k", download=True
)
print(json.dumps(provenance, indent=2))
PY

The cache contains three original publisher files:

Local file	Purpose	SHA-256
model.safetensors	ImageNet-1k encoder/classifier tensors	80c49dee3da4822c009c5a7fe591e9223c5a2cfcf95a4067ca4dfb5a7b89c612
model-card.md	Pinned publisher description, license declaration, and citations	e96d7547dfaff6f0a0d74209819c461ec2d66556a9d6b059f1c2a50d413da5bb
config.json	Pinned architecture and preprocessing metadata	a15887d36094642dbffa237f2c7b66101d7f8ab704c45981b9fafb107cd99d28

All three files are size- and hash-checked. Downloads stream to temporary files and are published only after verification, with a 64 MiB per-file upper bound. Existing files with mismatched bytes fail instead of being replaced. Keeping the cache under the ignored data/ directory keeps weights outside Git.

Subsequent calls use download=False, the default, to verify the cache offline:

from pvextent.models.resnet_unet import (
    build_resnet_unet, normalize_rgb, prepare_encoder_weights,
)

provenance = prepare_encoder_weights("data/pretrained/resnet18.a1_in1k")
model = build_resnet_unet(provenance["path"])
# images: float32 RGB tensors [N,3,H,W], already scaled to [0,1].
logits = model(normalize_rgb(images))

Model construction never downloads weights. build_resnet_unet(None) explicitly constructs a random encoder for isolated tests or restoration of a complete local segmentation checkpoint; it is not the pretrained initialization path.

Architecture and preprocessing

The built-in model identifier is unet-resnet18-v1. Construction passes encoder_name="resnet18" and encoder_weights=None to SMP U-Net, preventing implicit downloads. The model uses encoder depth 5, decoder widths 256, 128, 64, 32, 16, batch normalization, nearest-neighbour decoder interpolation, three input channels, and one output channel with no activation. The complete model has 14,328,209 trainable parameters with SMP 0.5.0.

The adapter loads the verified safetensors bytes directly, checks finite floating-point tensors, and removes exactly fc.weight and fc.bias, the ImageNet classification head. All 120 remaining encoder entries must match the SMP ResNet18 encoder's keys, shapes, and dtypes. They are loaded strictly; missing or additional encoder tensors fail instead of being silently skipped. The decoder and segmentation head are not loaded from the classifier.

The pinned publisher configuration specifies RGB means [0.485, 0.456, 0.406] and standard deviations [0.229, 0.224, 0.225]. normalize_rgb applies those channel statistics to finite float32 RGB values in [0,1], preserving the grid and gradients. It does not resize, crop, change channel order, or modify the input tensor.

PVExtent retains the canonical 320 × 320 pixel, 0.1 m/pixel context for this experiment. This differs from the classifier's original image-size/crop settings: applying classification centre crops would remove segmentation context and require changing the masks. Model tests use dimensions divisible by 32; training-mode batch normalization also requires sufficient samples at the deepest resolution. The baseline's native 320-pixel crops and batch size 2 satisfy those constraints.

Software licenses and retained attribution

The dependency versions used for initial compatibility checks are recorded in uv.lock; exact versions also belong in experiment reports. The primary code license sources checked for this adapter are:

Dependency	Checked version	Code license source
Segmentation Models PyTorch	0.5.0	MIT, copyright 2019 Pavel Iakubovskii
Torchvision	0.29.1	BSD-3-Clause, copyright Soumith Chintala 2016
timm	1.0.30	Apache-2.0
safetensors	0.8.0	Apache-2.0

Dependency code licenses do not substitute for a checkpoint's own license declaration. Preserve the pinned publisher model card, its citations, applicable Apache-2.0 license text and notices, source weight hash, and the training-data provenance with any derived weight release. Preserve dependency copyright and license notices when redistributing their code. PVExtent's original adapter code remains copyright Valeford under Apache-2.0.

The checkpoint's ImageNet-1k training origin remains part of its provenance; this weight license declaration does not relicense the underlying image set. No ImageNet images or upstream weights are committed to this repository.