Nestmodels for Linnet
Models / resnet-18

ResNet-18

The 18-layer residual network for ImageNet-1k: a 7x7 stem, four stages of two residual blocks, and a linear classifier.

11.7M parametersresnetApache-2.0image-classificationvisionconvolutional
Hugging FaceGitHubarXiv Registry entry

The 18-layer residual network, fine-tuned on ImageNet-1k: a 7x7 stem at stride 2, max pooling, four stages of two residual blocks each, spatial averaging, and a linear classifier. Widths double and resolution halves at each stage, so the first block of a stage projects its shortcut with a 1x1 convolution.

This is the first convolutional model in Nest, and the first to use std.nn.conv. A convolution in Linnet is a contraction over a window:

linnet
let mixed[b, co, oh, ow] = sum[ci, kh, kw]
    padded[b, ci, rows[oh, kh], columns[ow, kw]] * weight[co, ci, kh, kw]

rows and columns are index tensors built from iota, so the window arithmetic is ordinary indexing that the checker verifies. That body is the normative meaning; a backend with F.conv2d (or Conv in ONNX) is free to call it instead, and the compiler recovers the stride and padding from the shapes to do so.

Batch normalization appears as what inference actually computes — an affine transform along the channel axis from the statistics the checkpoint stores — so running_mean and running_var bind like any other parameter.

Loading ​

python
from linnet import nest

model = nest.load("resnet-18", backend="torch")
logits = model(images)      # Tensor[B, 3, 224, 224; f32] -> Tensor[B, 1000; f32]

Images are ImageNet-normalized in [B, C, H, W]; preprocessor_config.json in the Hub repository has the mean and standard deviation. Other input sizes work as long as both sides are multiples of 32, which the block's where clause states.

Entries ​

Entry
forward<B>(image)class logits

Provenance ​

Backends ​

The same source and checkpoint in each framework; numerics and compile are the loaders' options.

PyTorch ​

python
from linnet import nest

model = nest.load("resnet-18", backend="torch", numerics="fast", compile="inductor")
logits = model(tokens)

JAX ​

python
from linnet import nest

f = nest.load("resnet-18", backend="jax")          # StableHLO compiled by XLA
logits = f(tokens)

Flax NNX ​

python
from linnet import nest

model = nest.load("resnet-18", backend="nnx")
logits = model(tokens)

Command line ​

bash
python -m linnet.nest pull resnet-18
linnet stablehlo --root Model --entry forward --bind Height=224 --bind Width=224 --bind Classes=1000 --bind T=f32 --bind B=1 <model dir>/src/lib.linnet