Nestmodels for Linnet
Models / resnet-50

ResNet-50

The 50-layer residual network (v1.5) for ImageNet-1k: a 7x7 stem, four stages of bottleneck blocks 3, 4, 6, and 3 deep, and a linear classifier.

25.6M parametersresnetApache-2.0image-classificationvisionconvolutional
Hugging FaceGitHubarXiv Registry entry

The 50-layer residual network, fine-tuned on ImageNet-1k: a 7x7 stem at stride 2, max pooling, four stages of bottleneck blocks three, four, six, and three deep, spatial averaging, and a linear classifier. A bottleneck reduces to a quarter of its stage's width with a 1x1 convolution, convolves 3x3 there, and expands back with another 1x1; the first bottleneck of a stage also projects its shortcut with a 1x1 convolution (plus its own normalization) whenever the width or the resolution changes, which is true of every stage's first block since the width always changes there.

This is v1.5, the variant microsoft/resnet-50 publishes and the one most current ResNet-50 checkpoints use: the stride that halves the resolution sits on the stage's 3x3 convolution rather than its first 1x1 (the original v1 placement), so more of the stage's work happens at full resolution before the channel count grows. Only the first stage keeps stride 1 throughout, since max pooling has already halved the resolution once; the other three each halve it again on their first block.

Reuses resnet-18's ConvNorm (a convolution plus the affine batch normalization inference computes) and stem, adapted for std.nn.conv's convolution as a contraction over an index-built window. Stage is generic over the input and output width, the stride, and the depth, so all four stages share one block definition; the per-stage counts come from the checkpoint's depths config ([3, 4, 6, 3]).

Loading ​

python
from linnet import nest

model = nest.load("resnet-50", 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-50", backend="torch", numerics="fast", compile="inductor")
logits = model(tokens)

JAX ​

python
from linnet import nest

f = nest.load("resnet-50", backend="jax")          # StableHLO compiled by XLA
logits = f(tokens)

Flax NNX ​

python
from linnet import nest

model = nest.load("resnet-50", backend="nnx")
logits = model(tokens)

Command line ​

bash
python -m linnet.nest pull resnet-50
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