---

name: resnest description: "Use ResNeSt split-attention CNN models across PyTorch, optional MXNet Gluon, and optional Detectron2 backbones with safe install checks, model routing, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

ResNeSt Repo Skill

Use this skill when a task names ResNeSt or asks for split-attention ResNet-style CNNs, PyTorch classification factories, Torch Hub loading, optional MXNet Gluon models, or optional Detectron2 ResNeSt/FPN backbones.

ResNeSt is a Python package centered on ImageNet classification backbones and downstream detection/segmentation integration. The safest verified core path is PyTorch model construction with pretrained=False; pretrained weights, ImageNet/COCO metrics, MXNet Gluon, Horovod, and Detectron2 are optional/heavyweight surfaces that need the user's environment and data.

Start with an install/import check

For ordinary package use:

pip install resnest --pre
python - <<'PY'
from resnest.torch import resnest50
model = resnest50(pretrained=False)
print(model.__class__.__name__)
PY

For a fuller offline-first package probe, run the bundled root helper:

python scripts/check_resnest_install.py --model resnest50 --image-size 64

The helper does not download weights or train models by default. It reports optional Gluon and Detectron2 availability without making those backends required.

Route by workflow

User intent	Read next	Why
Load ResNeSt in PyTorch, use Torch Hub, run a no-pretrained inference smoke, inspect SplAtConv2d, or interpret PyTorch ImageNet training configs	sub-skills/pytorch-models/SKILL.md	This is the verified core surface and owns PyTorch APIs, model names, training config notes, and PyTorch troubleshooting.
Use MXNet Gluon builders, get_model, Gluon pretrained .params cache behavior, raw/RecordIO ImageNet validation, or Horovod Gluon training notes	sub-skills/gluon-models/SKILL.md	Gluon is optional and needs a compatible MXNet stack; the sub-skill keeps optional-backend guidance separate from PyTorch.
Register ResNeSt inside Detectron2, choose COCO Faster/Mask/Cascade/Panoptic configs, validate build_resnest_fpn_backbone, or debug SyncBN/DCN/COCO issues	sub-skills/detectron2-backbones/SKILL.md	Detectron2 is optional and heavy; this sub-skill owns config extension, backbone builders, COCO recipe catalog, and safe config probes.
Decide which model/depth/backend is appropriate before entering a backend-specific route	references/model-overview.md	Shared catalog of core and fast model variants, published crop sizes, and backend/pretrained notes.
Fix import, optional dependency, pretrained download/cache, dataset, or backend setup failures	references/troubleshooting.md	Cross-cutting failure map before using a backend-specific troubleshooting reference.
Check whether this skill matches the current repository revision	references/repo-provenance.md	Source commit, package version, evidence paths, and refresh guidance.

Core model names

Core classification factories: resnest50, resnest101, resnest200, resnest269.

Fast ablation factories: resnest50_fast_1s1x64d, resnest50_fast_2s1x64d, resnest50_fast_4s1x64d, resnest50_fast_1s2x40d, resnest50_fast_2s2x40d, resnest50_fast_4s2x40d, resnest50_fast_1s4x24d.

All public PyTorch and Gluon factories accept the pattern pretrained=False plus backend-specific root/ctx and **kwargs. Keep pretrained=False for offline or smoke-test work. Use pretrained=True only when network/cache behavior and ImageNet-1000 classifier shape are acceptable.

Backend boundaries

• PyTorch classification is the required core path. CPU fully validates package import, factory construction, and tiny tensor forwards.
• CUDA is useful for real training and throughput, but it is not required for the core package smoke unless the user explicitly asks for GPU behavior.
• Gluon/MXNet is optional. Do not install or claim Gluon runtime support until a compatible MXNet wheel imports in the user's environment.
• Detectron2 is optional. Config probing is lightweight; real train/eval needs a Detectron2 build compatible with PyTorch/CUDA, COCO data, and often SyncBN/DCN operators.
• ImageNet and COCO preparation are large-data workflows. Do not run dataset download/extraction or full validation/training as a default check.

Safe operating pattern

1. Run the root helper or PyTorch tiny inference helper with pretrained=False.
2. Use the model overview to choose depth, crop size, and backend.
3. Enter the backend-specific sub-skill for concrete APIs, commands, config fields, and troubleshooting.
4. Escalate to pretrained downloads, ImageNet/COCO validation, CUDA, Horovod, or Detectron2 training only after the user confirms data, hardware, runtime, and time budget.

What this skill deliberately avoids

• It does not bundle full ImageNet or COCO download/extraction launchers.
• It does not make optional Gluon or Detectron2 dependencies part of the minimum install path.
• It does not claim reproduced paper or benchmark metrics from tiny smoke checks.
• It does not require the original ResNeSt repository checkout at runtime; references and scripts here are distilled or adapted for the skill tree.