OpenSkillIndex← back to search
claude-code

AutoResearchClaw cv-classification

Best practices for image classification tasks. Use when working on CIFAR, ImageNet, or other classification benchmarks.

install
source · Clone the upstream repo
git clone https://github.com/aiming-lab/AutoResearchClaw
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/aiming-lab/AutoResearchClaw "$T" && mkdir -p ~/.claude/skills && cp -r "$T/researchclaw/skills/builtin/domain/cv-classification" ~/.claude/skills/aiming-lab-autoresearchclaw-cv-classification && rm -rf "$T"
manifest: researchclaw/skills/builtin/domain/cv-classification/SKILL.md
source content

Image Classification Best Practice

Architecture selection:

  • Small scale (CIFAR-10/100): ResNet-18/34, WideResNet, Simple ViT
  • Medium scale: ResNet-50, EfficientNet-B0/B1, DeiT-Small
  • Large scale: ViT-B/16, ConvNeXt, Swin Transformer

Training recipe:

  • Optimizer: AdamW (lr=1e-3 to 3e-4) or SGD (lr=0.1 with cosine decay)
  • Weight decay: 0.01-0.1 for AdamW, 5e-4 for SGD
  • Data augmentation: RandomCrop, RandomHorizontalFlip, Cutout/CutMix
  • Warmup: 5-10 epochs linear warmup for transformers
  • Batch size: 128-256 for CNNs, 512-1024 for ViTs (if memory allows)

Standard benchmarks:

  • CIFAR-10: ~96% (ResNet-18), ~97% (WideResNet)
  • CIFAR-100: ~80% (ResNet-18), ~84% (WideResNet)
  • ImageNet: ~76% (ResNet-50), ~81% (ViT-B/16)