---

name: classification description: "Image classification workflows for InternImage backbones: data layouts, configs, train/eval/throughput, DeepSpeed/Accelerate, feature extraction, Hugging Face, and export routing." disable-model-invocation: true metadata: disco-role: operating license: MIT

InternImage classification

Use this sub-skill when the user is working with InternImage image classification: ImageNet or iNaturalist data layout, model/config selection, evaluation, training, throughput checks, DeepSpeed or Accelerate launch planning, intermediate feature extraction, or Hugging Face Transformers inference/conversion.

Routing

• For standard classification runs, first gather: task (eval, train, throughput, extract features, Transformers), InternImage checkout location, config label, checkpoint or pretrained weight, data root, GPU count, launcher style, output directory, and any YACS overrides.
• Use the bundled command builder instead of rewriting upstream shell scripts by hand: python scripts/build_classification_command.py --help
• Put long command recipes, data layouts, and feature-extraction details in references/workflows.md.
• Put YACS config defaults, config-family selection, model-builder arguments, and override keys in references/configuration.md.
• Put Transformers model IDs, trust_remote_code usage, and conversion notes in references/huggingface.md.
• Put classification-specific failure handling in references/troubleshooting.md before recommending a rerun.

Boundaries

• Stay in this sub-skill for image-classification commands and classification-specific Hugging Face usage.
• Route DCNv3 CUDA-extension build diagnosis, TensorRT custom-op setup, and ONNX/TensorRT export execution to the sibling deployment sub-skill. Keep only the classification model name, config, and checkpoint selection here.
• Do not claim that full training, evaluation, throughput, feature extraction, Hugging Face downloads, DCNv3 CUDA builds, or TensorRT export were verified by this generated skill. The selected verification scope covered self-contained helper checks and static source distillation; dataset/GPU/network-heavy runs remain user-approved runtime actions.

Quick command-builder examples

# Build a one-GPU ImageNet evaluation template.
python scripts/build_classification_command.py \
  --mode eval \
  --config configs/internimage_b_1k_224.yaml \
  --checkpoint CHANGE_ME/internimage_b_1k_224.pth \
  --data-path CHANGE_ME/imagenet \
  --gpus 1

# Build an intermediate-feature extraction template.
python scripts/build_classification_command.py \
  --mode extract-features \
  --config configs/internimage_t_1k_224.yaml \
  --checkpoint CHANGE_ME/internimage_t_1k_224.pth \
  --image CHANGE_ME/image.png \
  --keys patch_embed levels.0.downsample \
  --save-features

# Build a Hugging Face Transformers inference template.
python scripts/build_classification_command.py \
  --mode hf-transformers \
  --hf-model OpenGVLab/internimage_t_1k_224 \
  --hf-task both \
  --image CHANGE_ME/image.png