computer-visionlisted
Install: claude install-skill TranBaDat2607/claude-code-kaggle-grandmaster
# Computer Vision Playbook
## 1. Baseline recipe (works for most classification tasks)
- Backbone from `timm`: start with a fast, strong ImageNet-pretrained model
(EfficientNet-B0/B3, ConvNeXt-Tiny/Small, EfficientNetV2-S, NFNet-F0, or `tf_efficientnet`
variants); ViT/EVA/DINOv2/SigLIP backbones for semantic tasks with enough data.
- Input: native-ish resolution of the objects that matter (resolution is often the #1
lever — try 384→512→768+ once the pipeline is solid).
- AdamW, lr 1e-4–3e-4 (backbone) with head ×10, cosine schedule with warmup (1 epoch),
weight decay 1e-2, mixed precision (bf16/fp16), batch as large as fits, EMA of weights.
- Augmentations (albumentations): flips, shift-scale-rotate, random resized crop,
brightness/contrast, hue/saturation, CoarseDropout; MixUp/CutMix for classification.
Choose augmentations that preserve the label (no vertical flips for text, careful with
colour in medical/satellite data).
- 5 folds (grouped by patient/source!), 10–30 epochs, save best-by-metric per fold, OOF preds.
- Template: `templates/train_image.py`.
## 2. Levers in rough order of typical impact
1. **Correct grouping in CV** (patients, studies, scenes, near-duplicates via image hashing).
2. **Data understanding**: look at hundreds of images per class, at the errors, at label noise.
3. **Resolution & cropping** (ROI crops via a first-stage detector/segmenter are huge in
medical/satellite comps).
4. **Backbone family & size** (scale after the pipeline