Model training and ONNX export code for the explorablecv computer-vision demos.
| Folder | App | What it produces |
|---|---|---|
| cnn-architecture-comparison | cnn-architecture-comparison | CIFAR-10 classifiers (LeNet, AlexNet, VGG, ResNet, ...) exported to ONNX |
| cnn-visualizer | cnn-visualizer | A small CNN plus per-layer activation dumps for the visualizer |
| cv-detection-playground | cv-detection-playground | YOLOv8n exported to ONNX with raw per-anchor output |
| cv-interpretability | cv-interpretability | Imagenette classifier plus CAM / LIME / attention-rollout / UMAP bundles |
| cv-segmentation-playground | cv-segmentation-playground | Semantic, instance, and panoptic segmentation ONNX exports plus precomputed sample results |
| embeddings-playground | embeddings-playground | CLIP image embeddings over COCO, reduced to 2D/3D with UMAP |
| vit-playground | vit-playground | ViT-tiny exported with logits, patch embeddings, per-layer attention |
scripts holds dataset downloads shared by more than one app:
download_cifar10.py: CIFAR-10 (used by cnn-architecture-comparison, cnn-visualizer)download_imagenette.py: Imagenette / Imagewoof (used by cv-interpretability, vit-playground)data_paths.py: resolves the shared data root
Datasets land in a shared data/ cache at the repo root so every app reads from one place. Override the
location with the EXPLORABLECV_DATA environment variable, or pass --dest for a private copy
(cv-interpretability does, since its preprocess renames class dirs in place). The shared scripts only
populate the cache. Each app keeps its own loaders, so no app depends on another.
python scripts/download_cifar10.py
python scripts/download_imagenette.py