AlexandreRocchi/imagenet-backbone-features-fp32
ImageNet backbone features (fp32) Activations pré-extraites d'ImageNet-1k pour trois backbones, mémoire-mappables, utilisées pour entraîner les SAE du repo SAE_CBM_unification. Contenu 136 shards .npy, 37.2 Go, fp32. Backbone Node Split Shards Dim Taille resnet50 avgpool train / val 14 / 2 2048 9.8 Go / 392 Mo resnet50 layer1 train / val 16 / 4 — 1.3 Go / 51 Mo resnet50 layer2 train / val 16 / 4 — 2.5 Go / 100 Mo resnet50 layer3 train / val 16 / 4… See the full description on the dataset page: https://huggingface.co/datasets/AlexandreRocchi/imagenet-backbone-features-fp32.
04
No card is published for this repository, or it could not be fetched from Hugging Face right now.
