README.md

metadata

license: other
license_name: imagenet-derived-research-only
license_link: https://www.image-net.org/download.php
viewer: false
pretty_name: ImageNet backbone features (fp32) for SAE training
tags:
  - embeddings
  - features
  - imagenet
  - sparse-autoencoder
extra_gated_heading: Request access to these features
extra_gated_description: >-
  Access requests are reviewed manually by the authors. Expect a few days for
  processing.
extra_gated_prompt: >
  These features are derived from ImageNet-1k. By requesting access, you agree
  to:

  1. Comply with the ImageNet terms of access — non-commercial research and
  education only.

  2. Not redistribute the features, in whole or in part, to any third party.

  3. Cite both ImageNet and the associated work in any publication that uses
  them.
extra_gated_fields:
  Full name: text
  Affiliation or company: text
  Work email: text
  Country: country
  Intended use:
    type: select
    options:
      - Academic research
      - Evaluation on an interpretability platform
      - label: Other
        value: other
  Describe your intended use in one sentence: text
  I have accepted the ImageNet terms of access: checkbox
  I agree not to redistribute these features: checkbox
extra_gated_button_content: Request access

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	—	5.0 Go / 198 Mo
resnet50	layer4	train / val	16 / 4	—	9.9 Go / 393 Mo
vit_b_16	penultimate	train / val	16 / 4	768	3.8 Go / 149 Mo
dinov2_vitb14_reg	penultimate	train / val	16 / 4	768	3.8 Go / 149 Mo

<backbone>/<node>/<split>/<split>_000NN.npy     # shards contigus, fp32
concept_vocab_thr0500.json                      # vocabulaire concepts, seuil 500
concept_vocab_thr5000.json                      # vocabulaire concepts, seuil 5000

Un shard train est de forme [100000, D], un shard val de forme [50000, D]. Les shards sont ordonnés : concaténés dans l'ordre lexicographique des noms, ils reconstituent le split complet dans l'ordre d'origine du DataLoader.

Chargement

import numpy as np, glob
shards = sorted(glob.glob("resnet50/avgpool/train/train_*.npy"))
X = np.concatenate([np.load(p, mmap_mode="r") for p in shards])   # ou lire shard par shard

Le pooling est un GAP (--pool gap), sans projection (--proj-dim 0), extrait en fp32.

Provenance

Extraites avec extract_features_fp32_blocks_imn_VIT.py (inclus dans le repo SAE_CBM_unification) depuis ImageNet-1k train/val. Le pipeline SAE ne lit jamais les images brutes : il consomme directement ces shards mémoire-mappés.

Licence

Features dérivées d'ImageNet-1k. La redistribution est soumise aux conditions d'accès d'ImageNet : usage recherche non commercial. L'accès est gaté pour cette raison — voir le formulaire ci-dessus.

Citation

TODO — référence de l'article + citation ImageNet (Deng et al., 2009).