timm documentation

EVA

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

EVA

カテゴリ: ViT · コード: timm/models/eva.py · アーキテクチャ: 65 · 事前学習済み重み: 83 (ImageNet-1kの結果を含む28)

Eva クラスを共有する EVA および EVA-02 Vision Transformer、その他の ViT 変種で、回転位置埋め込み (RoPE)、SwiGLU MLP、関連する追加機能を使用します。このモジュールには、Perception Encoder (PE)、ROPE-ViT、DINOv3、LingBot-Vision、Sapiens2 モデル、および RoPE を備えた timm のオリジナル ViT 設定も含まれます。

論文

使用方法

事前学習済み重みの読み込み、画像の前処理、ファインチューニングについてはクイックスタートを、モデルをバックボーンとして使用する方法については特徴量抽出を参照してください。

>>> import timm
>>> model = timm.create_model('eva02_small_patch14_336.mim_in22k_ft_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)

モデルAPIのサポート

機能Eva
features_only=True 特徴マップ✓ (strides 16, 16, 16)
forward_intermediates()✓
prune_intermediate_layers()✓
入力サイズは任意で、変更は不要です✓
set_input_size()✓
dynamic_img_size=True✓
set_grad_checkpointing()✓
レイヤーごとのLR減衰(group_matcher())✓
reset_classifier()✓

✓ 対応、✓ (square) 正方形サイズのみ動作、✗ 非対応、– 不要(モデルはすでに任意の入力サイズを受け入れる)。ランダム初期化されたモデル(Eva の vit_tiny_patch16_dinov3_qkvb、256)で各機能を実行して確認しました。同じクラスの他のアーキテクチャも同様に動作すると考えられます。詳細は Model API ページを参照してください。

アーキテクチャ

アーキテクチャパラメータ数 (M)GMACs入力サイズ事前学習済み重み
eva02_base_patch14_22485.7623.222241
eva02_base_patch14_44887.12107.114483
eva02_base_patch16_clip_22486.2617.622241
eva02_enormous_patch14_clip_2244350.561132.462242
eva02_large_patch14_224303.2781.152242
eva02_large_patch14_448305.08362.334486
eva02_large_patch14_clip_224304.1181.182241
eva02_large_patch14_clip_336304.43191.343361
eva02_small_patch14_22421.626.142241
eva02_small_patch14_33622.1315.483361
eva02_tiny_patch14_2245.501.702241
eva02_tiny_patch14_3365.764.683361
eva_giant_patch14_2241012.56267.182241
eva_giant_patch14_3361013.01620.643362
eva_giant_patch14_5601014.451906.765601
eva_giant_patch14_clip_2241012.59267.182242
vit_5b_patch16_sapiens25066.90-1024x7681
vit_7b_patch16_dinov36716.031775.102562
vit_base_patch16_dinov385.6423.602561
vit_base_patch16_dinov3_qkvb85.6623.602562
vit_base_patch16_lingbot85.66-5121
vit_base_patch16_rope_22486.4317.582241
vit_base_patch16_rope_ape_22486.5917.582241
vit_base_patch16_rope_mixed_22486.4417.582241
vit_base_patch16_rope_mixed_ape_22486.5917.582241
vit_base_patch16_rope_reg1_gap_25686.4323.222561
vit_base_patch16_sapiens2113.99-1024x7681
vit_betwixt_patch16_rope_reg4_gap_25660.2316.522561
vit_giant_patch16_lingbot1134.47-5121
vit_giant_patch16_sapiens21455.47-1024x7681
vit_huge_patch16_sapiens2814.28-1024x7681
vit_huge_plus_patch16_dinov3840.51224.882561
vit_huge_plus_patch16_dinov3_qkvb840.59224.882561
vit_large_patch16_dinov3303.0882.432562
vit_large_patch16_dinov3_qkvb303.1382.432562
vit_large_patch16_lingbot303.13-5121
vit_large_patch16_rope_224304.1761.602241
vit_large_patch16_rope_ape_224304.3761.602241
vit_large_patch16_rope_mixed_224304.2061.602241
vit_large_patch16_rope_mixed_ape_224304.4061.602241
vit_large_patch16_sapiens2395.45-1024x7681
vit_medium_patch16_rope_reg1_gap_25638.7410.632561
vit_mediumd_patch16_rope_reg1_gap_25663.9517.652561
vit_pe_core_base_patch16_22493.6717.822241
vit_pe_core_gigantic_patch14_4481882.032060.124481
vit_pe_core_large_patch14_336317.15192.333361
vit_pe_core_small_patch16_38423.7815.693841
vit_pe_core_tiny_patch16_3846.144.743841
vit_pe_lang_gigantic_patch14_4481740.921931.994482
vit_pe_lang_large_patch14_448291.42346.994482
vit_pe_spatial_base_patch16_51286.43107.135121
vit_pe_spatial_gigantic_patch14_4481851.892055.254481
vit_pe_spatial_large_patch14_448303.96362.054481
vit_pe_spatial_small_patch16_51221.9831.805121
vit_pe_spatial_tiny_patch16_5125.6810.465121
vit_small_patch16_dinov321.596.262561
vit_small_patch16_dinov3_qkvb21.606.262562
vit_small_patch16_lingbot21.60-5121
vit_small_patch16_rope_22421.984.612241
vit_small_patch16_rope_ape_22422.064.612241
vit_small_patch16_rope_mixed_22421.994.612241
vit_small_patch16_rope_mixed_ape_22422.064.612241
vit_small_plus_patch16_dinov328.688.112561
vit_small_plus_patch16_dinov3_qkvb28.698.112561
vit_tiny_patch16_dinov3_qkvb5.49-2561

GMACsとパラメータ数は、各アーキテクチャのデフォルトの入力サイズと分類器に基づいており、ベンチマーク結果から得られたものです。

事前学習済み重み

Top-1 は、results-imagenet.csv に記載された評価サイズでの ImageNet-1k 検証精度です。- は、ImageNet-1k の結果がない重み(例: プリトレーニング用または画像-テキスト用重み)を示します。Classes は事前学習済み分類器のサイズで、分類器を持たない重みについては none です。

重みはソースごとにグループ化されます。ソースは、事前学習設定で指定されている論文または出典、それ以外の場合はtimm以外の Hub 組織、さらにそれ以外の場合は事前学習タグの最初の部分(architecture.tag)です。重みが 1 つしかないソースは、other にリストされます。

mim タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva02_base_patch14_224.mim_in22k224--なしmittimm
eva02_base_patch14_448.mim_in22k_ft_in1k44888.2624481000mittimm
eva02_base_patch14_448.mim_in22k_ft_in22k448--21841mittimm
eva02_base_patch14_448.mim_in22k_ft_in22k_in1k44888.6784481000mittimm
eva02_large_patch14_224.mim_in22k224--なしmittimm
eva02_large_patch14_224.mim_m38m224--なしmittimm
eva02_large_patch14_448.mim_in22k_ft_in1k44889.6344481000mittimm
eva02_large_patch14_448.mim_in22k_ft_in22k448--21841mittimm
eva02_large_patch14_448.mim_in22k_ft_in22k_in1k44889.9564481000mittimm
eva02_large_patch14_448.mim_m38m_ft_in1k44889.5504481000mittimm
eva02_large_patch14_448.mim_m38m_ft_in22k448--21841mittimm
eva02_large_patch14_448.mim_m38m_ft_in22k_in1k44890.0564481000mittimm
eva02_small_patch14_224.mim_in22k224--なしmittimm
eva02_small_patch14_336.mim_in22k_ft_in1k33685.7703361000mittimm
eva02_tiny_patch14_224.mim_in22k224--なしmittimm
eva02_tiny_patch14_336.mim_in22k_ft_in1k33680.6483361000mittimm

fb タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_pe_core_base_patch16_224.fb224--1024apache-2.0timm
vit_pe_core_gigantic_patch14_448.fb448--1280apache-2.0timm
vit_pe_core_large_patch14_336.fb336--1024apache-2.0timm
vit_pe_core_small_patch16_384.fb384--512apache-2.0timm
vit_pe_core_tiny_patch16_384.fb384--512apache-2.0timm
vit_pe_lang_gigantic_patch14_448.fb448--なしapache-2.0timm
vit_pe_lang_gigantic_patch14_448.fb_tiling448--なしapache-2.0timm
vit_pe_lang_large_patch14_448.fb448--なしapache-2.0timm
vit_pe_lang_large_patch14_448.fb_tiling448--なしapache-2.0timm
vit_pe_spatial_base_patch16_512.fb512--なしapache-2.0timm
vit_pe_spatial_gigantic_patch14_448.fb448--なしapache-2.0timm
vit_pe_spatial_large_patch14_448.fb448--なしapache-2.0timm
vit_pe_spatial_small_patch16_512.fb512--なしapache-2.0timm
vit_pe_spatial_tiny_patch16_512.fb512--なしapache-2.0timm

naver タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_base_patch16_rope_224.naver_in1k22483.7142241000apache-2.0timm
vit_base_patch16_rope_ape_224.naver_in1k22483.7802241000apache-2.0timm
vit_base_patch16_rope_mixed_224.naver_in1k22483.8062241000apache-2.0timm
vit_base_patch16_rope_mixed_ape_224.naver_in1k22483.8982241000apache-2.0timm
vit_large_patch16_rope_224.naver_in1k22484.6462241000apache-2.0timm
vit_large_patch16_rope_ape_224.naver_in1k22484.6602241000apache-2.0timm
vit_large_patch16_rope_mixed_224.naver_in1k22484.8402241000apache-2.0timm
vit_large_patch16_rope_mixed_ape_224.naver_in1k22484.8462241000apache-2.0timm
vit_small_patch16_rope_224.naver_in1k22481.2222241000apache-2.0timm
vit_small_patch16_rope_ape_224.naver_in1k22481.0162241000apache-2.0timm
vit_small_patch16_rope_mixed_224.naver_in1k22481.2522241000apache-2.0timm
vit_small_patch16_rope_mixed_ape_224.naver_in1k22480.9742241000apache-2.0timm

lvd1689m タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_7b_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_base_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_base_patch16_dinov3_qkvb.lvd1689m256--なしdinov3-licensetimm
vit_huge_plus_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_huge_plus_patch16_dinov3_qkvb.lvd1689m256--なしdinov3-licensetimm
vit_large_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_large_patch16_dinov3_qkvb.lvd1689m256--なしdinov3-licensetimm
vit_small_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_small_patch16_dinov3_qkvb.lvd1689m256--なしdinov3-licensetimm
vit_small_plus_patch16_dinov3.lvd1689m256--なしdinov3-licensetimm
vit_small_plus_patch16_dinov3_qkvb.lvd1689m256--なしdinov3-licensetimm

Sapiens2

事前学習済み設定はSapiens2を参照しています。

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_5b_patch16_sapiens2.fb1024x768--なしsapiens2-licensetimm
vit_base_patch16_sapiens2.fb1024x768--なしsapiens2-licensetimm
vit_giant_patch16_sapiens2.fb1024x768--なしsapiens2-licensetimm
vit_huge_patch16_sapiens2.fb1024x768--なしsapiens2-licensetimm
vit_large_patch16_sapiens2.fb1024x768--なしsapiens2-licensetimm

merged2b タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva02_base_patch16_clip_224.merged2b224--512mittimm
eva02_large_patch14_clip_224.merged2b224--768mittimm
eva02_large_patch14_clip_336.merged2b336--768mittimm
eva_giant_patch14_clip_224.merged2b224--1024mittimm

sbb タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_base_patch16_rope_reg1_gap_256.sbb_in1k25683.8722561000mittimm
vit_betwixt_patch16_rope_reg4_gap_256.sbb_in1k25683.9062561000mittimm
vit_medium_patch16_rope_reg1_gap_256.sbb_in1k25683.8102561000mittimm
vit_mediumd_patch16_rope_reg1_gap_256.sbb_in1k25684.3162561000mittimm

密な空間知覚のためのビジョン事前学習

事前学習済み設定はVision Pretraining for Dense Spatial Perceptionを参照しています。

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_base_patch16_lingbot.robbyant512--なしapache-2.0belfner/vit_base_patch16_lingbot.robbyant
vit_giant_patch16_lingbot.robbyant512--なしapache-2.0belfner/vit_giant_patch16_lingbot.robbyant
vit_large_patch16_lingbot.robbyant512--なしapache-2.0belfner/vit_large_patch16_lingbot.robbyant
vit_small_patch16_lingbot.robbyant512--なしapache-2.0belfner/vit_small_patch16_lingbot.robbyant

eupe タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_base_patch16_dinov3_qkvb.eupe_lvd1689m256--なしfair-noncommercial-research-licensetimm
vit_small_patch16_dinov3_qkvb.eupe_lvd1689m256--なしfair-noncommercial-research-licensetimm
vit_tiny_patch16_dinov3_qkvb.eupe_lvd1689m256--なしfair-noncommercial-research-licensetimm

sat493m タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
vit_7b_patch16_dinov3.sat493m256--なしdinov3-licensetimm
vit_large_patch16_dinov3.sat493m256--なしdinov3-licensetimm
vit_large_patch16_dinov3_qkvb.sat493m256--なしdinov3-licensetimm

クリップタグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva_giant_patch14_224.clip_ft_in1k22488.8962241000mittimm
eva_giant_patch14_336.clip_ft_in1k33689.4623361000mittimm

laion2b タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva02_enormous_patch14_clip_224.laion2b224--1024mittimm
eva02_enormous_patch14_clip_224.laion2b_plus224--1024mittimm

m30m タグ

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva_giant_patch14_336.m30m_ft_in22k_in1k33689.5663361000mittimm
eva_giant_patch14_560.m30m_ft_in22k_in1k56089.7905601000mittimm

その他

重み入力サイズTop-1 (%)評価サイズクラスライセンスHub
eva_giant_patch14_clip_224.laion400m224--1024mittimm

引用

@article{EVA,
  title={EVA: Exploring the Limits of Masked Visual Representation Learning at Scale},
  author={Fang, Yuxin and Wang, Wen and Xie, Binhui and Sun, Quan and Wu, Ledell and Wang, Xinggang and Huang,
  Tiejun and Wang, Xinlong and Cao, Yue},
  journal={arXiv preprint arXiv:2211.07636},
  year={2022}
}

@article{EVA02,
  title={EVA-02: A Visual Representation for Neon Genesis},
  author={Fang, Yuxin and Sun, Quan and Wang, Xinggang and Huang, Tiejun and Wang, Xinlong and Cao, Yue},
  journal={arXiv preprint arXiv:2303.11331},
  year={2023}
}

@article{bolya2025perception,
  title={Perception encoder: The best visual embeddings are not at the output of the network},
  author={Bolya, Daniel and Huang, Po-Yao and Sun, Peize and Cho, Jang Hyun and Madotto, Andrea and Wei, Chen and Ma,
    Tengyu and Zhi, Jiale and Rajasegaran, Jathushan and Rasheed, Hanoona and others},
  journal={arXiv preprint arXiv:2504.13181},
  year={2025}
}

@inproceedings{heo2024rotary,
  title={Rotary position embedding for vision transformer},
  author={Heo, Byeongho and Park, Song and Han, Dongyoon and Yun, Sangdoo},
  booktitle={European Conference on Computer Vision},
  pages={289--305},
  year={2024},
  organization={Springer}
}

@article{simeoni2025dinov3,
  title={{DINOv3}},
  author={Sim{'e}oni, Oriane and Vo, Huy V. and Seitzer, Maximilian and Baldassarre, Federico and Oquab, Maxime
    and Jose, Cijo and Khalidov, Vasil and Szafraniec, Marc and Yi, Seungeun and Ramamonjisoa, Micha{"e}l
    and Massa, Francisco and Haziza, Daniel and Wehrstedt, Luca and Wang, Jianyuan and Darcet, Timoth{'e}e
    and Moutakanni, Th{'e}o and Sentana, Leonel and Roberts, Claire and Vedaldi, Andrea and Tolan, Jamie
    and Brandt, John and Couprie, Camille and Mairal, Julien and J{'e}gou, Herv{'e} and Labatut, Patrick
    and Bojanowski, Piotr},
  year={2025},
  eprint={2508.10104},
  url={https://arxiv.org/abs/2508.10104},
}

@article{lingbot-vision2026,
  title={Vision Pretraining for Dense Spatial Perception},
  author={Fu, Zelin and Tan, Bin and Sun, Changjiang and Liu, Shaohui and Zheng, Kecheng and Xu, Yinghao
    and Zhu, Xing and Shen, Yujun and Xue, Nan},
  journal={arXiv preprint arXiv:2607.05247},
  year={2026}
}

@article{khirodkarsapiens2,
  title={Sapiens2},
  author={Khirodkar, Rawal and Wen, He and Martinez, Julieta and Dong, Yuan and Su, Zhaoen and Saito, Shunsuke},
  journal={arXiv preprint arXiv:2604.21681},
  year={2026}
}
Update on GitHub