Hugging Face's logo

PyTorch Image Models
https://github.com/rwightman/pytorch-image-models
Activity Feed

AI & ML interests

Computer Vision

Recent Activity

rwightmanΒ  new activity about 15 hours ago
timm/vit_little_patch16_reg4_gap_256.sbb_in1k:Loss exploding to nan
rwightmanΒ  updated a model 3 days ago
timm/mobilenetv5_300m.gemma3n
rwightmanΒ  published a model 3 days ago
timm/mobilenetv5_300m.gemma3n
View all activity

timm 's collections 18

Perception Encoder
OpenCLIP (PE Core image + text) and timm PE Core, Spatial, Lang (ViT only) weights. NOTE: These weights do not work with original modeling code.
Searching for Better ViT Baselines
Exploring ViT hparams and model shapes for the GPU poor (between tiny and base).
MobileNetV4 pretrained weights
Weights for MobileNet-V4 pretrained in timm
timm Top-20 Fastest Models
Not the most accurate, but the highest throughput image classification models in timm
timm Takes on the Classics
timm includes the most popular convolutional and vision transformer models, many with new weights from updated training recipes.
Fastest timm models > 80% Top-1 ImageNet-1k
Fastest image classification models with 80% accuracy in ImageNet-1k .
Fastest timm models > 86% ImageNet-1k Top-1
Fastest image classification models with 86% accuracy in ImageNet-1k .
timm Backbones
Pre-trained feature extraction backbones available in timm.
MetaCLIP
MetaCLIP & MetaCLIP2 OpenCLIP and timm models. All models are dual timm + OpenCLIP (or just timm for specific vit encoders).
timm Top-20 ImageNet-1k Models
The 20 best models on ImageNet-1k validation set, all pretrained on datasets larger than ImageNet and fine-tuned on ImageNet-1k.
timm ImageNet-12k Models
timm has a number of unique and exclusive models trained on a 11821 (12k) subset of the full ImageNet-22k
Fastest timm models > 75.3% IN-1k Top-1 (Original ResNet-50)
Fastest image classification models with 75.3% accuracy in ImageNet-1k .
Fastest timm models > 83% ImageNet-1k Top-1
Fastest image classification models with 83% accuracy in ImageNet-1k .
Fastest timm models > 88% ImageNet-1k Top-1
Fastest image classification models with 88% accuracy in ImageNet-1k .
timm tiny test models
A collection of very small (~300-500k parameter) models at 160x160 resolution, for testing purposes. Trained on ImageNet-1k.
Perception Encoder
OpenCLIP (PE Core image + text) and timm PE Core, Spatial, Lang (ViT only) weights. NOTE: These weights do not work with original modeling code.
Searching for Better ViT Baselines
Exploring ViT hparams and model shapes for the GPU poor (between tiny and base).
MetaCLIP
MetaCLIP & MetaCLIP2 OpenCLIP and timm models. All models are dual timm + OpenCLIP (or just timm for specific vit encoders).
MobileNetV4 pretrained weights
Weights for MobileNet-V4 pretrained in timm
timm Top-20 ImageNet-1k Models
The 20 best models on ImageNet-1k validation set, all pretrained on datasets larger than ImageNet and fine-tuned on ImageNet-1k.
timm Top-20 Fastest Models
Not the most accurate, but the highest throughput image classification models in timm
timm ImageNet-12k Models
timm has a number of unique and exclusive models trained on a 11821 (12k) subset of the full ImageNet-22k
timm Takes on the Classics
timm includes the most popular convolutional and vision transformer models, many with new weights from updated training recipes.
Fastest timm models > 75.3% IN-1k Top-1 (Original ResNet-50)
Fastest image classification models with 75.3% accuracy in ImageNet-1k .
Fastest timm models > 80% Top-1 ImageNet-1k
Fastest image classification models with 80% accuracy in ImageNet-1k .
Fastest timm models > 83% ImageNet-1k Top-1
Fastest image classification models with 83% accuracy in ImageNet-1k .
Fastest timm models > 86% ImageNet-1k Top-1
Fastest image classification models with 86% accuracy in ImageNet-1k .
Fastest timm models > 88% ImageNet-1k Top-1
Fastest image classification models with 88% accuracy in ImageNet-1k .
timm Backbones
Pre-trained feature extraction backbones available in timm.
timm tiny test models
A collection of very small (~300-500k parameter) models at 160x160 resolution, for testing purposes. Trained on ImageNet-1k.