PyTorch Hub
For Researchers

HybridNets - End2End Perception Network

Resnet Style Video classification networks pretrained on the Kinetics 400 dataset

SlowFast networks pretrained on the Kinetics 400 dataset

X3D networks pretrained on the Kinetics 400 dataset

YOLOP pretrained on the BDD100K dataset

MiDaS models for computing relative depth from a single image.

classify birds using this fine-grained image classifier

The FastPitch model for generating mel spectrograms from text

GPUNet is a new family of Convolutional Neural Networks designed to max out the performance of NVIDIA GPU and TensorRT.

The HiFi GAN model for generating waveforms from mel spectrograms

Once-for-all (OFA) decouples training and search, and achieves efficient inference across various edge devices and resource constraints.

Reference implementation for music source separation

A set of compact enterprise-grade pre-trained STT Models for multiple languages.

A set of compact enterprise-grade pre-trained TTS Models for multiple languages

Pre-trained Voice Activity Detector

Ultralytics YOLOv5 🚀 for object detection, instance segmentation and image classification.

DeepLabV3 models with ResNet-50, ResNet-101 and MobileNet-V3 backbones

Transformer models for English-French and English-German translation.

ResNext models trained with billion scale weakly-supervised data.

A simple generative image model for 64x64 images

High-quality image generation of fashion, celebrity faces

ResNet and ResNext models introduced in the "Billion scale semi-supervised learning for image classification" paper

PyTorch implementations of popular NLP Transformers

U-Net with batch normalization for biomedical image segmentation with pretrained weights for abnormality segmentation in brain MRI

EfficientNets are a family of image classification models, which achieve state-of-the-art accuracy, being an order-of-magnitude smaller and faster....

ResNet50 model trained with mixed precision using Tensor Cores.

ResNet with bottleneck 3x3 Convolutions substituted by 3x3 Grouped Convolutions, trained with mixed precision using Tensor Cores.

ResNeXt with Squeeze-and-Excitation module added, trained with mixed precision using Tensor Cores.

Single Shot MultiBox Detector model for object detection

The Tacotron 2 model for generating mel spectrograms from text

WaveGlow model for generating speech from mel spectrograms (generated by Tacotron2)

A Robustly Optimized BERT Pretraining Approach

The 2012 ImageNet winner achieved a top-5 error of 15.3%, more than 10.8 percentage points lower than that of the runner up.

Dense Convolutional Network (DenseNet), connects each layer to every other layer in a feed-forward fashion.

Fully-Convolutional Network model with ResNet-50 and ResNet-101 backbones

Efficient networks by generating more features from cheap operations

GoogLeNet was based on a deep convolutional neural network architecture codenamed "Inception" which won ImageNet 2014.

Harmonic DenseNet pre-trained on ImageNet

Networks with domain/appearance invariance

Also called GoogleNetv3, a famous ConvNet trained on ImageNet from 2015

Boosting Tiny and Efficient Models using Knowledge Distillation.

Efficient networks optimized for speed and memory, with residual blocks

Proxylessly specialize CNN architectures for different hardware platforms.

A new ResNet variant.

Deep residual networks pre-trained on ImageNet

Next generation ResNets, more efficient and accurate

An efficient ConvNet optimized for speed and memory, pre-trained on ImageNet

Brain-inspired Multilayer Perceptron with Spiking Neurons

Alexnet-level accuracy with 50x fewer parameters.

Award winning ConvNets from 2014 ImageNet ILSVRC challenge

Wide Residual Networks

Lets Keep it simple, Using simple architectures to outperform deeper and more complex architectures