super-gradients 2.5.0

SuperGradients

Easily train or fine-tune SOTA computer vision models with one open source training library

Fill our 4-question quick survey! We will raffle free SuperGradients swag between those who will participate -> Fill Survey

---

Website • Why Use SG? • User Guide • Docs • Getting Started Notebooks • Transfer Learning • Pretrained Models • Community • License • Deci Platform

SuperGradients

Introduction

Welcome to SuperGradients, a free, open-source training library for PyTorch-based deep learning models. SuperGradients allows you to train or fine-tune SOTA pre-trained models for all the most commonly applied computer vision tasks with just one training library. We currently support object detection, image classification and semantic segmentation for videos and images.

Docs and full user guide

Why use SuperGradients?

Built-in SOTA Models

Easily load and fine-tune production-ready, pre-trained SOTA models that incorporate best practices and validated hyper-parameters for achieving best-in-class accuracy.

Easily Reproduce our Results

Why do all the grind work, if we already did it for you? leverage tested and proven recipes & code examples for a wide range of computer vision models generated by our team of deep learning experts. Easily configure your own or use plug & play hyperparameters for training, dataset, and architecture.

Production Readiness and Ease of Integration

All SuperGradients models’ are production ready in the sense that they are compatible with deployment tools such as TensorRT (Nvidia) and OpenVINO (Intel) and can be easily taken into production. With a few lines of code you can easily integrate the models into your codebase.

What's New

• 【06/9/2022】 PP-LiteSeg - new pre-trained checkpoints for Cityscapes with SOTA mIoU scores (~1.5% above paper)🎯
• 【07/08/2022】DDRNet23 - new pre-trained checkpoints and recipes for Cityscapes with SOTA mIoU scores (~1% above paper)🎯
• 【27/07/2022】YOLOX models (object detection) - recipes and pre-trained checkpoints.
• 【07/07/2022】SSD Lite MobileNet V2,V1 - Training recipes and pre-trained checkpoints on COCO - Tailored for edge devices! 📱
• 【07/07/2022】 STDC - new pre-trained checkpoints and recipes for Cityscapes with super SOTA mIoU scores (~2.5% above paper)🎯
• 【16/06/2022】 ResNet50 - new pre-trained checkpoint and recipe for ImageNet top-1 score of 81.9 💪
• 【09/06/2022】 ViT models (Vision Transformer) - Training recipes and pre-trained checkpoints (ViT, BEiT).
• 【09/06/2022】 Knowledge Distillation support - training module and notebook.
• 【06/04/2022】 Integration with professional tools - Weights and Biases and DagsHub.
• 【09/03/2022】 New quick start and transfer learning example notebooks for Semantic Segmentation.
• 【07/02/2022】 We added RegSeg recipes and pre-trained models to our Semantic Segmentation models.
• 【01/02/2022】 We added issue templates for feature requests and bug reporting.
• 【20/01/2022】 STDC family - new recipes added with even higher mIoU💪

Check out SG full release notes.

Coming soon

• PP-LiteSeg recipes for Cityscapes with SOTA mIoU scores (~1.5% above paper)🎯
• Single class detectors (recipes, pre-trained checkpoints) for edge devices deployment.
• Single class segmentation (recipes, pre-trained checkpoints) for edge devices deployment.
• QAT capabilities (Quantization Aware Training).
• Dali implementation.
• Integration with more professional tools.
• Improved pre-trained checkpoints and recipes (DDRNet, ResNet, RegSeg, etc.)

---

Table of Content

• Getting Started
  • Quick Start Notebook - Classification example
  • Quick Start Notebook - Semantic segmentation example

• Transfer Learning
  • Transfer Learning with SG Notebook - Semantic segmentation example
• Knowledge Distillation Training
  • Knowledge Distillation Training Quick Start with SG Notebook - ResNet18 example
• Installation Methods
  • Prerequisites
  • Quick Installation
• Computer Vision Models - Pretrained Checkpoints
  • Pretrained Classification PyTorch Checkpoints
  • Pretrained Object Detection PyTorch Checkpoints
  • Pretrained Semantic Segmentation PyTorch Checkpoints
• Implemented Model Architectures
• Training Recipes
• Contributing
• Citation
• Community
• License
• Deci Platform

Getting Started

Start Training with Just 1 Command Line

The most simple and straightforward way to start training SOTA performance models with SuperGradients reproducible recipes. Just define your dataset path and where you want your checkpoints to be saved and you are good to go from your terminal!

python -m super_gradients.train_from_recipe --config-name=imagenet_regnetY architecture=regnetY800 dataset_interface.data_dir=<YOUR_Imagenet_LOCAL_PATH> ckpt_root_dir=<CHEKPOINT_DIRECTORY>

Quickly Load Pre-Trained Weights for Your Desired Model with SOTA Performance

Want to try our pre-trained models on your machine? Import SuperGradients, initialize your Trainer, and load your desired architecture and pre-trained weights from our SOTA model zoo

# The pretrained_weights argument will load a pre-trained architecture on the provided dataset
# This is an example of loading COCO-2017 pre-trained weights for a YOLOX Nano object detection model

import super_gradients
from super_gradients.training import Trainer

trainer = SgModel(experiment_name="yoloxn_coco_experiment",ckpt_root_dir=<CHECKPOINT_DIRECTORY>)
trainer.build_model(architecture="yolox_n", arch_params={"pretrained_weights": "coco", num_classes": 80})

Quick Start Notebook - Classification

Get started with our quick start notebook for image classification tasks on Google Colab for a quick and easy start using free GPU hardware.

Classification Quick Start in Google Colab

Download notebook

View source on GitHub

Quick Start Notebook - Semantic Segmentation

Get started with our quick start notebook for semantic segmentation tasks on Google Colab for a quick and easy start using free GPU hardware.

Segmentation Quick Start in Google Colab

Download notebook

View source on GitHub

Transfer Learning

Transfer Learning with SG Notebook - Semantic Segmentation

Learn more about SuperGradients transfer learning or fine tuning abilities with our Citiscapes pre-trained RegSeg48 fine tuning into a sub-dataset of Supervisely example notebook on Google Colab for an easy to use tutorial using free GPU hardware

Segmentation Transfer Learning in Google Colab

Download notebook

View source on GitHub

Knowledge Distillation Training

Knowledge Distillation Training Quick Start with SG Notebook - ResNet18 example

Knowledge Distillation is a training technique that uses a large model, teacher model, to improve the performance of a smaller model, the student model. Learn more about SuperGradients knowledge distillation training with our pre-trained BEiT base teacher model and Resnet18 student model on CIFAR10 example notebook on Google Colab for an easy to use tutorial using free GPU hardware

KD Training in Google Colab

Download notebook

View source on GitHub

Installation Methods

Prerequisites

General requirements

• Python 3.7, 3.8 or 3.9 installed.
• torch>=1.9.0
  • https://pytorch.org/get-started/locally/
• The python packages that are specified in requirements.txt;

To train on nvidia GPUs

• Nvidia CUDA Toolkit >= 11.2
• CuDNN >= 8.1.x
• Nvidia Driver with CUDA >= 11.2 support (≥460.x)

Quick Installation

Install stable version using PyPi

See in PyPi

pip install super-gradients

That's it !

Install using GitHub

pip install git+https://github.com/Deci-AI/super-gradients.git@stable

Computer Vision Models - Pretrained Checkpoints

Pretrained Classification PyTorch Checkpoints

Model	Dataset	Resolution	Top-1	Top-5	Latency (HW)*T4	Latency (Production)**T4	Latency (HW)*Jetson Xavier NX	Latency (Production)**Jetson Xavier NX	Latency Cascade Lake
ViT base	ImageNet21K	224x224	84.15	-	4.46ms	4.60ms	- *	-	57.22ms
ViT large	ImageNet21K	224x224	85.64	-	12.81ms	13.19ms	- *	-	187.22ms
BEiT	ImageNet21K	224x224	-	-	-ms	-ms	- *	-	-ms
EfficientNet B0	ImageNet	224x224	77.62	93.49	0.93ms	1.38ms	- *	-	3.44ms
RegNet Y200	ImageNet	224x224	70.88	89.35	0.63ms	1.08ms	2.16ms	2.47ms	2.06ms
RegNet Y400	ImageNet	224x224	74.74	91.46	0.80ms	1.25ms	2.62ms	2.91ms	2.87ms
RegNet Y600	ImageNet	224x224	76.18	92.34	0.77ms	1.22ms	2.64ms	2.93ms	2.39ms
RegNet Y800	ImageNet	224x224	77.07	93.26	0.74ms	1.19ms	2.77ms	3.04ms	2.81ms
ResNet 18	ImageNet	224x224	70.6	89.64	0.52ms	0.95ms	2.01ms	2.30ms	4.56ms
ResNet 34	ImageNet	224x224	74.13	91.7	0.92ms	1.34ms	3.57ms	3.87ms	7.64ms
ResNet 50	ImageNet	224x224	81.91	93.0	1.03ms	1.44ms	4.78ms	5.10ms	9.25ms
MobileNet V3_large-150 epochs	ImageNet	224x224	73.79	91.54	0.67ms	1.11ms	2.42ms	2.71ms	1.76ms
MobileNet V3_large-300 epochs	ImageNet	224x224	74.52	91.92	0.67ms	1.11ms	2.42ms	2.71ms	1.76ms
MobileNet V3_small	ImageNet	224x224	67.45	87.47	0.55ms	0.96ms	2.01ms *	2.35ms	1.06ms
MobileNet V2_w1	ImageNet	224x224	73.08	91.1	0.46 ms	0.89ms	1.65ms *	1.90ms	1.56ms

NOTE:

• Latency (HW)* - Hardware performance (not including IO)
  
• Latency (Production)** - Production Performance (including IO)
• Performance measured for T4 and Jetson Xavier NX with TensorRT, using FP16 precision and batch size 1
• Performance measured for Cascade Lake CPU with OpenVINO, using FP16 precision and batch size 1

Pretrained Object Detection PyTorch Checkpoints

Model	Dataset	Resolution	mAPval 0.5:0.95	Latency (HW)*T4	Latency (Production)**T4	Latency (HW)*Jetson Xavier NX	Latency (Production)**Jetson Xavier NX	Latency Cascade Lake
SSD lite MobileNet v2	COCO	320x320	21.5	0.77ms	1.40ms	5.28ms	6.44ms	4.13ms
SSD lite MobileNet v1	COCO	320x320	24.3	1.55ms	2.84ms	8.07ms	9.14ms	22.76ms
YOLOX nano	COCO	640x640	26.77	2.47ms	4.09ms	11.49ms	12.97ms	-
YOLOX tiny	COCO	640x640	37.18	3.16ms	4.61ms	15.23ms	19.24ms	-
YOLOX small	COCO	640x640	40.47	3.58ms	4.94ms	18.88ms	22.48ms	-
YOLOX medium	COCO	640x640	46.4	6.40ms	7.65ms	39.22ms	44.5ms	-
YOLOX large	COCO	640x640	49.25	10.07ms	11.12ms	68.73ms	77.01ms	-

NOTE:

• Latency (HW)* - Hardware performance (not including IO)
  
• Latency (Production)** - Production Performance (including IO)
• Latency performance measured for T4 and Jetson Xavier NX with TensorRT, using FP16 precision and batch size 1
• Latency performance measured for Cascade Lake CPU with OpenVINO, using FP16 precision and batch size 1

Pretrained Semantic Segmentation PyTorch Checkpoints

Model	Dataset	Resolution	mIoU	Latency b1T4	Latency b1T4 including IO
PP-LiteSeg B50	Cityscapes	512x1024	76.48	4.18ms	31.22ms
PP-LiteSeg B75	Cityscapes	768x1536	78.52	6.84ms	33.69ms
PP-LiteSeg T50	Cityscapes	512x1024	74.92	3.26ms	30.33ms
PP-LiteSeg T75	Cityscapes	768x1536	77.56	5.20ms	32.28ms
DDRNet 23 slim	Cityscapes	1024x2048	78.01	5.74ms	32.01ms
DDRNet 23	Cityscapes	1024x2048	80.26	12.74ms	39.01ms
STDC 1-Seg50	Cityscapes	512x1024	75.11	3.34ms	30.12ms
STDC 1-Seg75	Cityscapes	768x1536	77.8	5.53ms	32.490ms
STDC 2-Seg50	Cityscapes	512x1024	76.44	4.12ms	30.94ms
STDC 2-Seg75	Cityscapes	768x1536	78.93	6.95ms	33.89ms
RegSeg (exp48)	Cityscapes	1024x2048	78.15	12.03ms	38.91ms
Larger RegSeg (exp53)	Cityscapes	1024x2048	79.2	22.00ms	48.96ms

NOTE: Performance measured on T4 GPU with TensorRT, using FP16 precision and batch size 1 (latency), and not including IO

NOTE: For resolutions below 1024x2048 we first resize the input to the inference resolution and then resize the predictions to 1024x2048. The time of resizing is included in the measurements so that the practical input-size is 1024x2048.

Implemented Model Architectures

Image Classification

• DensNet (Densely Connected Convolutional Networks) - Densely Connected Convolutional Networks https://arxiv.org/pdf/1608.06993.pdf
• DPN - Dual Path Networks https://arxiv.org/pdf/1707.01629
• EfficientNet - https://arxiv.org/abs/1905.11946
• GoogleNet - https://arxiv.org/pdf/1409.4842
• LeNet - https://yann.lecun.com/exdb/lenet/
• MobileNet - Efficient Convolutional Neural Networks for Mobile Vision Applications https://arxiv.org/pdf/1704.04861
• MobileNet v2 - https://arxiv.org/pdf/1801.04381
• MobileNet v3 - https://arxiv.org/pdf/1905.02244
• PNASNet - Progressive Neural Architecture Search Networks https://arxiv.org/pdf/1712.00559
• Pre-activation ResNet - https://arxiv.org/pdf/1603.05027
• RegNet - https://arxiv.org/pdf/2003.13678.pdf
• RepVGG - Making VGG-style ConvNets Great Again https://arxiv.org/pdf/2101.03697.pdf
• ResNet - Deep Residual Learning for Image Recognition https://arxiv.org/pdf/1512.03385
• ResNeXt - Aggregated Residual Transformations for Deep Neural Networks https://arxiv.org/pdf/1611.05431
• SENet - Squeeze-and-Excitation Networkshttps://arxiv.org/pdf/1709.01507
• ShuffleNet - https://arxiv.org/pdf/1707.01083
• ShuffleNet v2 - Efficient Convolutional Neural Network for Mobile Deviceshttps://arxiv.org/pdf/1807.11164
• VGG - Very Deep Convolutional Networks for Large-scale Image Recognition https://arxiv.org/pdf/1409.1556

Object Detection

• CSP DarkNet
• DarkNet-53
• SSD (Single Shot Detector) - https://arxiv.org/pdf/1512.02325
• YOLOX - https://arxiv.org/abs/2107.08430

Semantic Segmentation

• PP-LiteSeg - https://arxiv.org/pdf/2204.02681v1.pdf
• DDRNet (Deep Dual-resolution Networks) - https://arxiv.org/pdf/2101.06085.pdf
• LadderNet - Multi-path networks based on U-Net for medical image segmentation https://arxiv.org/pdf/1810.07810
• RegSeg - Rethink Dilated Convolution for Real-time Semantic Segmentation https://arxiv.org/pdf/2111.09957
• ShelfNet - https://arxiv.org/pdf/1811.11254
• STDC - Rethinking BiSeNet For Real-time Semantic Segmentation https://arxiv.org/pdf/2104.13188

Training Recipes

We defined recipes to ensure that anyone can reproduce our results in the most simple way.

Setup

To run recipes you first need to clone the super-gradients repository:

git clone https://github.com/Deci-AI/super-gradients

You then need to move to the root of the clone project (where you find "requirements.txt" and "setup.py") and install super-gradients:

pip install -e .

Finally, append super-gradients to the python path: (Replace "YOUR-LOCAL-PATH" with the path to the downloaded repo)

export PYTHONPATH=$PYTHONPATH:<YOUR-LOCAL-PATH>/super-gradients/

How to run a recipe

The recipes are defined in .yaml format and we use the hydra library to allow you to easily customize the parameters. The basic basic syntax is as follow:

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=<CONFIG-NAME> dataset_params.data_dir=<PATH-TO-DATASET>

But in most cases you will want to train on multiple GPU's using this syntax:

python -m torch.distributed.launch --nproc_per_node=<N-NODES> src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=<CONFIG-NAME> dataset_params.data_dir=<PATH-TO-DATASET>

Note: this script needs to be launched from the root folder of super_gradients Note: if you stored your dataset in the path specified by the recipe you can drop "dataset_params.data_dir=".

Explore our recipes

You can find all of our recipes here. You will find information about the performance of a recipe as well as the command to execute it in the header of its config file.

Example: Training of YoloX Small on Coco 2017, using 8 GPU

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_s dataset_params.data_dir=/home/coco2017

List of commands

All the commands to launch the recipes described here are listed below. Please make to "dataset_params.data_dir=" if you did not store the dataset in the path specified by the recipe (as showed in the example above).

- Classification

Cifar10

resnet:

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cifar10_resnet +experiment_name=cifar10

ImageNet

efficientnet

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_efficientnet

mobilenetv2

python -m torch.distributed.launch --nproc_per_node=2 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_mobilenetv2

mobilenetv3 small

python -m torch.distributed.launch --nproc_per_node=2 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_mobilenetv3_small

mobilenetv3 large

python -m torch.distributed.launch --nproc_per_node=2 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_mobilenetv3_large

regnetY200

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_regnetY architecture=regnetY200

regnetY400

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_regnetY architecture=regnetY400

regnetY600

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_regnetY architecture=regnetY600

regnetY800

python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_regnetY architecture=regnetY800

repvgg

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_repvgg

resnet50

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_resnet50

resnet50_kd

python -m torch.distributed.launch --nproc_per_node=8  src/super_gradients/examples/train_from_kd_recipe_example/train_from_kd_recipe.py --config-name=imagenet_resnet50_kd

vit_base

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_vit_base

vit_large

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=imagenet_vit_large

- Detection

Coco2017

ssd_lite_mobilenet_v2

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_ssd_lite_mobilenet_v2

yolox_n

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_n

yolox_t

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_t

yolox_s

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_s

yolox_m

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_m

yolox_l

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_l

yolox_x

python -m torch.distributed.launch --nproc_per_node=8 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolox architecture=yolox_x

- Segmentation

Cityscapes

DDRNet23

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_ddrnet

DDRNet23-Slim

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_ddrnet architecture=ddrnet_23_slim

RegSeg48

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_regseg48

STDC1-Seg50

python -m torch.distributed.launch --nproc_per_node=2 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_stdc_seg50

STDC2-Seg50

python -m torch.distributed.launch --nproc_per_node=2 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_stdc_seg50 architecture=stdc2_seg

STDC1-Seg75

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_stdc_seg75

STDC2-Seg75

python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_stdc_seg75 external_checkpoint_path=<stdc2-backbone-pretrained-path> architecture=stdc2_seg

Documentation

Check SuperGradients Docs for full documentation, user guide, and examples.

Contributing

To learn about making a contribution to SuperGradients, please see our Contribution page.

Our awesome contributors:

Made with contrib.rocks.

Citation

If you are using SuperGradients library or benchmarks in your research, please cite SuperGradients deep learning training library.

Community

If you want to be a part of SuperGradients growing community, hear about all the exciting news and updates, need help, request for advanced features, or want to file a bug or issue report, we would love to welcome you aboard!

• Slack is the place to be and ask questions about SuperGradients and get support. Click here to join our Slack

• To report a bug, file an issue on GitHub.

• Join the SG Newsletter for staying up to date with new features and models, important announcements, and upcoming events.

• For a short meeting with us, use this link and choose your preferred time.

License

This project is released under the Apache 2.0 license.

---

Deci Platform

Deci Platform is our end to end platform for building, optimizing and deploying deep learning models to production.

Sign up for our FREE Community Tier to enjoy immediate improvement in throughput, latency, memory footprint and model size.

Features:

• Automatically compile and quantize your models with just a few clicks (TensorRT, OpenVINO).
• Gain up to 10X improvement in throughput, latency, memory and model size.
• Easily benchmark your models’ performance on different hardware and batch sizes.
• Invite co-workers to collaborate on models and communicate your progress.
• Deci supports all common frameworks and Hardware, from Intel CPUs to Nvidia's GPUs and Jetsons.

Sign up for Deci Platform for free here