PaddleSeg Versions Save

Easy-to-use image segmentation library with awesome pre-trained model zoo, supporting wide-range of practical tasks in Semantic Segmentation, Interactive Segmentation, Panoptic Segmentation, Image Matting, 3D Segmentation, etc.

v2.9.0

6 months ago

New Project

  • Add support for multi-label segmentation, including code for data conversion and result visualization. (#3479)
  • Add quantitative awareness training support for PP-LiteSeg, PP-MobileSeg, OCRNet, and SegFormer-B0 to improve model inference speed. (#3457)
  • Add MobileSAM model to achieve faster SAM inference (#3346)

New feature

  • Add support for mlu, npu, xpu. (#3422)
  • Add more devices for TIPC including mlu, npu and xpu. (#3459)

Bug fixes

  • Solve the model accuracy problem of pp_mobileseg. (#3443)
  • Suppress Paddle 0-dimensional tensor warning. (#3423)
  • Replace concat with stack to solve the 0-dimensional tensor problem. (#3356)
  • Remove fluid API in rvm. (#3417)
  • Fixed multiple problems after stride is supported. (#3412, #3414, #3413)
  • Fix tool/model/analyze_model.py analysis flops problem. (#3404)
  • Fix image data reading problem, modify paddleseg/transforms/transforms.py. (#3377)
  • Fix CityscapesSOTA export error. (#3358)
  • Fix auc_roc configuration in predict. (#3446)
  • Remove invalid parameter num_class in losses/focal_loss.py. (#3451)
  • Fixed syntax issues in nnunet/infer.py. (#3460)
  • Optimize labelme2seg code and corresponding documentation. (#3497)

Documentation fix

  • Fixed grammatical errors in some Readme documents. (#3463)
  • Fixed invalid URL in doc/video.md.(#3452)
  • Add data set alias SegDataset.(#3427)
  • Update python version configuration of npu-tipc.(#3402)
  • Fix the device's configuration in TIPC and solve the mismatch problem between configuration and script function.(#3393)
  • Modify model_export_cn.md and add the steps to export dynamic_shape.pdtxt.(#3378)

New Contributors

PaddleSeg v2.9.0

新项目

  • 增加对多标签分割的支持、并提供数据转换代码及结果可视化。（#3479）
  • 对PP-LiteSeg、PP-MobileSeg、OCRNet、SegFormer-B0增加量化训练压缩功能，提升推理速度。（#3457）
  • 增加MobileSAM模型，实现更快速的SAM推理。（#3346）

新功能

  • 增加对mlu、npu、xpu的支持。（#3422）
  • 增加基于mlu、npu、xpu进行TIPC。（#3459）

Bug修复

  • 解决pp_mobileseg的模型精度问题。（#3443）
  • 抑制Paddle 0维tensor警告。（#3423）
  • 将concat替换为stack解决0维tensor问题。（#3356）
  • 移除rvm中的fluid api。（#3417）
  • 修复支持stride后，出现OOM、slice、split等共享显存降速、动转静检查误报问题。(#3412, #3414, #3413)
  • 修复tool/model/analyze_model.py中flops的问题。（#3404）
  • 修复image数据读入问题，修改paddleseg/transforms/transforms.py。（#3377）
  • 修复CityscapesSOTA导出错误。（#3358）
  • 修正predict中的auc_roc配置。（#3446）
  • 去除losses/focal_loss.py中无效参数num_class。(#3451)
  • 修正nnunet/infer.py中的语法问题。（#3460）
  • 优化labelme2seg的代码和对应文档。（#3497）

文档修复

  • 修正部分Readme文档的语法错误。（#3463）
  • 修正doc/video.md中无效url。（#3452）
  • 增加数据集别名SegDataset。（#3427）
  • 更新npu-tipc的python版本配置。（#3402）
  • 修复TIPC中device的配置，解决配置与脚本功能不匹配问题。 (#3393)
  • 修改model_export_cn.md，添加导出dynamic_shape.pdtxt的操作步骤。 (#3378)

Full Changelog: https://github.com/PaddlePaddle/PaddleSeg/compare/v2.8.0...v2.9.0

v2.8.0

1 year ago

New Features

Vision Foundation Model

  • Release Segment Anything Model (SAM) based on PaddlePaddle. As a vision foundation model, SAM has the powerful zero-shot capability to segment any objects and images. SAM can also segment specified object with prompt input.
  • Provide a gradio-based demo, which can be easily deployed to demonstrate the function of automatic full-image segmentation.
  • Provide a script-based demo, which segments specific objects with a point, box, or mask as prompt input.

Semantic Segmentation

  • Release PP-MobileSeg, a lightweight semantic segmentation model for mobile devices. Comparing PP-MobileSeg with the previous SOTA model on ADE20K dataset, the accuracy is increased by 1.5%, the speed is increased by 42.3%, and the number of parameters is reduced by 34.9%.
  • Add 7 semantic segmentation models, i.e., MaskFormer, HRFormer, ViT-Adapter, CAE, SegNeXt, K-Net, and LPSNet.
  • Enhance model training modules: Add Exponential Moving Average (EMA); refactor the optimizer as a customizable component; decouple Config from Builder, and strictly verify configuration information; move the user scripts into the tools directory.
  • Enhance model deployment modules: Add FastDeploy, a high-performance and all-scenario model deployment solution; add examples and tutorials for C++ deployment on Windows.

Panoptic Segmentation

  • Release PanopticSeg v0.5, a universal panoptic segmentation solution.
  • Provide full-process development capabilities for panoptic segmentation scenes, and support functions such as dataset preparation, model training, model deployment, and visual analysis.
  • Integrate Mask2Former and Panoptic-DeepLab models, and support Cityscapes and MS COCO datasets.

Quality Inspector

  • Release QualityInspector v0.5, a full-process solution for industrial quality inspection.
  • Support a unified and configurable pipeline that can flexibly use single-task and multi-task models, and integrate PaddleDetection and PaddleSeg models.
  • Provide 3 unsupervised quality inspection methods.
  • Support model evaluation and analysis functions, and one-click tuning by using the post-processing module.
  • Support functions such as data labeling, data analysis, and format conversion in industrial quality inspection scenes, and provide practical examples.

Others

  • Release EISeg v1.1, a semi-automatic tool for image annotation. Add manual labeling and automatic pre-labeling functions for detection objects, and support 3 dataset formats (COCO, VOC and YOLO).
  • Add a video matting model RVM, and support video matting and background replacement functions. Add a .NET deployment tutorial for matting models. Add DIY applications for ID photos and wedding photos based on PP-Matting.

Bug Fixes

  • Fix the precision error of multi-scale evaluation #2933 #2978
  • Fix the error of exporting the inference model for ESPNetV2 model #3003
  • Fix the error of repeatedly downloading datasets under multi GPUs #3126
  • Fix a bug in PortraitNet dataset #3024

新特性

视觉大模型

  • 开源飞桨版本视觉大模型Segment Anything Model (SAM)。SAM具有强大的zero-shot能力，可以分割任意物体和图像，也可以使用提示输入分割特定目标。
  • 提供基于Gradio的演示Demo，支持本地部署，可以快速体验SAM全图分割功能。
  • 提供脚本演示Demo，支持点、框、掩码作为提示输入，快速得到SAM交互分割结果。

语义分割

  • 发布超轻量级语义分割SOTA模型PP-MobileSeg，在ADE20K数据集上对比此前模型，精度提升1.5%、速度提升42.3%、参数量减少34.9%。
  • 新增7个前沿语义分割模型：MaskFormer、HRFormer、ViT-Adapter、CAE、SegNeXt、K-Net和LPSNet。
  • 增强训练功能：新增指数滑动平均EMA；支持自定义Optimizer组件，灵活配置训练超参；解耦Config和Builder，严格校验配置信息；训压推启动脚本统一到tools目录下。
  • 增强部署功能：新增支持高性能、全场景的模型部署方案FastDeploy；新增Windows上CPP部署的示例和教程。

全景分割

  • 发布通用的全景分割方案PanopticSeg v0.5版本。
  • 提供全景分割场景的全流程开发能力，打通数据集准备、模型训练、模型推理、可视化分析等功能。
  • 集成前沿模型Mask2Former和Panoptic-DeepLab，支持Cityscapes与MS COCO数据集。

工业质检

  • 发布工业质检全流程解决方案QualityInspector v0.5版本。
  • 支持统一可配置的算法方案，集成飞桨检测分割的能力和模型库，灵活使用单任务模型和多任务串联模型。
  • 支持三种无监督异常检测算法。
  • 支持工业级指标评测和分析功能，使用后处理模块可以一键调优。
  • 支持工业质检场景的数据标注、数据分析、格式转换等功能，提供全流程实践范例。

其他

  • 发布高性能智能标注工具EISeg v1.1版本，新增对检测目标的手工标注功能和自动预标注功能，支持COCO、VOC和YOLO的检测标注保存格式。
  • 新增视频抠图模型RVM，支持视频预测和背景替换；新增人像抠图.NET部署教程；新增基于PP-Matting的证件照/结婚照DIY应用。

Bug修复

  • 修复多尺度验证的精度错误 #2933 #2978
  • 修复ESPNetV2模型导出预测模型的错误 #3003
  • 修复多卡重复下载数据集的错误 #3126
  • 修复PortraitNet数据集的错误 #3024

v2.7.0

1 year ago

New Features

Semantic Segmentation

  • Release RTFormer, a real-time semantic segmentation model accepted by NeurIPS 2022. RTFormer combines the advantages of CNN and Transformer modules, and it achieves SOTA trade-off between performance and efficiency on several datasets.
  • Release UHRNet, a semantic segmentation model. The segmentation accuracy of UHRNet is higher than that of HRNet on Cityscapes.
  • Add 2 semantic segmentation models, i.e., TopFormer and MscaleOCRNet-PSA.
  • Enhance model training module, i.e., training for single channel images, setting different learning rate for backbone and head.
  • Add the tutorials of config preparation and training tricks.

Image Matting

  • Release PP-MattingV2, a real-time human matting model with SOTA performance. Compared to previous models, the mean error is reduced by 17.91%, the inference speed is improved by 44.6% on GPU.
  • Refine the tutorials and build the benchmark of Matting models.

3D Medical Segmentation

  • Release MedicalSegV2, a superior 3D medical image segmentation solution.
    • Release an intelligent annotation toolkit called EISeg-Med3D.
    • Release an optimized implementation of nnUNet named nnUNet-D, which has model deployment module.
    • Add 3 segmentation models, i.e., SwinUnet, TransUnet and nnFormer.
    • Refine the tutorials, add detailed information of model zoo and model introduction.

新特性

语义分割

  • 发布轻量级语义分割模型RTFormer，结合CNN和Transformer模块的优势，在公开数据集上实现性能SOTA，并发表于NeurIPS 2022。
  • 发布高精度语义分割模型UHRNet，Cityscapes数据集上精度超越HRNet。
  • 新增2个语义分割模型，TopFormer和MscaleOCRNet-PSA。
  • 增强训练功能，支持单通道图像训练，支持Backbone和Head设置不同学习率。
  • 优化安装步骤和文档，新增准备配置文件、高阶训练技巧的教程。

深度抠图

  • 发布自研的轻量级抠图模型PP-MattingV2，推理速度提升44.6%，平均误差减小17.91%，超越此前SOTA模型，支持零成本开箱即用。
  • 升级文档和教程，构建模型库Benchmark。

3D医疗分割

  • 发布3D医疗影像分割方案MedicalSegV2。
    • 发布3D医疗影像交互式标注工具EISeg-Med3D，具有算法创新、高效准确、用户友好等优势。
    • 新增前沿高精度分割方案nnUNet-D，涵盖数据分析、超参优化、模型构建、模型训练、模型融合等模块，而且新增模型部署的能力。
    • 新增3个医疗影像分割模型，SwinUnet、TransUnet和nnFormer，精度超过原论文最高达到3.6%。
    • 升级医疗分割的文档和教程，丰富ModelZoo和模型原理说明。

Bug Fix

  • Fix data transforms error in PanopticDeeplab. #2381
  • Fix evaluation error for cityscapes dataset in PanopticDeeplab. #2564
  • Replace _C_ops with _legacy_C_ops for basic api. #2494
  • Check and synchronize the num_classes in config. #2477
  • Replace with_softmax and with_argmax with output_op in export.py. #2547
  • Correct the use of dist.all_reduce in distributed training. #2606
  • Fix the error of releasing dataloader. #2650

v2.6.0

1 year ago

New Features

Semantic Segmentation

  • Release PP-HumanSeg v2, an off-the-shelf human segmentation model. It achieves 64.26 FPS on the mobile device, which is 45.5% faster than before.
  • Release PSSL, a novel pre-training method, including a large dataset that consists of 1.2M+ pseudo semantic segmentation labels (PSSL) corresponding to the whole ImageNet training set. It boosts the performances of various models on all downstream tasks.
  • Release the industrial model series: high-accuracy models, light-weight models, and super light-weight models, to help developers pick up the most suitable one.
  • Add 2 segmentation models: MobileNetV3_LRASPP，UperNet.
  • Add 1 initialization method: Xavier Uniform.
  • Upgrade data reading pipeline that allows using dict to pass the data.
  • Support PaddleSMRT which is a model selection tool that help developers to choose the best model according to the actual requirements.
  • Upgrade the homepage, and provide more easy-to-use quick-start tutorial.

Intelligent Labelling

  • Release EISeg v1.0, the stable-version semi-automatic tool for image, video and 3D slice data annotation. It achieves "Once for All" (training once, and labelling all) performance.
  • Add interactive video object segmentation for general scenes, this work is based on EISeg interactive segmentation model and MiVOS.
  • Add 3D segmentation capabilities for abdominal multi-organ and CT vertebral data, and provides 3D visualization tools.

Image Matting

  • Release PP-Matting source code and the pre-trained models.
  • Add the pymatting package that provides five traditional matting methods including ClosedFormMatting、KNNMatting, FastMatting, LearningBaseMatting, and RandomWalksMatting.
  • Add GCA model, update the ppmatting architecture, and support user-specified metrics evaluations.

3D Medical Segmentation

  • Add UNETR，we achieve Dice=71.8% in MSD-brain, which is 0.7% higher than the original implementation.
  • Add slicing window prediction to support large-scale input, which improves the inference accuracy.

新特性

语义分割

  • 发布实时人像分割模型PP-HumanSeg v2，移动端推理速度提升45.5%、达到64.26 FPS，分割精度更高、通用型更强、零成本开箱即用。
  • 发布120多万张ImageNet分割伪标签数据集，以及预训练方法PSSL，全面提升分割模型在下游任务上的性能。
  • 发布产业级语义分割模型，包括高精度、轻量级和超轻量级系列。
  • 新增2个语义分割模型，MobileNetV3_LRASPP，UperNet。
  • 新增1个初始化方法：Xavier Uniform。
  • 升级数据流程，通过字典形式进行数据的传递，提升数据流的可读性、灵活性与扩展性。
  • 接入飞桨产业模型选型工具PaddleSMRT，可以根据产业落地的诸多诉求，分析数据特点，推荐合适的模型和方案。
  • 全新升级文档主页，全流程使用教程更加详实。


智能标注

  • 发布高性能智能标注工具EISeg v1.0正式版，实现一次训练万物可标，加速提升图像、视频、3D医疗影像等领域的分割标注效率。
  • 新增用于通用场景视频交互式分割能力，以EISeg交互式分割模型及MiVOS算法为基础，全面提升视频标注体验。
  • 新增用于腹腔多器官及CT椎骨数据3D分割能力，并提供3D可视化工具，给予医疗领域3D标注新的思路。

深度抠图

  • 开源PP-Matting代码和预训练模型
  • 新增pymatting支持，引入ClosedFormMatting、KNNMatting、FastMatting、LearningBaseMatting和RandomWalksMatting传统机器学习算法。
  • 新增GCA模型，更新目录结构，支持指定指标进行评估。

3D医疗分割

  • 新增前沿模型UNETR，在MSD-brain 上Dice为71.8%，高于原论文0.7%。
  • 新增滑窗预测功能，支持大图推理提升精度。

Bug Fix

v2.5

2 years ago

New Features

  • Release PP-LiteSeg, a superior real-time semantic segmentation model. It achieves 273.6FPS on 1080Ti.
  • Release PP-Matting, a trimap-free image matting model for extremely fine-grained segmentation. It achieves SOTA performance on Composition-1k and Distinctions-646.
  • Release MedicalSeg, a newly easy-to-use toolkit for 3D medical imaging. It supports the whole segmentation process including data preprocessing, model training, and model deployment, and provides high-accuracy models on lung and spine segmentation.
  • Release EISeg v0.5, with three more interactive models for chest Xray, MRI spine and defect inspection.
  • Add 5 semantic segmentation models: ENet, CCNet, DDRNet, GloRe, PP-LiteSeg.
  • Add 1 loss function: MultiClassFocalLoss.
  • Support AMP training, including O1 and O2 levels.

新特性

  • 发布超轻量级语义分割模型PP-LiteSeg技术报告以及开源模型，在1080Ti GPU上推理速度达到273.6FPS。
  • 发布高精度抠图模型PP-Matting技术报告以及开源模型，在Composition-1K和Distinctions-646上实现SOTA。
  • 发布3D医疗影像开发套件MedicalSeg，支持数据预处理、模型训练、模型部署等全流程开发，并提供肺部、椎骨数据上的高精度分割模型。
  • 发布交互式分割工具EISeg v0.5版本，新增胸片X光、医学椎骨、工业质检标注垂类方向。
  • 新增5个分割模型：ENet, CCNet, DDRNet, GloRe, PP-LiteSeg。
  • 新增1个损失函数：MultiClassFocalLoss。
  • 完整支持混合精度训练能力，包括O1、O2模式训练、边训边评估。

Bug Fix

New Contributors

Full Changelog: https://github.com/PaddlePaddle/PaddleSeg/compare/v2.2.0...v2.5

v2.4

2 years ago

New Features

  • Release the upgraded interactive segmentation tool, EISeg 0.4, adding Remote Sensing and Medical annotation.
  • Publish the PP-HumanSeg paper for portrait segmentation, including semantic connectivity loss and large-scale video conferencing dataset named PP-HumanSeg14K.
  • Release PP-HumanMatting for extremely fine-grained human segmentation.
  • Provide 2 tutorials for domain adaptation algorithm PixMatch, and Lane Segmentation.
  • Add 9 semantic segmentation models: ESPNetV1, ESPNetV2, HRNet_W48_Contrast, DMNet, ENCNet, PFPNNet, FastFCN, BiSeNetV1, and SegMenter.
  • Add 2 loss functions: SECrossEntropyLoss，SemanticConnectivityLoss.
  • Add 1 transform method: RandomCenterCrop.
  • Add 4 medical segmentation datasets: STARE，DRIVE，CHASE DB1，HRF, and their pretrained models in UNet.
  • Provide a comprehensive performance table for all segmentation models in PaddleSeg.

新特性

  • 发布交互式分割工具EISeg v0.4版本，支持静态图预测，新增遥感、医疗标注垂类方向。
  • 发布人像分割论文PP-HumanSeg，并开源连通性学习（SCL）方法和大规模视频会议数据集PP-HumanSeg14K。
  • 开源人像发丝级抠图Matting算法，PP-HumanMatting。
  • 新增2个分割案例：无监督领域自适应模型PixMatch，车道线分割。
  • 新增9个分割模型：ESPNetV1, ESPNetV2, HRNet_W48_Contrast, DMNet, ENCNet, PFPNNet, FastFCN, BiSeNetV1, 以及Transformer系列模型SegMenter。
  • 新增2个损失函数：SECrossEntropyLoss，SemanticConnectivityLoss。
  • 新增1个数据增强方法：RandomCenterCrop。
  • 新增医疗分割数据集STARE，DRIVE，CHASE DB1，HRF，并提供UNet预训练模型。
  • 提供分割模型性能对比图，帮助用户全面了解分割模型性能，方便模型选型。

Bug Fix

New Contributors

Full Changelog: https://github.com/PaddlePaddle/PaddleSeg/compare/v2.2.0...v2.4

v2.3.0

2 years ago

新特性

  • 发布交互式分割SOTA算法论文，EdgeFlow。
  • 开源精细化分割Matting算法，DIM和MODNet。
  • 新增分割模型压缩高阶功能，蒸馏和量化。
  • 提供基于Paddle Inference的C++的分割模型预测指南。
  • 提供Paddle Servering部署和导出ONNX模型的示例和指南。
  • 新增经典模型SegNet，PointRend，图结构模型GINet，轻量级模型STDC，轻量级Transformer系列模型SegFormer。
  • 新增损失函数：RMI Loss，Focal Loss，KL Loss，Detail Aggregate Loss， Point CE Loss。
  • 支持自定义任意类别数量的color map，提升可视化效果。

问题修复

  • #1240 修复CrossEntropyLoss在加权情况下的值越界问题。
  • #1219 #1385 修复未训练完完整epoch退出时，dataloader随机抛出段错误的问题。
  • #1113 修复多进程dataloader在不同epoch下随机数种子相同的bug。

New Features

  • Published a paper on interactive segmentation named EdgeFlow.
  • Released two Matting algorithms, DIM and MODNet.
  • Provided advanced features on segmentation model compression, Knowledge Distillation and “Molde Quantization”.
  • Provided the model inference tutorial based on Paddle Inference and Paddle Serving.
  • Provided the ONNX exporting tutorial, which allows cross-platform deployment.
  • Added five models, SegNet, PointRend, GINet, STDC, SegFormer.
  • Added RMI Loss，Focal Loss，KL Loss，Detail Aggregate Loss， Point CE Loss.
  • Support customized color map.

Bug Fix

  • #1240 fix the problem of CrossEntropyLoss.
  • #1219 #1385 fix the segment problem of dataloader while exiting before a full epoch.
  • #1113 fix the problem of the same seed on multi-process dataloader.

v2.2.0

2 years ago

新特性

  • CVPR 2021 AutoNUE语义分割赛道 冠军方案 开源！
  • 全新开源的超轻量级人像分割模型PPSeg，基于自采的大规模半身人像数据训练，适用于 视频会议 等半身像场景
  • 新增交互式分割应用场景，基于seed-based SOTA模型RITM 提供了基于 人像 和 COCO+LVIS 训练的权重
  • 全新发布的交互式分割工具EISeg，可用于快速标注数据
  • 新增人像分割领域的经典模型 PortraitNet，新增Transformer系列 SOTA模型 SwinTransformer
  • 优化模型预测的后处理逻辑，提升模型预测精度

问题修复

  • #1123 修复模型剪枝时内存不足的问题
  • #1100 修复CrossEntropyLoss 使用 weight权重时训练无法收敛的问题
  • #1082 修复模型剪枝脚本运行失败的问题
  • #1081 修复了Windows系统下预测脚本输出保存路径不正确的问题
  • #1078 修复多卡训练模型时DataLoader未设置work_init_fn导致多进程中所使用同样的random seed的问题
  • #34c1bbf、#30860e 修复DecoupledSegNet和SFNet导出失败的问题

致谢

New Features

  • CVPR 2021 AutoNUE Semantic Segmentation Track Technical Report is open sourced!
  • An ultra-lightweight portrait segmentation model named PPSeg is open sourced, which is training based on large-scale portrait data and suitable for video conference
  • We provide interactive segmentation application scenarios, based on the seed-based SOTA model RITM and also provide weights training on portrait and COCO+LVIS.
  • A newly released interactive segmentation tool EISeg can be used to quickly label data
  • Added the popular model PortraitNet in the field of portrait segmentation, and added the Transformer series SOTA model SwinTransformer
  • We optimize the post-processing logic of model prediction to improve model prediction accuracy

Bug Fix

  • #1123 Fix the problem of insufficient memory during model pruning
  • #1100 Fix the problem that the weighted CrossEntropyLoss cannot converge during training
  • #1082 Fix the problem that the model pruning script fails to run
  • #1081 Fix the problem that the save path of the prediction script under Windows system is incorrect
  • #1078 Fix an issue where the DataLoader did not set work_init_fn when training the model in multi-card, which caused the same random seed to be used in multiple processes
  • #34c1bbf/#30860e Fix the problem that DecoupledSegNet and SFNet cannot be successfully exported

Thanks

v2.1.0

2 years ago

新特性

  • 语义分割方向新增医疗分割模型UNet3+、轻量级模型SFNet、ShuffleNetV2等模型。
  • 全新增加 全景分割 场景，支持训练、评估、预测以及可视化等能力，新增Anchor-Free的SOTA模型Panoptic-DeepLab。
  • 完善部署能力，新增 移动端部署 能力和 web端部署 能力，并支持添加后处理算子(argmax/softmax)。
  • 高精度的人像分割模型humanseg升级为动态图版，并显著优化边缘锯齿问题。
  • 升级学习率配置模块，新增10种学习率策略，涵盖了业界主流学习率调度方式。
  • 新增Weighted Cross Entropy Loss、L1 Loss、MSE Loss，适用于不同场景下的模型优化。

Bug修复

  • #1016 修复NonLocal2D模块在非gaussian模式下shape不一致的问题。
  • #1007 修复RandomRotation和RandomScaleAspect在未传入Label时无法正确调用的问题。
  • #1006 修复EMANet无法进行单卡训练的问题。
  • #995 修复了PaddleSeg在PaddlePaddle 2.1版本中存在的兼容性问题。
  • #980 修复DecoupledSegNet在PaddlePaddle 2.1版本中训练失败的问题。
  • #975 修复滑窗预测图像小于窗口大小时无法正确预测的问题。
  • #971 修复ResizeByLong进行数据增强，在预测时没有正确恢复尺寸的问题。

New Features

  • New semantic segmentation models such as the medical segmentation model UNet3+, the lightweight model SFNet, and ShuffleNetV2 have been added.
  • Newly added panoramic segmentation scenes, supporting training, evaluation, prediction and visualization capabilities, and new Anchor-Free SOTA model Panoptic-DeepLab.
  • Improve deployment capabilities, add mobile deployment and web deployment capabilities, and support the addition of post-processing operators (argmax/softmax).
  • The high-precision portrait segmentation model humanseg is upgraded to dynamic graph version, and the edge aliasing problem is significantly optimized.
  • Upgrade the learning rate configuration module and add 10 new learning rate strategies, covering the mainstream learning rate scheduling methods in the industry.
  • Added Weighted Cross Entropy Loss, L1 Loss, and MSE Loss, which are suitable for model optimization in different scenarios.

Bug Fix

  • #1016 Fix the problem that the shape of NonLocal2D module is inconsistent in non-gaussian mode.
  • #1007 Fixed an issue where RandomRotation and RandomScaleAspect could not be called correctly when Label was not passed in.
  • #1006 Fix the problem that EMANet cannot be trained in single card.
  • #995 Fixed the compatibility issue of PaddleSeg in PaddlePaddle 2.1 version.
  • #980 Fixed the problem that DecoupledSegNet failed to train in PaddlePaddle 2.1.
  • #975 Fix the problem that the sliding window prediction image cannot be correctly predicted when the image is smaller than the window size.
  • #971 Fix the problem that ResizeByLong does not restore the size correctly in predict phase.

v2.0.0

3 years ago

新特性

  • 全新发布2.0版本，全面升级至动态图，支持20+分割模型，4个骨干网络，5个数据集，9种Loss：
    • 分割模型：ANN、BiSeNetV2、DANet、DeeplabV3、DeeplabV3+、FCN、FastSCNN、Gated-scnn、GCNet、HarDNet、OCRNet、PSPNet、UNet、UNet++、U2Net、Attention UNet、Decoupled SegNet、EMANet、DNLNet、ISANet
    • 骨干网络：ResNet, HRNet, MobileNetV3, Xception
    • 数据集：Cityscapes, ADE20K, Pascal VOC, Pascal Context, COCO Stuff
    • Loss：CrossEntropy Loss、BootstrappedCrossEntropy Loss、Dice Loss、BCE Loss、OhemCrossEntropyLoss、RelaxBoundaryLoss、OhemEdgeAttentionLoss、Lovasz Hinge Loss、Lovasz Softmax Loss
  • 提供基于Cityscapes和Pascal Voc数据集的高质量预训练模型 50+
  • 支持多卡GPU并行评估，提供了高效的指标计算功能。支持多尺度评估/翻转评估/滑动窗口评估等多种评估方式。
  • 支持XPU模型训练，包括DeepLabv3、HRNet、UNet。
  • 开源了基于Hierarchical Multi-Scale Attention结构的语义分割模型，在Cityscapes验证集上达到87% mIoU。
  • 动态图模式支持模型在线量化、剪枝等模型压缩功能。
  • 动态图下支持模型动转静，实现高性能部署。

New Features

  • We newly released version 2.0 which has been fully upgraded to dynamic graphics. It supports more than 20 segmentation models, 4 backbone networks, , 5 datasets and 9 losses:
    • Segmentation models: ANN, BiSeNetV2, DANet, DeeplabV3, DeeplabV3+, FCN, FastSCNN, Gated-scnn, GCNet, HarDNet, OCRNet, PSPNet, UNet, UNet++, U2Net, Attention UNet, Decoupled SegNet, EMANet, DNLNet, ISANet
    • Backbone networks: ResNet, HRNet, MobileNetV3, and Xception
    • Datasets: Cityscapes, ADE20K, Pascal VOC, Pascal Context, COCO Stuff
    • Losses: CrossEntropy Loss, BootstrappedCrossEntropy Loss, Dice Loss, BCE Loss, OhemCrossEntropyLoss, RelaxBoundaryLoss, OhemEdgeAttentionLoss, Lovasz Hinge Loss, Lovasz Softmax Loss
  • We provide more than 50 high quality pre-trained models based on Cityscapes and Pascal Voc datasets.
  • The new version support multi-card GPU parallel evaluation for more efficient metrics calculation. It also support multiple evaluation methods such as multi-scale evaluation/flip evaluation/sliding window evaluation.
  • XPU model training including DeepLabv3, HRNet, UNet, is available now.
  • We open source a semantic segmentation model based on the Hierarchical Multi-Scale Attention structure, and it reached 87% mIoU on the Cityscapes validation set.
  • The dynamic graph mode supports model compression functions such as online quantification and pruning.
  • The dynamic graph mode supports model export for high-performance deployment.