深度学习全功能开发开源平台汇总

深度学习全功能开发平台

1. 微软 computervision-recipes平台

基本情况

框架：PyTorch

简介地址：

https://mp.weixin.qq.com/s/InFCT7CCGmHDly-IQcOqnA

开源地址：

https://github.com/microsoft/computervision-recipes

计算机视觉方向包括

图像分类、图像相似度计算、目标检测、物体及人体关键点检测、图像分割、动作识别、目标跟踪、拥挤人群计数等计算机视觉热门方向

Scenario	Support	Description
Classification	Base	Image Classification is a supervised machine learning technique to learn and predict the category of a given image.
Similarity	Base	Image Similarity is a way to compute a similarity score given a pair of images. Given an image, it allows you to identify the most similar image in a given dataset.
Detection	Base	Object Detection is a technique that allows you to detect the bounding box of an object within an image.
Keypoints	Base	Keypoint detection can be used to detect specific points on an object. A pre-trained model is provided to detect body joints for human pose estimation.
Segmentation	Base	Image Segmentation assigns a category to each pixel in an image.
Action recognition	Base	Action recognition to identify in video/webcam footage what actions are performed (e.g. “running”, “opening a bottle”) and at what respective start/end times. We also implemented the i3d implementation of action recognition that can be found under (contrib)[contrib].
Tracking	Base	Tracking allows to detect and track multiple objects in a video sequence over time.
Crowd counting	Contrib	Counting the number of people in low-crowd-density (e.g. less than 10 people) and high-crowd-density (e.g. thousands of people) scenarios.

2. 港中文、商汤科技MMLAB平台

基本情况

框架：PyTorch

简介地址：

https://zhuanlan.zhihu.com/p/160966882

https://mp.weixin.qq.com/s/nzhkVi0HT0wstewfUz4dWQ

开源地址：

https://github.com/open-mmlab

计算机视觉方向包括

• MMCV（基础支持）

更完善的训练流程支持，文件读取多后端支持，图片处理多后端支持，更丰富的 CNN 模块，20 种常用算子的高效 CUDA 实现。

• MMDetection（目标检测）

支持算法多达 40 个，300+ 预训练模型，速度和精度相比 1.0 版本有很大提升。通过更细粒度的模块化设计，MMDetection 的任务拓展性大大增强，成为了检测相关项目的基础平台。

• MMSegmentation（语义分割）

在统一的设定下，对 10+ 种语义分割算法的进行了统一 benchmark，并且达到了更高的精度。开源了 200+ 预训练模型和丰富的配置，方便进行横向纵向比较。支持丰富的训练和测试 trick，可以应对多样的使用场景。支持混合精度训练，节省显存 40%以上。

• MMDetection3D（3D目标检测）

MMDetection 的 3D 扩展，训练速度领先其他3D点云检测代码库。任务拓展性强，可作为3D检测相关任务的基础平台，支持单模态和多模态 3D 检测，同时支持室内和室外场景数据集SOTA。无缝使用 MMDetection 中的全部已有算法和模型，尽享丝滑用户体验。

• MMEditing（图像和视频编辑）

设计了统一的框架同时支持超分、修复、抠图、生成四大方向，方便用户在一个框架中调用不同的算法和模型。提供了丰富的底层视觉算法的高效算子，拥有高效的训练和测试速度，复现了多个未开源算法，并首次公开实现了基于 PyTorch 的 GAN 的分布式训练。

• MMAction2（动作识别）

支持 8 种模型和 8 个数据集，提供完善的数据处理脚本，支持多种解码器进行高效的在线视频解码，训练速度快，模型精度高。更细粒度的模块化设计，易于拓展。

• MMClassification（图像分类）

丰富的训练配置，支持常见网络的复现，并提供相应的预训练模型。

• MMPose（人体关键点检测）

首个同时支持 top-down 和 bottom-up 类型算法的开源人体姿态估计框架，并实现了目前学术界的最高训练精度和最快训练速度。基于模块化的设计，易于拓展和修改。后续将支持多人3D姿态估计、密集人群的姿态估计、人脸关键点等更多模块。

3. FaceBook Detectron2 平台

基本情况

框架：PyTorch

简介地址：

https://www.leiphone.com/news/201910/P2sq3UjmbsI6QdiW.html

开源地址：

https://github.com/facebookresearch/detectron2

计算机视觉方向包括

Projects by Facebook

Note that these are research projects, and therefore may not have the same level of support or stability as detectron2.

• DensePose: Dense Human Pose Estimation In The Wild
• Scale-Aware Trident Networks for Object Detection
• TensorMask: A Foundation for Dense Object Segmentation
• Mesh R-CNN
• PointRend: Image Segmentation as Rendering
• Momentum Contrast for Unsupervised Visual Representation Learning
• DETR: End-to-End Object Detection with Transformers

External Projects

External projects in the community that use detectron2:

• AdelaiDet, a detection toolbox including FCOS, BlendMask, etc.
• CenterMask
• Res2Net backbones
• VoVNet backbones
• FsDet, Few-Shot Object Detection.

4. 图森未来SimpleDet

基本情况

框架：MXNet

简介地址：

https://www.jiqizhixin.com/articles/2019-01-29-21

开源地址：

https://github.com/TuSimple/simpledet

Box, and Mask Detection

Model	Backbone	Head	Train Schedule	GPU	Image/GPU	FP16	Train MEM	Train Speed	Box AP(Mask AP)	Link
Faster	R50v1-C4	C5-512ROI	1X	8X 1080Ti	2	no	5.9G(4.5G)	20 img/s	34.2	model
Faster	R50v1-C4	C5-512ROI	1X	8X TitanV	2	yes	6.1G	49 img/s	34.4	model
Faster	R50v2-C4	C5-256ROI	1X	8X 1080Ti	2	no	5.1G	33 img/s	32.8	model
Cascade	R50v2-C5	2MLP	1X	8X 1080Ti	2	no	5.9G	25 img/s	38.8	model
Cascade	R50v1-FPN	2MLP	1X	8X 1080Ti	2	no	6.6G	21 img/s	40.3	model
Faster	R50v1-FPN	2MLP	1X	8X 1080Ti	2	no	4.2G(2.6G)	43 img/s	36.5	model
Mask	R50v1-FPN	2MLP+4CONV	1X	8X 1080Ti	2	no	5.7G(3.6G)	35 img/s	37.1(33.7)	model
Retina	R50v1-FPN	4Conv	1X	8X 1080Ti	2	no	4.7G(2.2G)	44 img/s	35.6	model
Trident	R50v2-C4	C5-128ROI	1X	8X 1080Ti	2	no	7.0G(5.3G)	20 img/s	37.1	model
Faster	R101v2-C4	C5-256ROI	1X	8X 1080Ti	2	no	6.7G	25 img/s	37.6	model
Faster-SyncBN	R101v2-C4	C5-256ROI	1X	8X 1080Ti	2	no	7.8G	17 img/s	38.6	model
Faster	R101v1-C4	C5-512ROI	1X	8X 1080Ti	2	no	10.2G	16 img/s	38.3	model
Faster	R101v1-C4	C5-512ROI	1X	8X TitanV	2	yes	7.0G	35 img/s	38.1	model
Faster	R101v1-FPN	2MLP	1X	8X 1080Ti	2	no	5.3G(3.4G)	31 img/s	38.7	model
Cascade	R101v2-C5	2MLP	1X	8X 1080Ti	2	no	7.6G	22 img/s	41.0	model
Cascade	R101v1-FPN	2MLP	1X	8X 1080Ti	2	no	8.7G	19 img/s	42.3	model
Trident	R101v2-C4	C5-128ROI	1X	8X 1080Ti	1	no	6.6G	9 img/s	40.6	model
Trident-Fast	R101v2-C4	C5-128ROI	1X	8X 1080Ti	1	no	6.6G	9 img/s	39.9	model
Retina	R101v1-FPN	4Conv	1X	8X 1080Ti	2	no	5.9G(3.0G)	31 img/s	37.8	model