From 2b503dc5a86875046ded0cd8c52b6e63b56ddad8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=99=88=E9=94=AE=E9=A3=9E?= Date: Sat, 19 Dec 2020 22:41:23 +0800 Subject: [PATCH] add sigs/mdp/roadmap_2021.md. --- sigs/mdp/roadmap_2021.md | 69 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 69 insertions(+) create mode 100644 sigs/mdp/roadmap_2021.md diff --git a/sigs/mdp/roadmap_2021.md b/sigs/mdp/roadmap_2021.md new file mode 100644 index 0000000..032a163 --- /dev/null +++ b/sigs/mdp/roadmap_2021.md @@ -0,0 +1,69 @@ +Technical Roadmap for 2021 +==== + +Probabilistic Programming Layer +---- + +Distribution: +[D1] add support to multivariate continous distributions +[D1.1] MultivariateNormal +[D1.2] GumbelSoftmax +[D1.3] Dirichlet +Expected time: Q1 + +Modeling Language: +[M1] Based on the new support of MindSpore, extend the BayesianNet modeling capacity to general +[M1.1] Implement with dictionaies +[M1.2] Integrate StochasticTensor with Distribution +Expected time: Q1 + +[M2] Hybrid BayesianNet + Bijector modeling language +[M2.1] Implement a bijector API, and integrate it with the existing bijector in distributions +[M2.2] Implement affine coupling layer +[M2.3] Reimplement the existing Glow / VFlow in the test folder with the new API +Depends on: M1 +Expected time: Q3 + +Inference algorithm: +[I1] Support common black-box algorithms +[I1.1] Importance sampling +[I1.2] SGLD +[I1.3] Reinforce +[I1.4] HMC +Depends on: M1 +Expected time: [I1.1]:Q1, [I1.2]:Q2, [I1.3]:Q3, [I1.4]:Q4 + +Model and Toolbox Layer +---- + +[T1] Reimplement existing model and toolbox with the probabilistic programming layer +[T1.1] Reimplement VAE +[T1.2] Reimplement BNN +[T1.3] Reimplement Uncertainty Estimation +[T1.4] Reimplement Anormaly Detection +Depends on: M1 +Expected time: Q2 + +[T2] Implement more SOTA algorithms of toolbox +[T2.1] BNN with SGLD for uncertainty estimation +Depends on: M1, I1.2 +Expected time: Q2 +[T2.2] SWAG for uncertainty estimation +Depends on: D1.1, M1 +Expected time: Q3 + +[T3] Add more tools (e.g. generative models, anormaly detection) +[T3.1] Anormaly detection +Expected time: Q3 + +Appliation +---- + +[A1] Apply our tools to representative applications +Depends on: T2 +Expected time: Q4 + +Miscellanous +---- + +Keep interacting with MindSpore to improve the coding style. \ No newline at end of file -- Gitee