zafarali / better-sampling Goto Github PK
View Code? Open in Web Editor NEWinvestigating sampling
better-sampling's People
Contributors
Watchers
better-sampling's Issues
Performance issues: When training is off, the model takes much longer to give samples (incorrect use of volatile?)
We probably want to use set_grad_enabled from the 0.4 release: http://pytorch.org/docs/stable/autograd.html?highlight=grad#torch.autograd.set_grad_enabled
Update to PyTorch 0.4
Inconsistency between branches
From rw-bugfixing:
*********************************************
Sampler: ISSampler
Start Estimate: 1.80106, Variance: 0.0213128, Prop Success: 1 ESS: 409.786
KL(true|est)=0.00403086, KL(obs|est)=0.00392562
*********************************************
*********************************************
Sampler: ABCSampler
Start Estimate: 1.23824, Variance: 10.6266, Prop Success: 0.319
KL(true|est)=0.0253446, KL(obs|est)=0.0261764
*********************************************
*********************************************
Sampler: MCSampler
Start Estimate: 1.47959, Variance: 9.19856, Prop Success: 0.392
KL(true|est)=0.00221258, KL(obs|est)=0.0022405
*********************************************
*********************************************
Sampler: RVISampler
Start Estimate: 1.33802, Variance: 0.0218989, Prop Success: 0.909 ESS: 443.872
KL(true|est)=0.00589845, KL(obs|est)=0.00582546
*********************************************
From batch-RVI:
*********************************************
Sampler: ISSampler
Start Estimate: 1.80106, Variance: 0.0213128, Prop Success: 1 ESS: 409.786
KL(true|est)=0.00403086, KL(obs|est)=0.00392562
*********************************************
*********************************************
Sampler: ABCSampler
Start Estimate: 1.23824, Variance: 10.6266, Prop Success: 0.319
KL(true|est)=0.0253446, KL(obs|est)=0.0261764
*********************************************
*********************************************
Sampler: MCSampler
Start Estimate: 1.26426, Variance: 9.13137, Prop Success: 0.333
KL(true|est)=0.0116782, KL(obs|est)=0.011906
*********************************************
*********************************************
Sampler: RVISampler
Start Estimate: 1.59196, Variance: 0.0232969, Prop Success: 0.374 ESS: 364.28
KL(true|est)=0.0106005, KL(obs|est)=0.0105296
*********************************************
From mergebrv-rwb:
*********************************************
Sampler: ISSampler
Start Estimate: 1.80106, Variance: 0.0213128, Prop Success: 1 ESS: 409.786
KL(true|est)=0.00403086, KL(obs|est)=0.00392562
*********************************************
*********************************************
Sampler: ABCSampler
Start Estimate: 1.23824, Variance: 10.6266, Prop Success: 0.319
KL(true|est)=0.0253446, KL(obs|est)=0.0261764
*********************************************
*********************************************
Sampler: MCSampler
Start Estimate: 1.47959, Variance: 9.19856, Prop Success: 0.392
KL(true|est)=0.00221258, KL(obs|est)=0.0022405
*********************************************
*********************************************
Sampler: RVISampler
Start Estimate: 1.59196, Variance: 0.0232969, Prop Success: 0.374 ESS: 364.28
KL(true|est)=0.0106005, KL(obs|est)=0.0105296
*********************************************
Continuous Integration
Write tests for random walk
I am noticing a discrepancy between branches rw-bugfixing and batch-RVI
GPU support
Right now we have an opportunity to learn from multiple samples at the same time. This naturally would leverage GPUs. We should think about adding support for this.
cProfiling
- What are the bottlenecks of the code? Where are things slow? What can we optimize?
- https://docs.python.org/3/library/profile.html
Evaluate posterior at the end of training
Optimize Chi-squared objective rather than KL objective
See A.4
https://drive.google.com/file/d/1foEpoVJ7tsiqVGUoehZA93ZXcZp4Bykb/view?usp=sharing
However, now it will become a sparse reward problem
Add checkpointing
Scale walk to higher dimensions.
Use of GAE
To reduce the variance in the gradient updates we might be able to take advantage of using the GAE:
https://danieltakeshi.github.io/2017/04/02/notes-on-the-generalized-advantage-estimation-paper/
This would require us to switch to a neural network baseline.
Schulman, John, et al. "Gradient estimation using stochastic computation graphs." Advances in Neural Information Processing Systems. 2015.
Make ISSampler and MCSampler work in batch mode for more efficient sample collection
Use Learned Value function
Use a learned value function as a baseline. Does performance improve? do the training curves become less noisy?
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.