Skip to content

Commit 0d30b43

Browse files
mandar2812mandar2812
committed
Added v1.4.1 release notes
1 parent 338fbae commit 0d30b43

File tree

2 files changed

+282
-0
lines changed

2 files changed

+282
-0
lines changed

‎docs/_data/sidebars/mydoc_sidebar.yml‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -43,6 +43,10 @@ entries:
4343
output: web, pdf
4444
folderitems:
4545

46+
- title: 1.4.1 Release notes
47+
url: /mydoc_release_notes_141.html
48+
output: web, pdf
49+
4650
- title: 1.4 Release notes
4751
url: /mydoc_release_notes_14.html
4852
output: web, pdf

‎docs/pages/mydoc/mydoc_release_notes_141.md‎

Lines changed: 278 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,278 @@
1+
---
2+
title: Release notes 1.4.1
3+
tags: [getting_started,release_notes]
4+
keywords: release notes, announcements, what's new, new features
5+
last_updated: March 26, 2017
6+
summary: "Version 1.4.1 of DynaML, released March 26, 2017, implements a number of new models (Extended Skew GP, student T process, generalized least squares, etc) and features."
7+
sidebar: mydoc_sidebar
8+
permalink: mydoc_release_notes_141.html
9+
folder: mydoc
10+
---
11+
12+
## Pipes API
13+
14+
### Additions
15+
16+
The pipes API has been vastly extended by creating pipes which encapsulate functions of multiple arguments leading to the following end points.
17+
18+
- `DataPipe2[A, B, C]`: Pipe which takes 2 arguments
19+
- `DataPipe3[A, B, C, D]` : Pipe which takes 3 arguments
20+
- `DataPipe4[A, B, C, D, E]`: Pipe which takes 4 arguments
21+
22+
Furthermore there is now the ability to create pipes which return pipes, something akin to curried functions in functional programming.
23+
24+
- `MetaPipe`: Takes an argument returns a `DataPipe`
25+
- `MetaPipe21`: Takes 2 arguments returns a `DataPipe`
26+
- `MetaPipe12`: Takes an argument returns a `DataPipe2`
27+
28+
29+
A new kind of Stream data pipe, `StreamFlatMapPipe` is added to represent data pipelines which can perform flat map like operations on streams.
30+
31+
32+
```scala
33+
val mapFunc: (I) => Stream[J] = ...
34+
val streamFMPipe = StreamFlatMapPipe(mapFunc)
35+
```
36+
37+
- Added Data Pipes API for Apache Spark RDDs.
38+
39+
```scala
40+
val num = 20
41+
val numbers = sc.parallelize(1 to num)
42+
val convPipe = RDDPipe((n: Int) => n.toDouble)
43+
44+
val sqPipe = RDDPipe((x: Double) => x*x)
45+
46+
val sqrtPipe = RDDPipe((x: Double) => math.sqrt(x))
47+
48+
val resultPipe = RDDPipe((r: RDD[Double]) => r.reduce(_+_).toInt)
49+
50+
val netPipeline = convPipe > sqPipe > sqrtPipe > resultPipe
51+
netPipeline(numbers)
52+
```
53+
54+
- Added `UnivariateGaussianScaler` class for gaussian scaling of univariate data.
55+
56+
57+
58+
59+
## Core API
60+
61+
### Additions
62+
63+
64+
**Package** `dynaml.models.bayes`
65+
66+
This new package will house stochastic prior models, currently there is support for GP and Skew GP priors, to see a starting example see `stochasticPriors.sc` in the `scripts` directory of the DynaML source.
67+
68+
----
69+
70+
**Package** `dynaml.kernels`
71+
72+
- Added `evaluateAt(h)(x,y)` and `gradientAt(h)(x,y)`; expressing `evaluate(x,y)` and `gradient(x,y)` in terms of them
73+
- Added `asPipe` method for Covariance Functions
74+
- For backwards compatibility users are advised to extend
75+
`LocalSVMKernel` in their custom Kernel implementations incase they do
76+
not want to implement the `evaluateAt` API endpoints.
77+
- Added [`FeatureMapKernel`](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/lm/SparkLogisticGLM.scala), representing kernels which can be explicitly decomposed into feature mappings.
78+
- Added Matern half integer kernel [`GenericMaternKernel[I]`](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/kernels/GenericMaternKernel.scala)
79+
- Added `block(S: String*)` method to block any hyper-parameters of kernels.
80+
- Added `NeuralNetworkKernel` and `GaussianSpectralKernel`.
81+
- Added `DecomposableCovariance`
82+
83+
```scala
84+
import io.github.mandar2812.dynaml.DynaMLPipe._
85+
import io.github.mandar2812.dynaml.kernels._
86+
87+
88+
implicit val ev = VectorField(6)
89+
implicit val sp = breezeDVSplitEncoder(2)
90+
implicit val sumR = sumReducer
91+
92+
val kernel = new LaplacianKernel(1.5)
93+
val other_kernel = new PolynomialKernel(1, 0.05)
94+
95+
val decompKernel = new DecomposableCovariance(kernel, other_kernel)(sp, sumReducer)
96+
97+
val other_kernel1 = new FBMKernel(1.0)
98+
99+
val decompKernel1 = new DecomposableCovariance(decompKernel, other_kernel1)(sp, sumReducer)
100+
101+
val veca = DenseVector.tabulate[Double](8)(math.sin(_))
102+
val vecb = DenseVector.tabulate[Double](8)(math.cos(_))
103+
104+
decompKernel1.evaluate(veca, vecb)
105+
106+
```
107+
108+
----
109+
110+
**Package** `dynaml.algebra`
111+
112+
Partitioned Matrices/Vectors and the following operations
113+
- Addition, Subtraction
114+
- Matrix, vector multiplication
115+
- LU, Cholesky
116+
- A\y, A\Y
117+
118+
Added calculation of quadratic forms, namely:
119+
120+
- `quadraticForm` which calculates x<sup>T</sup> A<sup>-1</sup> x
121+
- `crossQuadraticForm` which calculates y<sup>T</sup> A<sup>-1</sup> x
122+
123+
Where A is assumed to be a symmetric positive semi-definite matrix
124+
125+
Usage:
126+
127+
```scala
128+
import io.github.mandar2812.dynaml.algebra._
129+
130+
val x: DenseVector[Double] = ...
131+
val y: DenseVector[Double] = ...
132+
val a: DenseMatrix[Double] = ...
133+
134+
quadraticForm(a,x)
135+
crossQuadraticForm(y, a, x)
136+
137+
```
138+
139+
----
140+
141+
**Package** `dynaml.modelpipe`
142+
143+
New package created, moved all inheriting classes of `ModelPipe` to this package.
144+
145+
Added the following:
146+
147+
- `GLMPipe2` A pipe taking two arguments and returning a `GeneralizedLinearModel` instance
148+
- `GeneralizedLeastSquaresPipe2`:
149+
- `GeneralizedLeastSquaresPipe3`:
150+
151+
----
152+
153+
**Package** `dynaml.models`
154+
155+
- Added a new Neural Networks API: `NeuralNet` and `GenericFFNeuralNet`, for an example refer to `TestNNDelve` in `dynaml-examples`.
156+
- `GeneralizedLeastSquaresModel`: The [GLS](https://en.wikipedia.org/wiki/Generalized_least_squares) model.
157+
- `ESGPModel`: The implementation of a skew gaussian process regression model
158+
- [Warped Gaussian Process](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/gp/WarpedGPModel.scala) models **WIP**
159+
- Added mean function capability to Gaussian Process and Student T process models.
160+
- Added Apache Spark implementation of Generalized Linear Models; see [SparkGLM](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/lm/SparkGLM.scala), [SparkLogisticModel](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/lm/SparkLogisticGLM.scala), [SparkProbitGLM](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/lm/SparkLogisticGLM.scala)
161+
162+
----
163+
164+
**Package** `dynaml.probability`
165+
166+
- `MultivariateSkewNormal` as specified in [Azzalani et. al](https://arxiv.org/pdf/0911.2093.pdf)
167+
- `ExtendedMultivariateSkewNormal`
168+
- `UESN` and `MESN` representing an alternative formulation of the skew gaussian family from Adcock and Shutes.
169+
- `TruncatedGaussian`: Truncated version of the Gaussian distribution.
170+
- [Matrix Normal Distribution](https://en.wikipedia.org/wiki/Matrix_normal_distribution)
171+
- Added _Expectation_ operator for `RandomVariable` implementations in the `io.github.mandar2812.dynaml.probability` package object. Usage example given below.
172+
- `SkewGaussian`, `ExtendedSkewGaussian`: An breeze implementation of the SkewGaussian and extended Skew-Gaussian distributions respectively
173+
- `PushforwardMap`, `DifferentiableMap` added: `PushforwardMap` enables creating new random variables with defined density from base random variables.
174+
175+
``` scala
176+
import io.github.mandar2812.dynaml.analysis._
177+
import io.github.mandar2812.dynaml.probability._
178+
import io.github.mandar2812.dynaml.probability.distributions._
179+
180+
val g = GaussianRV(0.0, 0.25)
181+
val sg = RandomVariable(SkewGaussian(1.0, 0.0, 0.25))
182+
183+
//Define a determinant implementation for the Jacobian type (Double in this case)
184+
implicit val detImpl = identityPipe[Double]
185+
186+
//Defines a homeomorphism y = exp(x) x = log(y)
187+
val h: PushforwardMap[Double, Double, Double] = PushforwardMap(
188+
DataPipe((x: Double) => math.exp(x)),
189+
DifferentiableMap(
190+
(x: Double) => math.log(x),
191+
(x: Double) => 1.0/x)
192+
)
193+
194+
//Creates a log-normal random variable
195+
val p = h->g
196+
197+
//Creates a log-skew-gaussian random variable
198+
val q = h->sg
199+
200+
//Calculate expectation of q
201+
println("E[Q] = "+E(q))
202+
```
203+
- Added _Markov Chain Monte Carlo_ (MCMC) based inference schemes `ContinuousMCMC` and the underlying sampling implementation in `GeneralMetropolisHastings`.
204+
- Added implementation of _Approximate Bayesian Computation_ (ABC) in the `ApproxBayesComputation` class.
205+
206+
207+
``` scala
208+
//The mean
209+
val center: DenseMatrix[Double] = ...
210+
//Covariance (positive semi-def) matrix among rows
211+
val sigmaRows: DenseMatrix[Double] = ...
212+
//Covariance (positive semi-def) matrix among columns
213+
val sigmaCols: DenseMatrix[Double] = ...
214+
val matD = MatrixNormal(center, sigmaRows, sigmaCols)
215+
```
216+
- [Matrix T Distribution](https://en.wikipedia.org/wiki/Matrix_t-distribution) (_Experimental_)
217+
218+
``` scala
219+
//The degrees of freedom (must be > 2.0 for existence of finite moments)
220+
val mu: Double = ...
221+
//The mean
222+
val center: DenseMatrix[Double] = ...
223+
//Covariance (positive semi-def) matrix among rows
224+
val sigmaRows: DenseMatrix[Double] = ...
225+
//Covariance (positive semi-def) matrix among columns
226+
val sigmaCols: DenseMatrix[Double] = ...
227+
val matD = MatrixT(mu, center, sigmaCols, sigmaRows)
228+
```
229+
230+
----
231+
232+
**Package** `dynaml.optimization`
233+
234+
- Added `ProbGPCommMachine` which performs grid search or CSA and then instead of selecting a single hyper-parameter configuration calculates a weighted Gaussian Process committee where the weights correspond to probabilities or confidence on each model instance (hyper-parameter configuration).
235+
236+
**Package** `dynaml.utils`
237+
238+
- Added [multivariate gamma function](https://en.wikipedia.org/wiki/Multivariate_gamma_function)
239+
240+
``` scala
241+
//Returns logarithm of multivariate gamma function
242+
val g = mvlgamma(5, 1.5)
243+
```
244+
245+
----
246+
247+
**Package** `dynaml.dataformat`
248+
249+
- Added support for reading _MATLAB_ `.mat` files in the [`MAT`](https://github.com/transcendent-ai-labs/DynaML/blob/master/dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/dataformat/MAT.scala) object.
250+
251+
252+
### Improvements/Bug Fixes
253+
254+
**Package** `dynaml.probability`
255+
256+
- Removed `ProbabilityModel` and replaced with `JointProbabilityScheme` and `BayesJointProbabilityScheme`, major refactoring to `RandomVariable` API.
257+
258+
----
259+
260+
**Package** `dynaml.optimization`
261+
262+
- Improved logging of `CoupledSimulatedAnnealing`
263+
- Refactored `GPMLOptimizer` to `GradBasedGlobalOptimizer`
264+
265+
----
266+
267+
**Package** `dynaml.utils`
268+
- Correction to `utils.getStats` method used for calculating mean and variance of data sets consisting of `DenseVector[Double]`.
269+
- `minMaxScalingTrainTest` `minMaxScaling` of `DynaMLPipe` using `GaussianScaler` instead of `MinMaxScaler` for processing of features.
270+
271+
----
272+
273+
**Package** `dynaml.kernels`
274+
275+
- Fix to `CoRegCauchyKernel`: corrected mismatch of hyper-parameter string
276+
- Fix to `SVMKernel` objects matrix gradient computation in the case when kernel dimensions are not multiples of block size.
277+
- Correction to gradient calculation in RBF kernel family.
278+
- Speed up of kernel gradient computation, kernel and kernel gradient matrices with respect to the model hyper-parameters now calculated in a single pass through the data.

0 commit comments

Comments
 (0)