nb updated and added to examples

Author: shkr

RELEASE-NOTES.md

@@ -8,7 +8,7 @@
 - Improve NUTS initialization `advi+adapt_diag_grad` and add `jitter+adapt_diag_grad` (#2643) 
 - Update loo, new improved algorithm (#2730)
 - New CSG (Constant Stochastic Gradient) approximate posterior sampling
-  algorithm added
+  algorithm (#2544)
 ### Fixes 
 - Fixed `compareplot` to use `loo` output. 
 - Add test for `model.logp_array` and `model.bijection` (#2724) 
@@ -202,6 +202,7 @@ Taku Yoshioka <[email protected]>
 Peadar Coyle (springcoil) <[email protected]>
 Austin Rochford <[email protected]>
 Osvaldo Martin <[email protected]>
+Shashank Shekhar <[email protected]>
 
 In addition, the following community members contributed to this release:
 
@@ -250,7 +251,6 @@ Patricio Benavente <[email protected]>
 Raymond Roberts
 Rodrigo Benenson <[email protected]>
 Sergei Lebedev <[email protected]>
-Shashank Shekhar <[email protected]>
 Skipper Seabold <[email protected]>
 Thomas Kluyver <[email protected]>
 Tobias Knuth <[email protected]>

docs/source/examples.rst

@@ -78,3 +78,12 @@ Variational Inference
    notebooks/convolutional_vae_keras_advi.ipynb
    notebooks/empirical-approx-overview.ipynb
    notebooks/normalizing_flows_overview.ipynb
+   
+
+Stochastic Gradient
+===================
+
+.. toctree::
+   notebooks/constant_stochastic_gradient.ipynb
+   notebooks/sgfs_simple_optimization.ipynb
+   notebooks/bayesian_neural_network_with_sgfs.ipynb

docs/source/notebooks/constant_stochastic_gradient.ipynb

@@ -4,7 +4,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Introduction"
+    "# Constant Stochastic Gradient"
    ]
   },
   {
@@ -59,14 +59,8 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Sampling: Constant Stochastic Gradient"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Problem: A multivariate regression problem on the Protein Structure Properties dataset available at the [uci repo](https://archive.ics.uci.edu/ml/datasets/Physicochemical+Properties+of+Protein+Tertiary+Structure)."
+    "We will take a regression on a protein dataset that is used to show results in Figure 1. \n",
+    "It is a multivariate regression problem on the Protein Structure Properties dataset available at the [uci repo](https://archive.ics.uci.edu/ml/datasets/Physicochemical+Properties+of+Protein+Tertiary+Structure)."
    ]
   },
   {
@@ -378,7 +372,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## NUTS Trace Plot"
+    "### NUTS Trace Plot"
    ]
   },
   {
@@ -420,7 +414,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Preconditioned CSG Trace Plot"
+    "### Preconditioned CSG Trace Plot"
    ]
   },
   {
@@ -462,7 +456,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## SGFS Trace Plot"
+    "### SGFS Trace Plot"
    ]
   },
   {
@@ -504,7 +498,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Mean Absolute Error on Test Dataset"
+    "### Mean Absolute Error on Test Dataset"
    ]
   },
   {
@@ -622,7 +616,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Sample covariance projections on the smallest and largest components"
+    "### Sample covariance projections on the smallest and largest components"
    ]
   },
   {
@@ -760,7 +754,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Result"
+    "### Result"
    ]
   },
   {

docs/source/notebooks/sgfs_simple_optimization.ipynb

@@ -1,5 +1,12 @@
 {
  "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Stochastic Gradient Fisher Scoring"
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},
@@ -185,21 +192,21 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 2",
    "language": "python",
-   "name": "python3"
+   "name": "python2"
   },
   "language_info": {
    "codemirror_mode": {
     "name": "ipython",
-    "version": 3
+    "version": 2
    },
    "file_extension": ".py",
    "mimetype": "text/x-python",
    "name": "python",
    "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.6.1"
+   "pygments_lexer": "ipython2",
+   "version": "2.7.13"
   },
   "toc": {
    "colors": {

pymc3/sampling.py

@@ -10,7 +10,7 @@
 from .backends.base import BaseTrace, MultiTrace
 from .backends.ndarray import NDArray
 from .model import modelcontext, Point
-from .step_methods import (NUTS, HamiltonianMC, SGFS, CSG, Metropolis, BinaryMetropolis,
+from .step_methods import (NUTS, HamiltonianMC, Metropolis, BinaryMetropolis,
                            BinaryGibbsMetropolis, CategoricalGibbsMetropolis,
                            Slice, CompoundStep)
 from .util import update_start_vals
@@ -23,7 +23,7 @@
 
 __all__ = ['sample', 'iter_sample', 'sample_ppc', 'sample_ppc_w', 'init_nuts']
 
-STEP_METHODS = (NUTS, HamiltonianMC, SGFS, CSG, Metropolis, BinaryMetropolis,
+STEP_METHODS = (NUTS, HamiltonianMC, Metropolis, BinaryMetropolis,
                 BinaryGibbsMetropolis, Slice, CategoricalGibbsMetropolis)
 
 

pymc3/step_methods/sgmcmc.py

@@ -70,7 +70,7 @@ class BaseStochasticGradient(ArrayStepShared):
     variables of type `GeneratorOp`
 
     Parameters
-    -------
+    ----------
     vars : list
         List of variables for sampler
     batch_size`: int
@@ -206,7 +206,7 @@ class SGFS(BaseStochasticGradient):
     StochasticGradientFisherScoring
 
     Parameters
-    -----
+    ----------
     vars : list
         model variables
     B : np.array
@@ -216,10 +216,11 @@ class SGFS(BaseStochasticGradient):
         to the half of the previous step size
  
     References
-    -----
+    ----------
     -   Bayesian Posterior Sampling via Stochastic Gradient Fisher Scoring
         Implements Algorithm 1 from the publication http://people.ee.duke.edu/%7Elcarin/782.pdf
     """
+    name = 'stochastic_gradient_fisher_scoring'
 
     def __init__(self, vars=None, B=None, step_size_decay=100, **kwargs):
         """
@@ -323,17 +324,20 @@ class CSG(BaseStochasticGradient):
     discusses a proof for the optimal preconditioning matrix
     based on variational inference, so there is no parameter tuning required
     like in the case of 'B' matrix used for preconditioning in SGFS.
+    Take a look at this example notebook 
+    https://github.com/pymc-devs/pymc3/tree/master/docs/source/notebooks/constant_stochastic_gradient.ipynb
      
     Parameters
-    -----
+    ----------
     vars : list
         model variables
 
     References
-    -----
+    ----------
     -   Stochastic Gradient Descent as Approximate Bayesian Inference
         https://arxiv.org/pdf/1704.04289v1.pdf
     """
+    name = 'constant_stochastic_gradient'
 
     def __init__(self, vars=None, **kwargs):
         """