Changed to support TensorFlow Probability

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: sgmcmc
 Type: Package
 Title: Stochastic Gradient Markov Chain Monte Carlo
-Version: 0.2.2
+Version: 0.2.3
 Authors@R: c(
     person("Jack", "Baker", email = "[email protected]", role = c("aut", "cre", "cph")),
     person( "Christopher", "Nemeth", role = c("aut", "cph") ), 

R/setup.R

@@ -1,5 +1,7 @@
 # Create generic sgmcmc object
 createSGMCMC = function( logLik, logPrior, dataset, params, stepsize, minibatchSize, seed ) { 
+    # First check tf installation using tf_status. Throw error if it isn't installed.
+    checkTFInstall()
     # Set seed if required, TensorFlow seeds set inside dynamics
     if ( !is.null( seed ) ) {
         tf$set_random_seed(seed)
@@ -16,7 +18,7 @@ createSGMCMC = function( logLik, logPrior, dataset, params, stepsize, minibatchS
     # Check for tf$float64 errors (we only use tf$float32), if so throw a more explanatory error
     estLogPost = tryCatch({ 
         setupEstLogPost( logLik, logPrior, paramstf, placeholders, N, minibatchSize )
-    }, error = function ( e ) throwFloat64Error( e ) )
+    }, error = function ( e ) getPosteriorBuildError( e ) )
     # Check stepsize tuning constants are in list format
     stepsize = convertList( stepsize, params )
     # Declare sgmcmc object as list, return for custom tuning constants to be added by calling method
@@ -166,8 +168,3 @@ convertList = function( tuningConst, params ) {
     }
     return( convertedConsts )
 }
-
-# If tf$float64 error encountered, provide a more useful message
-throwFloat64Error = function( e ) {
-    stop(paste0("Problem building log posterior estimate from supplied logLik and logPrior functions. This is usually due to one of two things: a problem with the TensorFlow code used to declare logLik or logPrior functions; or that some constants declared in this function are read as type tf$float64. Maybe try to specify all constants as tf$float32. Full TensorFlow output below should make it clear if the issue is a float64 issue, or just a generic TensorFlow code issue.\n\n", e))
-}

R/tfErrors.R

@@ -0,0 +1,38 @@
+# Check tensorflow is installed. If it isn't throw an error.
+checkTFInstall <- function() {
+    if ( !get("TF", envir = tf_status) ) {
+        stop(tfErrorMsg(), call. = FALSE)
+    } else if ( !get("TFP", envir = tf_status) ) {
+        stop(tfpErrorMsg(), call. = FALSE)
+    }
+}
+
+# If there is an error building the posterior print a hopefully more readable error message
+getPosteriorBuildError <- function(e) {
+    stop(buildErrorMsg(e), call. = FALSE)
+}
+
+
+tfErrorMsg <- function() {
+    msg <- "\nNo TensorFlow python installation found.\n"
+    msg <- paste0(msg, "This can be installed using the installTF() function.\n")
+    return(msg)
+}
+
+
+tfpErrorMsg <- function() {
+    msg <- "\nNo TensorFlow Probability python installation found.\n"
+    msg <- paste0(msg, "This can be installed using the installTF() function.\n")
+    return(msg)
+}
+
+
+buildErrorMsg = function(e) {
+    msg <- "Problem building log posterior estimate from supplied logLik and logPrior functions.\n\n"
+    msg <- paste0(msg, "Python error output:\n", e)
+    msg <- paste0(msg, "\n", 
+            "Check your tensorflow code specifying the logLik and logPrior functions is correct.\n")
+    msg <- paste0(msg, "Ensure constants in logLik and logPrior functions are specified as ",
+            "type float32 using \ntf$constant(.., dtype = tf$float32) -- see the tutorials for some examples.")
+    return(msg)
+}

R/zzz.R

@@ -1,51 +1,52 @@
+# Environment determining status of the TensorFlow Installation.
+# This allows a custom error message to be displayed.
+tf_status <- new.env()
+
+
 # Load TensorFlow Probability and add the contents to tf$distributions.
 .onLoad <- function(libname, pkgname) {
-    # Check TensorFlow is installed and load tensorflow_probability
-    # If either are not installed display a custom install message
-    # Set tf$distributions to be tfp$distributions
-    tryCatch(tfp <- loadTF(),
-            error = function(e) {
-                tfMissing()
-            }
-    )
-
+    # Set default tf_status that everything is installed correctly.
+    assign("TF", TRUE, envir = tf_status)
+    assign("TFP", TRUE, envir = tf_status)
+    # Check TensorFlow is installed. Update tf_status accordingly.
+    checkTF()
+    # If checkTF was not successful, return to avoid printing multiple messages
+    if (!get("TF", envir = tf_status)) {
+        return()
+    }
+    # Check TensorFlow Probability is installed, and load in. Update tf_status accordingly.
+    tryCatch(loadTFP(), error = function(e) tfpMissing(e))
 }
 
 
-# Build message if TensorFlow Probability missing
-tfMissing <- function() {
-    message("\nNo TensorFlow or TensorFlow Probability python installation found.")
-    message("This can be installed using the installTF() function.\n")
-    # Set custom error message incase user still tries to use tf
-    assign("on_error", function (e) error_fn(e), env = tf)
+# Check tensorflow installed by doing a dummy operation that will throw an error
+checkTF = function() {
+    tryCatch(temp <- tf$constant(4),
+            error = function (e) tfMissing())
 }
 
 
-# Check TensorFlow installed and load TensorFlow probability
-loadTF = function() {
-    # Check tensorflow installed by doing a dummy operation that will throw an error
-    temp <- tf$constant(4)
-
-    # Delay load tensorflow_probability as tfp using reticulate package.
-    tfp <- reticulate::import("tensorflow_probability", delay_load = list(
-            priority = 5,
-            environment = "r-tensorflow"
-    ))
-
-    # Set tfp$distributions to be tf$distributions
+# Load tensorflow probability and assign distns to tf$distributions.
+# If this fails, print message and update tf_status
+loadTFP <- function() {
+    import_opts <- list(priority = 5, environment = "r-tensorflow")
+    tfp <- reticulate::import("tensorflow_probability", delay_load = import_opts)
     tf$distributions <- tfp$distributions
 }
 
 
-error_fn = function(e) {
-    stop(tfErrorMsg(), call. = FALSE)
+# Build message if TensorFlow missing. Update tf_status
+tfMissing <- function() {
+    message("\nNo TensorFlow python installation found.")
+    message("This can be installed using the installTF() function.\n")
+    assign("TF", FALSE, envir = tf_status)
+    assign("TFP", FALSE, envir = tf_status)
 }
 
 
-# Build error message for TensorFlow configuration errors
-tfErrorMsg <- function() {
-    message <- "Installation of TensorFlow or TensorFlow Probability not found.\n"
-    message <- paste0(message,
-            "These can be installed by running the installTF() function.")
-    return(message)
+# Build message if TensorFlow Probability missing. Update tf_status
+tfpMissing <- function(e) {
+    message("\nNo TensorFlow Probability python installation found.")
+    message("This can be installed using the installTF() function.\n")
+    assign("TFP", FALSE, envir = tf_status)
 }

README.md

@@ -5,15 +5,16 @@
 
 `sgmcmc` implements popular stochastic gradient Markov chain Monte Carlo (SGMCMC) methods including [stochastic gradient Langevin dynamics (SGLD)](http://people.ee.duke.edu/~lcarin/398_icmlpaper.pdf), [stochastic gradient Hamiltonian Monte Carlo (SGHMC)](https://arxiv.org/pdf/1402.4102v2.pdf) and [stochastic gradient Nosé-Hoover thermostat (SGNHT)](http://papers.nips.cc/paper/5592-bayesian-sampling-using-stochastic-gradient-thermostats.pdf). The package uses automatic differentiation, so all the differentiation needed for the methods is calculated automatically. Control variate methods can be used in order to improve the efficiency of the methods as proposed in the [recent publication](https://arxiv.org/pdf/1706.05439.pdf).
 
-The package is built on top of the [TensorFlow library for R](https://tensorflow.rstudio.com/), which has a lot of support for statistical distributions and operations, which allows a large class of posteriors to be built. For more details can be found at the [TensorFlow R library webpage](https://tensorflow.rstudio.com/), also see the [TensorFlow API](https://www.tensorflow.org/api_docs/) for full documentation.
+The package is built on top of the [TensorFlow library for R](https://tensorflow.rstudio.com/), which has a lot of support for statistical distributions and operations, which allows a large class of posteriors to be built. More details can be found at the [TensorFlow R library webpage](https://tensorflow.rstudio.com/), also see the [TensorFlow API](https://www.tensorflow.org/api_docs/) for full documentation.
 
 ## Installation
 
 `sgmcmc` requires [TensorFlow for R](https://github.com/rstudio/tensorflow) to be installed, which requires packages that can't be automatically built by `R`, so has a few steps:
-- Install [TensorFlow for R](https://github.com/rstudio/tensorflow), first run `install.packages("tensorflow")`, then after run `tensorflow::install_tensorflow()`.
-- Now run `install.packages("sgmcmc")`.
+- Install [TensorFlow for R](https://github.com/rstudio/tensorflow), run `install.packages("tensorflow")`.
+- Install the `sgmcmc` R package: `install.packages("sgmcmc")`.
+- Install the required python packages (including TensorFlow and TensorFlow Probability) by running: `sgmcmc::installTF()`.
 
-The TensorFlow API changes a lot, so it's best to make sure your TensorFlow for python installation is up to date.
+If you already have the TensorFlow and TensorFlow probability packages installed, then this should be autodetected by the package and you can skip the final step. Make sure these are up to date though, as the TensorFlow API is under active development and still changes quite regularly.
 
 ## Documentation
 

_pkgdown.yml

@@ -37,6 +37,10 @@ reference:
         - initSess
         - sgmcmcStep
         - getParams
+  - title: "Installation"
+    desc: "Install python packages depended on by sgmcmc, TensorFlow and TensorFlow Probability."
+    contents:
+        - installTF
   - title: "Datasets"
     desc: "Download and load datasets used in the examples and vignettes"
     contents: