public FunctionOutput <string>[] FastReplace(Excel.Range com, DAG dag, InputSample original, InputSample sample, AST.Address[] outputs, bool replace_original)
{
FunctionOutput <string>[] fo_arr;
if (!_d.TryGetValue(sample, out fo_arr))
{
// replace the COM value
ReplaceExcelRange(com, sample);
// initialize array
fo_arr = new FunctionOutput <string> [outputs.Length];
// grab all outputs
for (var k = 0; k < outputs.Length; k++)
{
// save the output
fo_arr[k] = new FunctionOutput <string>(dag.readCOMValueAtAddress(outputs[k]), sample.GetExcludes());
}
// Add function values to cache
// Don't care about return value
_d.Add(sample, fo_arr);
// restore the COM value
if (replace_original)
{
ReplaceExcelRange(com, original);
}
}
return(fo_arr);
}
public FunctionOutput<string>[] FastReplace(Excel.Range com, DAG dag, InputSample original, InputSample sample, AST.Address[] outputs, bool replace_original)
{
FunctionOutput<string>[] fo_arr;
if (!_d.TryGetValue(sample, out fo_arr))
{
// replace the COM value
ReplaceExcelRange(com, sample);
// initialize array
fo_arr = new FunctionOutput<string>[outputs.Length];
// grab all outputs
for (var k = 0; k < outputs.Length; k++)
{
// save the output
fo_arr[k] = new FunctionOutput<string>(dag.readCOMValueAtAddress(outputs[k]), sample.GetExcludes());
}
// Add function values to cache
// Don't care about return value
_d.Add(sample, fo_arr);
// restore the COM value
if (replace_original)
{
ReplaceExcelRange(com, original);
}
}
return fo_arr;
}
public static void ReplaceExcelRange(Range com, InputSample input)
{
bool done = false;
while (!done)
{
try
{
com.Value2 = input.GetInputArray();
done = true;
}
catch (Exception)
{
}
}
}
public static void ReplaceExcelRange(Range com, InputSample input)
{
bool done = false;
while (!done)
{
try
{
com.Value2 = input.GetInputArray();
done = true;
}
catch (Exception)
{
}
}
}
private static Dictionary <AST.Range, InputSample> StoreInputs(AST.Range[] inputs, DAG dag)
{
var d = new Dictionary <AST.Range, InputSample>();
foreach (AST.Range input_range in inputs)
{
var com = dag.getCOMRefForRange(input_range);
var s = new InputSample(com.Height, com.Width);
// store the entire COM array as a multiarray
// in one fell swoop.
s.AddArray(com.Range.Value2);
// add stored input to dict
d.Add(input_range, s);
// this is to force excel to recalculate its outputs
// exactly the same way that it will for our bootstraps
BootMemo.ReplaceExcelRange(com.Range, s);
}
return(d);
}
public static InputSample[] Resample(int num_bootstraps, InputSample orig_vals, Random rng)
{
// the resampled values go here
var ss = new InputSample[num_bootstraps];
// sample with replacement to get i bootstrapped samples
for (var i = 0; i < num_bootstraps; i++)
{
var s = new InputSample(orig_vals.Rows(), orig_vals.Columns());
// make a vector of index counters
var inc_count = new int[orig_vals.Length()];
// randomly sample j values, with replacement
for (var j = 0; j < orig_vals.Length(); j++)
{
// randomly select a value from the original values
int input_idx = rng.Next(0, orig_vals.Length());
inc_count[input_idx] += 1;
if (input_idx >= orig_vals.Length())
{
throw new Exception("input_idx >= orig_vals.Length()");
}
string value = orig_vals.GetInput(input_idx);
s.Add(value);
}
// indicate which indices are excluded
s.SetIncludes(inc_count);
// add the new InputSample to the output array
ss[i] = s;
}
return(ss);
}
public override bool Equals(object obj)
{
InputSample other = (InputSample)obj;
return(_includes.SequenceEqual(other.GetIncludes()));
}
// num_bootstraps: the number of bootstrap samples to get
// inputs: a list of inputs; each TreeNode represents an entire input range
// outputs: a list of outputs; each TreeNode represents a function
public static TreeScore DataDebug(int num_bootstraps,
DAG dag,
Excel.Application app,
bool weighted,
bool all_outputs,
long max_duration_in_ms,
Stopwatch sw,
double significance,
ProgBar pb)
{
// this modifies the weights of each node
PropagateWeights(dag);
// filter out non-terminal functions
var output_fns = dag.terminalFormulaNodes(all_outputs);
// filter out non-terminal inputs
var input_rngs = dag.terminalInputVectors();
// first idx: the index of the TreeNode in the "inputs" array
// second idx: the ith bootstrap
var resamples = new InputSample[input_rngs.Length][];
// RNG for sampling
var rng = new Random();
// we save initial inputs and outputs here
var initial_inputs = StoreInputs(input_rngs, dag);
var initial_outputs = StoreOutputs(output_fns, dag);
// Set progress bar max
pb.setMax(input_rngs.Length * 2);
#region RESAMPLE
// populate bootstrap array
// for each input range (a TreeNode)
for (int i = 0; i < input_rngs.Length; i++)
{
// this TreeNode
var t = input_rngs[i];
// resample
resamples[i] = Resample(num_bootstraps, initial_inputs[t], rng);
// update progress bar
pb.IncrementProgress();
}
#endregion RESAMPLE
#region INFERENCE
return Inference(
num_bootstraps,
resamples,
initial_inputs,
initial_outputs,
input_rngs,
output_fns,
dag,
weighted,
significance,
pb);
#endregion INFERENCE
}
private static Dictionary<AST.Range, InputSample> StoreInputs(AST.Range[] inputs, DAG dag)
{
var d = new Dictionary<AST.Range, InputSample>();
foreach (AST.Range input_range in inputs)
{
var com = dag.getCOMRefForRange(input_range);
var s = new InputSample(com.Height, com.Width);
// store the entire COM array as a multiarray
// in one fell swoop.
s.AddArray(com.Range.Value2);
// add stored input to dict
d.Add(input_range, s);
// this is to force excel to recalculate its outputs
// exactly the same way that it will for our bootstraps
BootMemo.ReplaceExcelRange(com.Range, s);
}
return d;
}
public static InputSample[] Resample(int num_bootstraps, InputSample orig_vals, Random rng)
{
// the resampled values go here
var ss = new InputSample[num_bootstraps];
// sample with replacement to get i bootstrapped samples
for (var i = 0; i < num_bootstraps; i++)
{
var s = new InputSample(orig_vals.Rows(), orig_vals.Columns());
// make a vector of index counters
var inc_count = new int[orig_vals.Length()];
// randomly sample j values, with replacement
for (var j = 0; j < orig_vals.Length(); j++)
{
// randomly select a value from the original values
int input_idx = rng.Next(0, orig_vals.Length());
inc_count[input_idx] += 1;
if (input_idx >= orig_vals.Length())
{
throw new Exception("input_idx >= orig_vals.Length()");
}
string value = orig_vals.GetInput(input_idx);
s.Add(value);
}
// indicate which indices are excluded
s.SetIncludes(inc_count);
// add the new InputSample to the output array
ss[i] = s;
}
return ss;
}
public static TreeScore Inference(
int num_bootstraps,
InputSample[][] resamples,
Dictionary<AST.Range, InputSample> initial_inputs,
Dictionary<AST.Address, string> initial_outputs,
AST.Range[] input_arr,
AST.Address[] output_arr,
DAG dag,
bool weighted,
double significance,
ProgBar pb)
{
// synchronization token
object lock_token = new Object();
// init thread event notification array
var mres = new ManualResetEvent[input_arr.Length];
// init job storage
var ddjs = new DataDebugJob[input_arr.Length];
// init started jobs count
var sjobs = 0;
// init completed jobs count
var cjobs = 0;
// last-ditch effort flag
bool last_try = false;
// init score storage
var scores = new TreeScore();
for (int i = 0; i < input_arr.Length; i++)
{
try
{
#region BOOTSTRAP
// bootstrapping is done in the parent STA thread because
// the .NET threading model prohibits thread pools (which
// are MTA) from accessing STA COM objects directly.
// alloc bootstrap storage for each output (f), for each resample (b)
FunctionOutput<string>[][] bs = new FunctionOutput<string>[initial_outputs.Count][];
for (int f = 0; f < initial_outputs.Count; f++)
{
bs[f] = new FunctionOutput<string>[num_bootstraps];
}
// init memoization table for input vector i
var memo = new BootMemo();
// fetch the input range TreeNode
var input = input_arr[i];
// fetch the input range COM object
var com = dag.getCOMRefForRange(input).Range;
// compute outputs
// replace the values of the COM object with the jth bootstrap,
// save all function outputs, and
// restore the original input
for (var b = 0; b < num_bootstraps; b++)
{
// lookup outputs from memo table; otherwise do replacement, compute outputs, store them in table, and return them
FunctionOutput<string>[] fos = memo.FastReplace(com, dag, initial_inputs[input], resamples[i][b], output_arr, false);
for (var f = 0; f < output_arr.Length; f++)
{
bs[f][b] = fos[f];
}
}
// restore the original inputs; faster to do once, after bootstrapping is done
BootMemo.ReplaceExcelRange(com, initial_inputs[input]);
// TODO: restore formulas if it turns out that they were overwrittern
// this should never be the case
#endregion BOOTSTRAP
#region HYPOTHESIS_TEST
// cancellation token
mres[i] = new ManualResetEvent(false);
// set up job
ddjs[i] = new DataDebugJob(
dag,
bs,
initial_outputs,
input_arr[i],
output_arr,
weighted,
significance,
mres[i]
);
sjobs++;
// hand job to thread pool
ThreadPool.QueueUserWorkItem(ddjs[i].threadPoolCallback, i);
#endregion HYPOTHESIS_TEST
// update progress bar
pb.IncrementProgress();
}
catch (System.OutOfMemoryException e)
{
if (!last_try)
{
// If there are no more jobs running, but
// we still can't allocate memory, try invoking
// GC and then trying again
cjobs = mres.Count(mre => mre.WaitOne(0));
if (sjobs - cjobs == 0)
{
GC.Collect();
last_try = true;
}
}
else
{
// we just don't have enough memory
throw e;
}
// wait for any of the 0..i-1 work items
// to complete and try again
WaitHandle.WaitAny(mres.Take(i).ToArray());
}
}
// Do not proceed until all hypothesis tests are done.
// WaitHandle.WaitAll cannot be called on an STA thread which
// is why we call WaitOne in a loop.
// Merge scores as data becomes available.
for (int i = 0; i < input_arr.Length; i++)
{
mres[i].WaitOne();
scores = DictAdd(scores, ddjs[i].Result);
}
return scores;
}
// num_bootstraps: the number of bootstrap samples to get
// inputs: a list of inputs; each TreeNode represents an entire input range
// outputs: a list of outputs; each TreeNode represents a function
public static TreeScore DataDebug(int num_bootstraps,
DAG dag,
Excel.Application app,
bool weighted,
bool all_outputs,
long max_duration_in_ms,
Stopwatch sw,
double significance,
ProgBar pb)
{
// this modifies the weights of each node
PropagateWeights(dag);
// filter out non-terminal functions
var output_fns = dag.terminalFormulaNodes(all_outputs);
// filter out non-terminal inputs
var input_rngs = dag.terminalInputVectors();
// first idx: the index of the TreeNode in the "inputs" array
// second idx: the ith bootstrap
var resamples = new InputSample[input_rngs.Length][];
// RNG for sampling
var rng = new Random();
// we save initial inputs and outputs here
var initial_inputs = StoreInputs(input_rngs, dag);
var initial_outputs = StoreOutputs(output_fns, dag);
// Set progress bar max
pb.setMax(input_rngs.Length * 2);
#region RESAMPLE
// populate bootstrap array
// for each input range (a TreeNode)
for (int i = 0; i < input_rngs.Length; i++)
{
// this TreeNode
var t = input_rngs[i];
// resample
resamples[i] = Resample(num_bootstraps, initial_inputs[t], rng);
// update progress bar
pb.IncrementProgress();
}
#endregion RESAMPLE
#region INFERENCE
return(Inference(
num_bootstraps,
resamples,
initial_inputs,
initial_outputs,
input_rngs,
output_fns,
dag,
weighted,
significance,
pb));
#endregion INFERENCE
}