Peter M. Groen / openbr

Commit b52a58b69dcf13583f9482a32e41feb30eaa9a7a

Authored by Josh Klontz
1 parent 830083ad

optimized pipe training

Inline Side-by-side
Showing 2 changed files with 36 additions and 8 deletions
openbr/openbr_plugin.cpp
  View file @b52a58b
... ... @@ -1162,6 +1162,7 @@ public:
1162 1162 transform->setParent(this);
1163 1163 transforms.append(transform);
1164 1164 file = transform->file;
  1165 + trainable = transform->trainable;
1165 1166 setObjectName(transforms.first()->objectName());
1166 1167 }
1167 1168  
... ... @@ -1178,9 +1179,8 @@ private:
1178 1179  
1179 1180 void train(const TemplateList &data)
1180 1181 {
1181   - // Don't bother constructing datasets if the transform is untrainable
1182   - if (dynamic_cast<UntrainableTransform*>(transforms.first()))
1183   - return;
  1182 + // Don't bother if the transform is untrainable
  1183 + if (!trainable) return;
1184 1184  
1185 1185 QList<TemplateList> templatesList;
1186 1186 foreach (const Template &t, data) {
... ...
openbr/plugins/meta.cpp
  View file @b52a58b
... ... @@ -75,14 +75,42 @@ class PipeTransform : public CompositeTransform
75 75 {
76 76 Q_OBJECT
77 77  
  78 + void _projectPartial(Template *srcdst, int startIndex, int stopIndex)
  79 + {
  80 + for (int i=startIndex; i<stopIndex; i++)
  81 + *srcdst >> *transforms[i];
  82 + }
  83 +
78 84 void train(const TemplateList &data)
79 85 {
80 86 TemplateList copy(data);
81   - for (int i=0; i<transforms.size(); i++) {
82   - fprintf(stderr, "%s training... ", qPrintable(transforms[i]->objectName()));
83   - transforms[i]->train(copy);
84   - fprintf(stderr, "projecting...\n");
85   - copy >> *transforms[i];
  87 + int i = 0;
  88 + while (i < transforms.size()) {
  89 + fprintf(stderr, "%s", qPrintable(transforms[i]->objectName()));
  90 +
  91 + // Conditional statement covers likely case that first transform is untrainable
  92 + if (transforms[i]->trainable) {
  93 + fprintf(stderr, " training...");
  94 + transforms[i]->train(copy);
  95 + }
  96 +
  97 + // We project through any subsequent untrainable transforms at once
  98 + // as a memory optimization in case any of these intermediate
  99 + // transforms allocate a lot of memory (like OpenTransform)
  100 + // then we don't want all the training templates to be processed
  101 + // by that transform at once if we can avoid it.
  102 + int nextTrainableTransform = i+1;
  103 + while ((nextTrainableTransform < transforms.size()) &&
  104 + !transforms[nextTrainableTransform]->trainable)
  105 + nextTrainableTransform++;
  106 +
  107 + fprintf(stderr, " projecting...\n");
  108 + QFutureSynchronizer<void> futures;
  109 + for (int j=0; j<copy.size(); j++)
  110 + if (Globals->parallelism) futures.addFuture(QtConcurrent::run(this, &PipeTransform::_projectPartial, &copy[j], i, nextTrainableTransform));
  111 + else _projectPartial( &copy[j], i, nextTrainableTransform);
  112 + futures.waitForFinished();
  113 + i = nextTrainableTransform;
86 114 }
87 115 }
88 116  
... ...