[REVIEW] Obey initialize_embeddings parameter in B-H tSNE

CHANGELOG.md

@@ -13,6 +13,7 @@
 
 ## Bug Fixes
 - PR #2983: Fix seeding of KISS99 RNG
+- PR #3011: Fix unused initialize_embeddings parameter in Barnes-Hut t-SNE
 - PR #3008: Check number of columns in check_array validator
 - PR #3012: Increasing learning rate for SGD log loss and invscaling pytests
 

cpp/include/cuml/manifold/tsne.h

@@ -59,7 +59,7 @@ namespace ML {
  *                                 or >= 0 for reproducible outputs.
  * @param[in] verbosity            verbosity level for logging messages during
  *                                 execution
- * @param[in] intialize_embeddings Whether to overwrite the current Y vector
+ * @param[in] initialize_embeddings Whether to overwrite the current Y vector
  *                                 with random noise.
  * @param[in] barnes_hut           Whether to use the fast Barnes Hut or use the
  *                                 slower exact version.
@@ -85,6 +85,6 @@ void TSNE_fit(const raft::handle_t &handle, const float *X, float *Y,
               const float pre_momentum = 0.5, const float post_momentum = 0.8,
               const long long random_state = -1,
               int verbosity = CUML_LEVEL_INFO,
-              const bool intialize_embeddings = true, bool barnes_hut = true);
+              const bool initialize_embeddings = true, bool barnes_hut = true);
 
 }  // namespace ML

cpp/src/tsne/barnes_hut.cuh

@@ -46,6 +46,7 @@ namespace TSNE {
  * @param[in] pre_momentum: The momentum used during the exaggeration phase.
  * @param[in] post_momentum: The momentum used after the exaggeration phase.
  * @param[in] random_state: Set this to -1 for pure random intializations or >= 0 for reproducible outputs.
+ * @param[in] initialize_embeddings: Whether to overwrite the current Y vector with random noise.
  */
 void Barnes_Hut(float *VAL, const int *COL, const int *ROW, const int NNZ,
                 const raft::handle_t &handle, float *Y, const int n,
@@ -56,7 +57,8 @@ void Barnes_Hut(float *VAL, const int *COL, const int *ROW, const int NNZ,
                 const float post_learning_rate = 500.0f,
                 const int max_iter = 1000, const float min_grad_norm = 1e-7,
                 const float pre_momentum = 0.5, const float post_momentum = 0.8,
-                const long long random_state = -1) {
+                const long long random_state = -1,
+                const bool initialize_embeddings = true) {
   auto d_alloc = handle.get_device_allocator();
   cudaStream_t stream = handle.get_stream();
 
@@ -131,9 +133,13 @@ void Barnes_Hut(float *VAL, const int *COL, const int *ROW, const int NNZ,
     cudaMemsetAsync(old_forces.data(), 0, sizeof(float) * n * 2, stream));
 
   MLCommon::device_buffer<float> YY(d_alloc, stream, (nnodes + 1) * 2);
-  // TODO bug #2549: this should be conditional on bool initialize_embeddings.
-  random_vector(YY.data(), -0.0001f, 0.0001f, (nnodes + 1) * 2, stream,
-                random_state);
+  if (initialize_embeddings) {
+    random_vector(YY.data(), -0.0001f, 0.0001f, (nnodes + 1) * 2, stream,
+                  random_state);
+  } else {
+    raft::copy(YY.data(), Y, n, stream);
+    raft::copy(YY.data() + nnodes + 1, Y + n, n, stream);
+  }
 
   // Set cache levels for faster algorithm execution
   //---------------------------------------------------

cpp/src/tsne/exact_tsne.cuh

@@ -44,7 +44,7 @@ namespace TSNE {
  * @param[in] pre_momentum: The momentum used during the exaggeration phase.
  * @param[in] post_momentum: The momentum used after the exaggeration phase.
  * @param[in] random_state: Set this to -1 for pure random intializations or >= 0 for reproducible outputs.
- * @param[in] intialize_embeddings: Whether to overwrite the current Y vector with random noise.
+ * @param[in] initialize_embeddings: Whether to overwrite the current Y vector with random noise.
  */
 void Exact_TSNE(float *VAL, const int *COL, const int *ROW, const int NNZ,
                 const raft::handle_t &handle, float *Y, const int n,
@@ -55,11 +55,11 @@ void Exact_TSNE(float *VAL, const int *COL, const int *ROW, const int NNZ,
                 const int max_iter = 1000, const float min_grad_norm = 1e-7,
                 const float pre_momentum = 0.5, const float post_momentum = 0.8,
                 const long long random_state = -1,
-                const bool intialize_embeddings = true) {
+                const bool initialize_embeddings = true) {
   auto d_alloc = handle.get_device_allocator();
   cudaStream_t stream = handle.get_stream();
 
-  if (intialize_embeddings)
+  if (initialize_embeddings)
     random_vector(Y, -0.0001f, 0.0001f, n * dim, stream, random_state);
 
   // Allocate space

cpp/src/tsne/tsne.cu

@@ -36,7 +36,8 @@ void TSNE_fit(const raft::handle_t &handle, const float *X, float *Y,
               const float post_learning_rate, const int max_iter,
               const float min_grad_norm, const float pre_momentum,
               const float post_momentum, const long long random_state,
-              int verbosity, const bool intialize_embeddings, bool barnes_hut) {
+              int verbosity, const bool initialize_embeddings,
+              bool barnes_hut) {
   ASSERT(n > 0 && p > 0 && dim > 0 && n_neighbors > 0 && X != NULL && Y != NULL,
          "Wrong input args");
   ML::Logger::get().setLevel(verbosity);
@@ -118,12 +119,13 @@ void TSNE_fit(const raft::handle_t &handle, const float *X, float *Y,
     TSNE::Barnes_Hut(VAL, COL, ROW, NNZ, handle, Y, n, theta, epssq,
                      early_exaggeration, exaggeration_iter, min_gain,
                      pre_learning_rate, post_learning_rate, max_iter,
-                     min_grad_norm, pre_momentum, post_momentum, random_state);
+                     min_grad_norm, pre_momentum, post_momentum, random_state,
+                     initialize_embeddings);
   } else {
     TSNE::Exact_TSNE(VAL, COL, ROW, NNZ, handle, Y, n, dim, early_exaggeration,
                      exaggeration_iter, min_gain, pre_learning_rate,
                      post_learning_rate, max_iter, min_grad_norm, pre_momentum,
-                     post_momentum, random_state, intialize_embeddings);
+                     post_momentum, random_state, initialize_embeddings);
   }
 }
 

python/cuml/manifold/t_sne.pyx

@@ -67,7 +67,7 @@ cdef extern from "cuml/manifold/tsne.h" namespace "ML" nogil:
         const float post_momentum,
         const long long random_state,
         int verbosity,
-        const bool intialize_embeddings,
+        const bool initialize_embeddings,
         bool barnes_hut) except +