[TRACKER] Algorithm issues and tech debt

[dantegd]

Algorithm issues and tech debt

T-SNE

Numerical issues

• [BUG] t-SNE is not deterministic even with random_state #2980
• TSNE: Isolate condition that freezes on certain data & providing warning/assertion #3360
• [BUG] Why tSNE is stuck depending on the distribution of data? #3865

Tech debt

• [FEA] Promote FITSNE from experimental #3805
• Shared with UMAP [FEA] UMAP / TSNE should warn when knn_graph argument is invalid #3665

UMAP

Crash related issues

• [BUG] Semi-Supervised UMAP, with euclidean target_metric, reduction errors when input passes a certain size. #2333

Numerical issues

• [BUG] UMAP transform accuracy. #3864
• [BUG] UMAP failures in CUDA 11.2 #3943
• [BUG] Flaky UMAP MNMG Pytest #3905
• [BUG] UMAP transform accuracy. #3864

Tech debt

• [BUG] Re-enable reproducibility assertions in UMAP gtest for CTK 11.4 #4129
• Documentation about [BUG] Increasing memory usage leading to OOM when running UMAP in loop #4068
• Can't get UMAP to be distributed with DASK #3991
• Shared with T-SNE [FEA] UMAP / TSNE should warn when knn_graph argument is invalid #3665

RF

Crash related issues

• [QST] Is there a size limit to the input data for the RandomForestClassifier's fit function? #4132
• [BUG] SIGABRT in CUML RF, out of bounds memory usage #4046

Numerical issues

• [BUG] Overly-confident RF probabilities with unbalanced data #3023
• [BUG] Negative R-Squared with Nightly (0.17) #3305

Tech debt

• [TRACKER] RCA for RF accuracy issue #3526
• [BUG] Flaky RF Classification test (test_random_forest.py::test_rf_classification_seed) #3740
• [BUG] cuml.dask random forest pytest OOM's on 10GB GPU #3618
• [TRACKER] Clearing technical debt of RF/DT code #3999
• [BUG] Random Forest stress test failing: test_rf_host_memory_leak #3787

KMeans

Crash related issues

• Older so need to see if still reproduces with latest cuML/NCCL [BUG] Kmeans MNMG Notebook: NCCL Error #2458

Numerical issues

• [BUG] Significant difference between cuML KMeans and sklearn KMeans #3412
• [BUG] KMeans random state differs from sklearn implementation #2771

Tech debt

• [QST]Using cuML kmeans from C++ application is falied #3283
• [BUG] Intermittent test failure for dask kmeans score method #2472

Nearest Neighbors

Crash related issues

• [BUG] NearestNeighbors crashes with non-brute algorithms #4020
• [BUG] cuml.dask.neighbors.NearestNeighbors: Crash with core dump on compute.... #3571
• [BUG] KNeighborsClassifier prediction fails  #3346

Linear Models

Crash related issues

• [BUG] cuml multi GPU Ridge and Linear Regression fails with cudaErrorMisalignedAddress:misaligned address #3965

Numerical issues

• [BUG] ElasticNet fit time is slow with high variance on close to rank deficient matrices #4122
• [BUG] Mismatch between cuml and sklearn LinearRegression for specific inputs #3465

DBSCAN

• [BUG] MemoryError: std::bad_alloc: CUDA error at: /home/jon/minicondadai_py38/include/rmm/mr/device/cuda_memory_resource.hpp:69: cudaErrorMemoryAllocation out of memory #4067
• [DISCUSS] Use of cudaMemGetInfo for computing potential allocatable memory can be dangerous w/ RMM  #3979

Others

• [BUG] MNMG PCA test failing due to timeout #3955
• [BUG] HDBSCAN is marking points in dense regions as noise #4054
• Why are the results rapids much worse than sklearn? #3945
• [BUG] SparseRandomProjection cudaErrorInvalidValue:invalid argument #3857
• [BUG]MBSGD accuracy much lower than scikit-learn #2612
• [BUG] Sporadic single linkage agglomerative clustering pytest fail  #4044
• cuml PCA shows cuda out of memory while GPU shows lots of vacancy.  #4141 being tackled by @lowener

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.