cuML

cuML (GPU) path

In the following, we identify the hurdles that appear when enablimg a cuML (GPU) path and we outline minimal, targeted changes that unlock NVIDIA T4‑GPU performance in our RunModels Google Colab while preserving the notebook’s familiar layout.

The RunModels colab already delivers a well‑structured CPU workflow—data ingestion, preprocessing, SMOTE for class imbalance, model training and evaluation. Parameter widgets let analysts experiment easily.

SMOTE - Improves generalization, but accuracy will decline a little

Before SMOTE:

• The dataset may initially be heavily imbalanced, so models easily memorized the dominant class.

• As the validation split preserved the imbalance, models posted inflated high scores without truly learning minority patterns.

After SMOTE:

• SMOTE balanced the classes, forcing models to actually learn.

• Consequently, accuracy and F1 scores dropped, but generalization improved.

• Models now have to predict more than just the obvious — exposing weaknesses (especially in SVM).

Updates made for cuML

We re‑ordered to one install, added a single setup cell, converted data once, swapped four import lines for GPU acceleration without rewriting the notebook. The updates preserve accuracy and position the project for future GPU‑centric enhancements.
- Srivatsav and Gayathri

IDENTIFIED TECHNICAL ISSUES

1. imbalanced‑learn installation never executes in a clean, sequential run
   • In Colab, Run all executes cells strictly from top to bottom and stops on the first error.
   • Cell 11 tries from imblearn.over_sampling import SMOTE before the package exists.
   • Python raises ModuleNotFoundError, execution halts, and the later pip install imbalanced‑learn cell (cell 42) never runs—so the package remains uninstalled.
   • The problem is purely the cell order; moving the install command ahead of the import—or putting all installs in a single setup cell—solves it for every fresh run.

2. GPU‑aware libraries not yet invoked
   • Training code still imports scikit‑learn classes (RandomForestClassifier, XGBClassifier, SVC, LogisticRegression).
   • Without explicit calls to GPU‑aware libraries (cuML, cuDF), model fitting stays on the CPU.

3. Version‑collision risk: cuML vs. imbalanced‑learn
   • RAPIDS/cuML 24.04 pins scikit‑learn at 1.2.
   • Newer imbalanced‑learn releases (0.13 +) require scikit‑learn 1.3 +.
   • Installing cuML first downgrades scikit‑learn and breaks imbalanced‑learn; installing imbalanced‑learn first blocks cuML.

PROPOSED ADJUSTMENTS

A. Unified installation cell

• Install RAPIDS 24.04 CUDA‑12 wheels: cudf‑cu12, cuml‑cu12, cupy‑cuda12, xgboost‑cu12.
• Install imbalanced‑learn 0.11.0 (compatible with scikit‑learn 1.2).
• Import pandas, NumPy, seaborn in the same cell.

B. One‑time DataFrame conversion

• Immediately after train_test_split, convert feature DataFrames to cudf.DataFrame and label arrays to cupy.ndarray.

C. Drop‑in estimator swaps

Current CPU class	GPU replacement (cuML)	Why safe	Benefit on T4
RandomForestClassifier	cuml.ensemble.RandomForestClassifier	API identical	90 s → 11 s
XGBClassifier (CPU)	cuml.experimental.XGBClassifier	Wraps CUDA build	5 min → 30 s
SVC	cuml.svm.SVC	Same methods	Minutes → seconds
LogisticRegression	cuml.linear_model.LogisticRegression	Only penalty='l2' needed	Sub‑second fit

D. Class‑imbalance handling

• Default: keep SMOTE on CPU.
• Optional: flag use_gpu_smote enables cuML’s experimental SMOTENC for all‑GPU oversampling.

IMPLEMENTATION STEPS

1. Create branch feature/gpu‑refactor.
2. Insert unified install cell; remove scattered pip install lines.
3. Add helpers for one‑time pandas→cuDF and NumPy→cupy conversion.
4. Replace CPU estimator imports with cuML versions.
5. Add use_gpu_smote flag.
6. Validate notebook end‑to‑end in Colab (T4 GPU); confirm ROC‑AUC matches CPU baseline.
7. Peer‑review and merge.

EXPECTED OUTCOMES

• Random Forest training: ~90 s CPU → ~11–12 s GPU.
• XGBoost: ~5 min CPU → ~30 s GPU.
• Same user interface; only import lines change.
• Version pinning prevents silent breakage.
• Data already in cuDF sets the stage for GPU hyper‑parameter tuning or multi‑GPU scaling.

RISK ASSESSMENT

Risk	Mitigation
RAPIDS API changes	Pin RAPIDS 24.04; monthly smoke‑test against nightly build.
SMOTENC instability	Off by default; flag toggles it.
Colab image drift	Install cell pins versions every run.