scikit-learn-bench 0.1.2

ML Benchmark used to compare platforms

scikit-learn-bench

Benchmark 100+ scikit-learn Machine Learning algorithms at once — in just a few seconds.

This tool offers an easy way to evaluate models across multiple ML categories and profiling strategies.

---

📦 Installation

pip3 install scikit_learn_bench

🚀 Features

📊 Data Generation

Easily control the characteristics of synthetic datasets:

• num_samples: Number of samples (rows)
• num_features: Number of input features (columns)
• num_output: Target shape — used for regression, classification, clusters, or transformed outputs

---

🧠 ML Algorithm Types

Type	Label	Description
Regressors	"reg"	55 algorithms with 2 methods .fit and .predict
Classifiers	"cla"	42 algorithms with 2 methods .fit and .predict
Clustering	"clu"	12 clustering algorithms (.predict supported for 6)
Transformations	"tra"	68 transform functions (e.g. MinMaxScaler, PCA, TSNE, etc.)

In total, the tool allows benchmarking 280 scikit-learn functions (55*2+42*2+12+6+68).

The exact counts may vary depending on your installed scikit-learn version (here 1.6.1), the data characteristics (#samples, #features, #output), and other software dependencies.

---

⏱ Profiling Strategies

Choose one of three profiler types:

• "time":
  Measures training and inference throughput (samples/sec).
  Output: (train_throughput, infer_throughput)

• "timememory":
  Adds peak memory (kB) with tracemalloc.
  Output: (train_throughput, infer_throughput, train_peak_memory, infer_peak_memory)

• "timeline":
  Fine-grained cProfile analysis saved as .prof files for each algorithm.
  Output: .prof file per model

---

⚙️ Other Parameters

• min_prof_time: Minimum time in seconds to run the profiling of each function (reduces noise)
• max_prof_time: Maximum time in seconds to run the profiling. If the delay is reached, a timeout interrupts the computation.
• table_print: Display formatted results in console
• table_print_sort_crit: Sort results (e.g., by training speed)
• line_profiler_path: Path to store .prof files for "timeline" profiler

---

🧪 Example Usage

CLI

After installing scikit_learn_bench, you can invoke it directly from the command line:

pierrick@laptop:~$ pip3 install scikit_learn_bench
pierrick@laptop:~$ scikit_learn_bench

For detailed usage and options, run:

scikit_learn_bench --help

Programming interface (advanced)

from scikit_learn_bench.core import bench

scores = bench(
  num_samples=10,
  num_features=2,
  num_output=2,
  min_prof_time=0.1,
  ml_type="cla",
  profiler_type="timememory",
  table_print=True
)

This function returns a dictionary with performance metrics for each algorithm, such as:

{
    'AdaBoostClassifier': (4454.128, 48093.051, 94.06, 19.29),
    'BaggingClassifier': (282.16, 6696.015, 96.019, 162.843),
    ...
}

The output includes:

• Train/s: Training speed (samples per second)
• Train Mem: Memory usage during training (MB)
• Infer/s: Inference speed (samples per second)
• Infer Mem: Memory usage during inference (MB)

Advanced performance analysis

dAditionally, the usage_example/ directory contains scripts for advanced analyses, including:

• 2D cloud points comparing throughput and memory consumption across all algorithms
• Scalability studies examining how performance varies with data size (samples, features, output size)
• Analysis of algorithm performance as the number of CPU cores increases, helping identify which algorithms benefit most from parallel processing

📚 Citing scikit-learn-bench

Work in progress: BibTeX / reference for citing this repo.

🙏 Acknowledgments

ULHPC Platform for computing support and motivating this project.