Skip to content

Getting Started with Benchmarking

This guide will walk you through the basics of benchmarking DNA language models using DNALLM. You'll learn how to set up your first benchmark, configure models and datasets, and interpret results.

Overview

Benchmarking in DNALLM allows you to: - Compare multiple DNA language models on the same tasks - Evaluate performance across different datasets - Measure accuracy, speed, and resource usage - Generate comprehensive performance reports

Prerequisites

Ensure you have the following installed and configured:

# Install DNALLM
pip install dnallm

# Or with uv (recommended)
uv pip install dnallm

Basic Setup

1. Import Required Modules

from dnallm import load_config, Benchmark
from dnallm.inference import load_model_and_tokenizer
from dnallm.datahandling import DNADataset

2. Create a Simple Configuration

Create a benchmark_config.yaml file:

# benchmark_config.yaml
benchmark:
  name: "My First Benchmark"
  description: "Comparing DNA models on promoter prediction"

  models:
    - name: "Plant DNABERT"
      path: "zhangtaolab/plant-dnabert-BPE"
      source: "huggingface"
      task_type: "classification"

    - name: "Plant DNAGPT"
      path: "zhangtaolab/plant-dnagpt-BPE"
      source: "huggingface"
      task_type: "generation"

  datasets:
    - name: "promoter_data"
      path: "path/to/your/data.csv"
      task: "binary_classification"
      text_column: "sequence"
      label_column: "label"

  metrics:
    - "accuracy"
    - "f1_score"
    - "precision"
    - "recall"

  evaluation:
    batch_size: 16
    max_length: 512
    device: "cuda"  # or "cpu"

  output:
    format: "html"
    path: "benchmark_results"

3. Load Your Data

# Load your dataset
dataset = DNADataset.load_local_data(
    "path/to/your/data.csv",
    seq_col="sequence",
    label_col="label",
    max_length=512
)

# Split if needed
if not dataset.is_split:
    dataset.split_data(test_size=0.2, val_size=0.1)

4. Run the Benchmark

# Load configuration
config = load_config("benchmark_config.yaml")

# Initialize benchmark
benchmark = Benchmark(config=config)

# Run benchmark
results = benchmark.run()

# Display results
print("Benchmark Results:")
print("=" * 50)
for model_name, model_results in results.items():
    print(f"\n{model_name}:")
    for dataset_name, metrics in model_results.items():
        print(f"  {dataset_name}:")
        for metric, value in metrics.items():
            print(f"    {metric}: {value:.4f}")

Command Line Interface

DNALLM also provides a convenient command-line interface:

# Basic benchmark run
dnallm-benchmark --config benchmark_config.yaml

# Generate detailed report
dnallm-benchmark --config config.yaml --output report.html

# Run with custom parameters
dnallm-benchmark --config config.yaml --batch-size 32 --device cuda

Understanding Results

Basic Metrics

Metric Description Range Best Value
Accuracy Correct predictions / Total predictions 0.0 - 1.0 1.0
F1 Score Harmonic mean of precision and recall 0.0 - 1.0 1.0
Precision True positives / (True positives + False positives) 0.0 - 1.0 1.0
Recall True positives / (True positives + False negatives) 0.0 - 1.0 1.0

Performance Metrics

Metric Description Unit
Inference Time Time to process one batch seconds
Memory Usage GPU/RAM memory consumption MB/GB
Throughput Samples processed per second samples/sec

Example: Complete Benchmark

Here's a complete working example:

import os
from dnallm import load_config, Benchmark
from dnallm.datahandling import DNADataset

# 1. Prepare your data
data_path = "path/to/your/dna_sequences.csv"
if not os.path.exists(data_path):
    print("Please provide a valid data path")
    exit()

# 2. Load and prepare dataset
dataset = DNADataset.load_local_data(
    data_path,
    seq_col="sequence",
    label_col="label",
    max_length=512
)

# 3. Create configuration
config = {
    "benchmark": {
        "name": "DNA Model Comparison",
        "models": [
            {
                "name": "Plant DNABERT",
                "path": "zhangtaolab/plant-dnabert-BPE",
                "source": "huggingface",
                "task_type": "classification"
            },
            {
                "name": "Plant DNAGPT",
                "path": "zhangtaolab/plant-dnagpt-BPE", 
                "source": "huggingface",
                "task_type": "generation"
            }
        ],
        "datasets": [dataset],
        "metrics": ["accuracy", "f1_score", "precision", "recall"],
        "evaluation": {
            "batch_size": 16,
            "max_length": 512,
            "device": "cuda"
        },
        "output": {
            "format": "html",
            "path": "my_benchmark_results"
        }
    }
}

# 4. Run benchmark
benchmark = Benchmark(config=config)
results = benchmark.run()

# 5. Generate report
benchmark.generate_report(
    output_path="my_benchmark_results",
    format="html",
    include_predictions=True
)

print("Benchmark completed! Check 'my_benchmark_results' folder for results.")

Data Format Requirements

Your dataset should be in one of these formats:

CSV/TSV Format

sequence,label
ATCGATCGATCG,1
GCTAGCTAGCTA,0
TATATATATATA,1

JSON Format

[
  {"sequence": "ATCGATCGATCG", "label": 1},
  {"sequence": "GCTAGCTAGCTA", "label": 0}
]

FASTA Format

>sequence1|label:1
ATCGATCGATCG
>sequence2|label:0
GCTAGCTAGCTA

Common Tasks

Binary Classification

task: "binary_classification"
num_labels: 2
label_names: ["Negative", "Positive"]
threshold: 0.5

Multi-class Classification

task: "multiclass"
num_labels: 4
label_names: ["Class_A", "Class_B", "Class_C", "Class_D"]

Regression

task: "regression"
num_labels: 1

Next Steps

After completing this basic tutorial:

  1. Explore Advanced Features: Learn about cross-validation and custom metrics
  2. Optimize Performance: Discover performance profiling techniques
  3. Customize Output: Learn about advanced configuration options
  4. Real-world Examples: See practical use cases

Troubleshooting

Common Issues

"Model not found" error 

# Check if model exists on Hugging Face
# Visit: https://huggingface.co/models?search=dna

Memory errors 

# Reduce batch size in config
evaluation:
  batch_size: 8  # Reduced from 16

Slow performance 

# Enable mixed precision
evaluation:
  use_fp16: true

Additional Resources

Ready for more? Continue to Advanced Techniques to learn about cross-validation, custom metrics, and performance profiling.