run2-parallel-grid-search

name	run2_parallel_grid_search
description	Efficient parallel hyperparameter grid search using joblib with 847 combinations

Parallel Grid Search Optimization

Grid Search Space

For Mars cloud clustering, evaluate all combinations of:

• min_samples: 3, 4, 5, 6, 7, 8, 9 (7 values)
• epsilon: 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24 (11 values)
• shape_weight: 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9 (11 values)

Total combinations: 7 × 11 × 11 = 847

Parallelization with joblib

Setup

from joblib import Parallel, delayed
import pandas as pd
import numpy as np

# Define hyperparameter ranges
min_samples_vals = list(range(3, 10))
epsilon_vals = list(range(4, 26, 2))
shape_weight_vals = list(np.round(np.arange(0.9, 2.0, 0.1), 1))

# Generate all combinations
hp_list = []
for ms in min_samples_vals:
    for eps in epsilon_vals:
        for sw in shape_weight_vals:
            hp_list.append((ms, eps, sw))

print(f"Total combinations: {len(hp_list)}")  # Should be 847

Evaluation Function

Each worker evaluates one hyperparameter combination:

def evaluate_hyperparams(min_samples, epsilon, shape_weight, citsci_df, expert_df, expert_images):
    """
    Evaluate one hyperparameter combination across all expert images.

    Returns: (F1_avg, delta_avg, min_samples, epsilon, shape_weight)
    """
    f1_scores = []
    delta_scores = []

    for image_id in expert_images:
        f1, delta = evaluate_image(image_id, citsci_df, expert_df,
                                    min_samples, epsilon, shape_weight)
        f1_scores.append(f1)
        if not np.isnan(delta):
            delta_scores.append(delta)

    avg_f1 = np.mean(f1_scores) if f1_scores else 0.0
    avg_delta = np.mean(delta_scores) if delta_scores else np.nan

    return avg_f1, avg_delta, min_samples, epsilon, shape_weight

Parallel Execution

# Load data once, before parallelization
citsci_df = pd.read_csv('/root/data/citsci_train.csv')
expert_df = pd.read_csv('/root/data/expert_train.csv')
expert_images = sorted(expert_df['file_rad'].unique())

# Evaluate all combinations in parallel
results = Parallel(n_jobs=-1, verbose=1)(
    delayed(evaluate_hyperparams)(ms, eps, sw, citsci_df, expert_df, expert_images)
    for ms, eps, sw in hp_list
)

# Convert results to DataFrame
results_df = pd.DataFrame(results,
    columns=['F1', 'delta', 'min_samples', 'epsilon', 'shape_weight'])

Joblib Parameters

n_jobs=-1

Uses all available CPU cores. For typical 22-core machines:

• ~22 parallel workers
• 847 combinations ÷ 22 ≈ 39 tasks per worker
• Total time: ~1.7 minutes (vs. ~60+ minutes sequentially)

verbose=1

Prints progress:

[Parallel(n_jobs=-1)]: Done   6 tasks      | elapsed:    3.0s
[Parallel(n_jobs=-1)]: Done 156 tasks      | elapsed:   20.7s
[Parallel(n_jobs=-1)]: Done 847 out of 847 | elapsed:  1.7min finished

Memory Efficiency

Data Sharing

Using Unix fork (default for Loky backend):

• Data loaded once in parent process
• Child processes inherit via copy-on-write
• No duplication overhead

Per-Evaluation Memory

Each worker needs:

• DBSCAN per image: O(n²) for distance matrix (n = points per image)
• Typical image: 50-100 points → 3KB-10KB distance matrix
• All workers combined: reasonable memory usage

# Load data BEFORE parallel section
citsci_df = pd.read_csv('/root/data/citsci_train.csv')
expert_df = pd.read_csv('/root/data/expert_train.csv')
expert_images = sorted(expert_df['file_rad'].unique())

# Pass as parameters (more efficient than loading in each worker)
results = Parallel(n_jobs=-1)(
    delayed(evaluate_hyperparams)(ms, eps, sw,
                                   citsci_df, expert_df, expert_images)
    for ms, eps, sw in hp_list
)

Progress Tracking Alternatives

Option 1: joblib verbose

results = Parallel(n_jobs=-1, verbose=10)(
    delayed(fn)(args) for args in arguments
)
# verbose levels: 1=minimal, 5=moderate, 10=detailed

Option 2: tqdm progress bar

from tqdm import tqdm

results = Parallel(n_jobs=-1)(
    delayed(fn)(args)
    for args in tqdm(arguments, total=len(arguments))
)

Option 3: Custom progress tracking

from joblib import Parallel, delayed, parallel_backend

def progress_callback(n_tasks):
    """Called every n_tasks completed."""
    print(f"Progress: {n_tasks} tasks completed")

results = Parallel(n_jobs=-1)(
    delayed(fn)(args) for args in arguments
)

Results Aggregation

# All results collected after parallel section completes
results_df = pd.DataFrame(results,
    columns=['F1', 'delta', 'min_samples', 'epsilon', 'shape_weight'])

# Verify completeness
assert len(results_df) == 847, "Missing results"

# Analyze
print(f"F1 range: [{results_df['F1'].min():.5f}, {results_df['F1'].max():.5f}]")
print(f"Delta range: [{results_df['delta'].min():.5f}, {results_df['delta'].max():.5f}]")
print(f"Solutions with F1 > 0.5: {(results_df['F1'] > 0.5).sum()}")

Error Handling

Timeout Protection

# Set timeout for long-running tasks
results = Parallel(n_jobs=-1, timeout=300)(  # 5 minutes per task
    delayed(evaluate_hyperparams)(ms, eps, sw, citsci_df, expert_df, expert_images)
    for ms, eps, sw in hp_list
)

Worker Crash Recovery

# If a worker crashes, task fails and exception propagates
# Typical causes:
# - Out of memory
# - Segmentation fault
# - Infinite loop (timeout helps)

# Recovery: Increase memory, reduce n_jobs, or fix underlying issue

Performance Optimization

Batch Size

# Parallel can batch tasks for efficiency
results = Parallel(n_jobs=-1, batch_size='auto')(
    delayed(fn)(args) for args in arguments
)

Backend Selection

from joblib import parallel_backend

# Loky backend (default, most robust)
with parallel_backend('loky', n_jobs=-1):
    results = Parallel()(delayed(fn)(args) for args in arguments)

# Threading backend (for I/O-bound tasks, lower overhead)
with parallel_backend('threading', n_jobs=-1):
    results = Parallel()(delayed(fn)(args) for args in arguments)

Reproducibility

# Set random seed for reproducibility
import numpy as np
np.random.seed(42)

# Note: DBSCAN is deterministic (no randomness)
# Parallelization doesn't affect results (only speed)

Typical Performance

For 847 DBSCAN evaluations across 369 images:

Configuration	Time
Sequential (1 core)	~60-90 minutes
Parallel (22 cores)	~1.5-2 minutes
Speedup	~40-45×

Debugging Parallel Code

# Test single evaluation first
result = evaluate_hyperparams(3, 4, 0.9, citsci_df, expert_df, expert_images)
print(result)

# Then enable parallelization
results = Parallel(n_jobs=-1, verbose=10)(
    delayed(evaluate_hyperparams)(ms, eps, sw, citsci_df, expert_df, expert_images)
    for ms, eps, sw in hp_list[:10]  # Test on subset first
)

# Finally, full grid search
results = Parallel(n_jobs=-1, verbose=1)(
    delayed(evaluate_hyperparams)(ms, eps, sw, citsci_df, expert_df, expert_images)
    for ms, eps, sw in hp_list
)

Key Takeaways

1. Load data before parallelization: Avoids redundant I/O in each worker
2. Return simple types: Tuples faster than dictionaries or objects
3. Use n_jobs=-1: Automatic CPU count detection and full utilization
4. Monitor progress: verbose=1 or tqdm for user feedback
5. Test sequentially first: Easier debugging before parallel execution
6. Check results completeness: Verify all 847 combinations evaluated