breaking-vision-clip-classifier / scripts /hyperparameter_tuning.py
bawolf's picture
wip
c850c95
raw
history blame
9.89 kB
import optuna
import os
from datetime import datetime
import pandas as pd
from pathlib import Path
import json
import math
import sys
sys.path.append(os.path.dirname(os.path.dirname(__file__)))
from scripts.train import train_and_evaluate
from src.utils.utils import create_run_directory
def create_hyperparam_directory():
"""Create a parent directory for all hyperparameter searches"""
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
base_dir = "runs_hyperparam"
hyperparam_dir = os.path.join(base_dir, f"hyperparam_{timestamp}")
os.makedirs(hyperparam_dir, exist_ok=True)
return hyperparam_dir
def objective(trial, hyperparam_run_dir, data_path):
"""Objective function for a single dataset"""
# Then suggest parameters using the model-specific ranges
config = {
"clip_model": trial.suggest_categorical("clip_model", ["openai/clip-vit-base-patch32", "openai/clip-vit-large-patch14"]),
"batch_size": trial.suggest_categorical("batch_size", [8,16,32]),
"unfreeze_layers": trial.suggest_int("unfreeze_layers", 1, 4),
"learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
"weight_decay": trial.suggest_float("weight_decay", 1e-8, 1e-1, log=True),
"gradient_clip_max_norm": trial.suggest_float("gradient_clip_max_norm", 0.1, 1.0),
"augmentation_strength": trial.suggest_float("augmentation_strength", 0.0, 1.0),
"crop_scale_min": trial.suggest_float("crop_scale_min", 0.6, 0.9),
"max_frames": trial.suggest_int("max_frames", 5, 15),
"sigma": trial.suggest_float("sigma", 0.1, 0.5),
}
class_labels = ["windmill", "halo", "swipe", "baby_mill"][:3]
# Fixed configurations
config.update({
"class_labels": class_labels,
"num_classes": len(class_labels),
"data_path": data_path,
"num_epochs": 50,
"patience": 10,
"image_size": 224,
"crop_scale_max": 1.0,
"normalization_mean": [0.485, 0.456, 0.406],
"normalization_std": [0.229, 0.224, 0.225],
"overfitting_threshold": 10,
})
# Derive augmentation parameters
config.update({
"flip_probability": 0.5 * config["augmentation_strength"],
"rotation_degrees": int(15 * config["augmentation_strength"]),
"brightness_jitter": 0.2 * config["augmentation_strength"],
"contrast_jitter": 0.2 * config["augmentation_strength"],
"saturation_jitter": 0.2 * config["augmentation_strength"],
"hue_jitter": 0.1 * config["augmentation_strength"],
})
# Create dataset-specific run directory
dataset_label = '_'.join(Path(data_path).parts[-2:]) # Get last two parts of path
trial_dir = create_run_directory(
prefix=f"trial_{dataset_label}",
parent_dir=hyperparam_run_dir
)
config["run_dir"] = trial_dir
# Run training and evaluation with device cleanup
try:
val_accuracy, vis_dir = train_and_evaluate(config)
if val_accuracy is None or math.isnan(val_accuracy) or math.isinf(val_accuracy):
raise ValueError(f"Invalid accuracy value: {val_accuracy}")
# Save trial info
trial_info = {
'dataset': data_path,
'dataset_label': dataset_label,
'trial_number': trial.number,
'parameters': trial.params,
'accuracy': val_accuracy,
'visualization_dir': vis_dir,
'trial_dir': trial_dir
}
with open(os.path.join(trial_dir, 'trial_info.json'), 'w') as f:
json.dump(trial_info, f, indent=4)
return val_accuracy
except Exception as e:
print(f"Error in trial for {data_path}: {str(e)}")
# Log detailed error information
error_log_path = os.path.join(hyperparam_run_dir, 'error_log.txt')
with open(error_log_path, 'a') as f:
f.write(f"\nError in trial at {datetime.now()}:\n")
f.write(f"Dataset: {data_path}\n")
f.write(f"Error: {str(e)}\n")
f.write(f"Trial params: {trial.params}\n")
f.write("Stack trace:\n")
import traceback
f.write(traceback.format_exc())
f.write("\n" + "="*50 + "\n")
return float('-inf')
def run_hyperparameter_search(data_paths, n_trials=100):
"""Run hyperparameter search for multiple datasets"""
# Create parent directory for all searches
parent_hyperparam_dir = create_hyperparam_directory()
# Store results for all datasets
all_results = {}
for data_path in data_paths:
print(f"\nStarting hyperparameter search for dataset: {data_path}")
# Create dataset-specific directory
dataset_label = '_'.join(Path(data_path).parts[-2:])
dataset_dir = os.path.join(parent_hyperparam_dir, f"search_{dataset_label}")
os.makedirs(dataset_dir, exist_ok=True)
# Create and run study with explicit trial count tracking
study = optuna.create_study(direction="maximize")
completed_trials = 0
failed_trials = []
total_attempts = 0
max_attempts = n_trials * 2
while completed_trials < n_trials and total_attempts < max_attempts:
try:
total_attempts += 1
study.optimize(
lambda trial: objective(trial, dataset_dir, data_path),
n_trials=1
)
# Only increment if the trial actually succeeded
if study.trials[-1].value != float('-inf'):
completed_trials += 1
print(f"Completed trial {completed_trials}/{n_trials} for {dataset_label}")
else:
error_info = {
'trial_number': completed_trials + len(failed_trials) + 1,
'error': "Trial returned -inf",
'timestamp': datetime.now().isoformat()
}
failed_trials.append(error_info)
print(f"Failed trial for {dataset_label}: returned -inf")
except Exception as e:
error_info = {
'trial_number': completed_trials + len(failed_trials) + 1,
'error': str(e),
'timestamp': datetime.now().isoformat()
}
failed_trials.append(error_info)
print(f"Error in trial for {dataset_label}: {str(e)}")
# Log the error
with open(os.path.join(dataset_dir, 'failed_trials.json'), 'w') as f:
json.dump(failed_trials, f, indent=4)
if total_attempts >= max_attempts:
print(f"Warning: Reached maximum attempts ({max_attempts}) for {dataset_label}")
# Save study results
results_df = study.trials_dataframe()
results_df.to_csv(os.path.join(dataset_dir, 'study_results.csv'))
# Save trial statistics
trial_stats = {
'completed_trials': completed_trials,
'failed_trials': len(failed_trials),
'total_attempts': completed_trials + len(failed_trials)
}
with open(os.path.join(dataset_dir, 'trial_statistics.json'), 'w') as f:
json.dump(trial_stats, f, indent=4)
# Save best trial info
best_trial = study.best_trial
best_params_path = os.path.join(dataset_dir, 'best_params.txt')
with open(best_params_path, 'w') as f:
f.write(f"Best trial value: {best_trial.value}\n\n")
f.write("Best parameters:\n")
for key, value in best_trial.params.items():
f.write(f"{key}: {value}\n")
# Store results
all_results[data_path] = {
'best_value': best_trial.value,
'best_params': best_trial.params,
'study': study,
'results_df': results_df,
'failed_trials': failed_trials,
'trial_stats': trial_stats
}
print(f"\nResults for {data_path}:")
print(f"Completed trials: {completed_trials}")
print(f"Failed trials: {len(failed_trials)}")
print(f"Best trial value: {best_trial.value}")
print("Best parameters:")
for key, value in best_trial.params.items():
print(f" {key}: {value}")
# Create overall summary with additional statistics
summary_data = []
for data_path, result in all_results.items():
summary_data.append({
'dataset': data_path,
'best_accuracy': result['best_value'],
'completed_trials': result['trial_stats']['completed_trials'],
'failed_trials': result['trial_stats']['failed_trials'],
**result['best_params']
})
summary_df = pd.DataFrame(summary_data)
summary_df.to_csv(os.path.join(parent_hyperparam_dir, 'overall_summary.csv'), index=False)
return parent_hyperparam_dir, all_results
if __name__ == "__main__":
# List of dataset paths to optimize
data_paths = [
'./data/blog/datasets/bryant/random',
'./data/blog/datasets/bryant/adjusted',
'./data/blog/datasets/youtube/random',
'./data/blog/datasets/youtube/adjusted',
'./data/blog/datasets/combined/random',
'./data/blog/datasets/combined/adjusted',
'./data/blog/datasets/bryant_train_youtube_val/default'
]
# Run hyperparameter search
hyperparam_dir, results = run_hyperparameter_search(
data_paths,
n_trials=8 # Adjust as needed
)
print(f"\nHyperparameter search complete!")
print(f"Results are saved in: {hyperparam_dir}")