ICL_code/ICL_Jay_final/data/prepare_image_data.sh

#!/bin/bash
#SBATCH --partition=scavenger-gpu
#SBATCH --exclude=dcc-youlab-gpu-28,dcc-gehmlab-gpu-56
#SBATCH --array=2-3
#SBATCH --ntasks=1
#SBATCH --nodes=1
#SBATCH --cpus-per-task=16
#SBATCH --mem-per-cpu=64G
#SBATCH --gres=gpu:1
#SBATCH --time=72:00:00
#SBATCH --output=/work/jf381/code/code/ICL_LOG/dataset_generation/%A_%a_%j.out
#SBATCH --error=/work/jf381/code/code/ICL_LOG/dataset_generation/%A_%a_%j.err
#SBATCH --requeue
SLURM_ARRAY_TASK_ID=3
# Enhanced Dataset Generation Script for SLURM Array Jobs
# This script uses SLURM_ARRAY_TASK_ID to control which dataset to generate
set -e  # Exit on any error

# Configuration
BASE_OUTPUT_DIR="/work/jf381/data/icl_jay"
RAW_DATA_DIR="/work/jf381/data/raw_datasets"
NUM_PROMPTS=1000
SAMPLES_PER_CLASS=50
IMAGE_SIZE="256 256"
NUM_WORKERS=8  # Use all available CPUs

# Create necessary directories
mkdir -p "$BASE_OUTPUT_DIR"
mkdir -p "$RAW_DATA_DIR"
mkdir -p "/work/jf381/code/code/ICL_LOG/dataset_generation"

# Function to log with timestamp
log_message() {
    echo "[$(date '+%Y-%m-%d %H:%M:%S')] $1"
}

# Function to prepare a specific dataset
prepare_dataset() {
    local dataset_type=$1
    local output_dir="$BASE_OUTPUT_DIR/$dataset_type"
    
    log_message "Starting preparation of $dataset_type dataset..."
    log_message "SLURM Job ID: $SLURM_JOB_ID"
    log_message "Array Task ID: $SLURM_ARRAY_TASK_ID"
    log_message "Output directory: $output_dir"
    log_message "Using $NUM_WORKERS worker threads"
    
    # Record start time
    start_time=$(date +%s)
    
    # Run dataset preparation
    python /work/jf381/code/code/ICL_Jay/data/prepare_datasets.py \
        --dataset "$dataset_type" \
        --data_dir "$RAW_DATA_DIR" \
        --output_dir "$output_dir" \
        --num_prompts $NUM_PROMPTS \
        --samples_per_class $SAMPLES_PER_CLASS \
        --image_size $IMAGE_SIZE \
        --num_workers $NUM_WORKERS \
        --force
    
    # Check if preparation was successful
    if [ $? -eq 0 ]; then
        # Record end time and calculate duration
        end_time=$(date +%s)
        duration=$((end_time - start_time))
        
        log_message "✅ $dataset_type preparation completed successfully!"
        log_message "⏱️  Processing time: ${duration} seconds ($(($duration / 60)) minutes)"
        
        # Run verification
        log_message "🔍 Running verification for $dataset_type..."
        python /work/jf381/code/code/ICL_Jay/data/prepare_datasets.py \
            --dataset "$dataset_type" \
            --output_dir "$output_dir" \
            --verify
        
        if [ $? -eq 0 ]; then
            log_message "✅ $dataset_type verification passed!"
        else
            log_message "❌ $dataset_type verification failed!"
            return 1
        fi
        
        # Show dataset statistics
        log_message "📊 Dataset statistics for $dataset_type:"
        if [ -f "$output_dir/dataset_metadata.json" ]; then
            python3 -c "
import json
import os

try:
    with open('$output_dir/dataset_metadata.json', 'r') as f:
        meta = json.load(f)
    
    print(f'  Total prompts: {meta.get(\"num_prompts\", \"N/A\")}')
    print(f'  Training prompts: {len(meta.get(\"train_prompts\", []))}')
    print(f'  Test prompts: {len(meta.get(\"test_prompts\", []))}')
    print(f'  Classes used: {len(meta.get(\"train_classes_used\", []))}')
    
    if 'data_split_info' in meta:
        split_info = meta['data_split_info']
        print(f'  Train split: {split_info.get(\"train_split_used\", \"N/A\")}')
        print(f'  Test split: {split_info.get(\"test_split_used\", \"N/A\")}')
        print(f'  No data leakage: {split_info.get(\"no_data_leakage\", False)}')
    
    if 'processing_stats' in meta:
        stats = meta['processing_stats']
        print(f'  Completed tasks: {stats.get(\"completed_tasks\", \"N/A\")}')
        print(f'  Failed tasks: {stats.get(\"failed_tasks\", \"N/A\")}')
        
except Exception as e:
    print(f'  Error reading metadata: {e}')
"
        fi
        
        # Calculate dataset size
        dataset_size=$(du -sh "$output_dir" 2>/dev/null | cut -f1 || echo 'N/A')
        log_message "  Dataset size: $dataset_size"
        
        return 0
    else
        log_message "❌ $dataset_type preparation failed!"
        return 1
    fi
}

# Main execution based on SLURM_ARRAY_TASK_ID
log_message "==================================================="
log_message "SLURM Array Dataset Generation Job"
log_message "==================================================="
log_message "SLURM_JOB_ID: $SLURM_JOB_ID"
log_message "SLURM_ARRAY_JOB_ID: $SLURM_ARRAY_JOB_ID"
log_message "SLURM_ARRAY_TASK_ID: $SLURM_ARRAY_TASK_ID"
log_message "HOSTNAME: $(hostname)"
log_message "GPU INFO: $(nvidia-smi --query-gpu=name --format=csv,noheader | head -1)"
log_message "==================================================="

# Check for required Python packages
log_message "Checking Python dependencies..."
python3 -c "
import sys
required_packages = ['datasets', 'torch', 'torchvision', 'PIL', 'numpy', 'tqdm']
missing_packages = []

for package in required_packages:
    try:
        if package == 'PIL':
            import PIL
        else:
            __import__(package)
        print(f'✅ {package}')
    except ImportError:
        missing_packages.append(package)
        print(f'❌ {package}')

if missing_packages:
    print(f'Missing packages: {missing_packages}')
    sys.exit(1)
else:
    print('All dependencies satisfied!')
"

if [ $? -ne 0 ]; then
    log_message "❌ Dependency check failed!"
    exit 1
fi

# Map SLURM_ARRAY_TASK_ID to dataset type
case $SLURM_ARRAY_TASK_ID in
    0)
        DATASET_TYPE="cifar10"
        log_message "📋 Task 0: Preparing CIFAR-10 dataset"
        log_message "    - 10 classes with automatic train/test split"
        log_message "    - Fast processing, good for initial testing"
        ;;
    1)
        DATASET_TYPE="cifar100"
        log_message "📋 Task 1: Preparing CIFAR-100 dataset"
        log_message "    - 100 classes with automatic train/test split"
        log_message "    - Moderate complexity, comprehensive testing"
        ;;
    2)
        DATASET_TYPE="imagenet10"
        log_message "📋 Task 2: Preparing ImageNet10 dataset"
        log_message "    - 10 carefully selected classes from ImageNet100"
        log_message "    - Optimized for in-context learning performance"
        log_message "    - Uses HuggingFace ilee0022/ImageNet100 dataset"
        ;;
    3)
        DATASET_TYPE="imagenet100"
        log_message "📋 Task 3: Preparing ImageNet100 dataset"
        log_message "    - 100 classes from HuggingFace dataset"
        log_message "    - Comprehensive evaluation dataset"
        log_message "    - Proper train/validation split from HuggingFace"
        ;;
    *)
        log_message "❌ Invalid SLURM_ARRAY_TASK_ID: $SLURM_ARRAY_TASK_ID"
        log_message "Valid IDs are 0-3 (cifar10, cifar100, imagenet10, imagenet100)"
        exit 1
        ;;
esac

# Execute dataset preparation
log_message "Starting dataset preparation for $DATASET_TYPE..."
prepare_dataset "$DATASET_TYPE"

if [ $? -eq 0 ]; then
    log_message "==================================================="
    log_message "✅ SUCCESS: $DATASET_TYPE dataset preparation completed!"
    log_message "==================================================="
    log_message "Dataset location: $BASE_OUTPUT_DIR/$DATASET_TYPE"
    log_message ""
    log_message "To use this dataset in your training script:"
    log_message "python your_training_script.py \\"
    log_message "    --data_type image_data \\"
    log_message "    --dataset_type $DATASET_TYPE \\"
    log_message "    --image_dir $BASE_OUTPUT_DIR/$DATASET_TYPE \\"
    log_message "    --use_vgg_features \\"
    log_message "    --image_noise_level 0.1"
    log_message ""
    log_message "For combined manifolds:"
    log_message "python your_training_script.py \\"
    log_message "    --data_type combine_manifold \\"
    log_message "    --manifold_list swiss_roll,sphere,image_data \\"
    log_message "    --dataset_type $DATASET_TYPE \\"
    log_message "    --image_dir $BASE_OUTPUT_DIR/$DATASET_TYPE \\"
    log_message "    --use_vgg_features"
    log_message "==================================================="
else
    log_message "==================================================="
    log_message "❌ FAILURE: $DATASET_TYPE dataset preparation failed!"
    log_message "==================================================="
    log_message "Check the error logs above for details."
    log_message "Common issues:"
    log_message "  - Network connectivity for HuggingFace datasets"
    log_message "  - Insufficient disk space"
    log_message "  - Missing Python dependencies"
    log_message "  - GPU memory issues"
    log_message "==================================================="
    exit 1
fi

log_message "Job completed at $(date)"
log_message "SLURM_ARRAY_TASK_ID $SLURM_ARRAY_TASK_ID finished successfully"