Material Data Science Model Garden (MDSMG)

A comprehensive deep learning framework for material image segmentation using various encoder-decoder architectures.

Features

• Multiple encoder architectures: ResNet50, SE-ResNeXt101
• Multiple decoder architectures: UNet, DeepLabV3, DeepLabV3+
• Multiple pretraining strategies: ImageNet, MicroNet
• Support for various material datasets
• Command-line interface for easy experimentation
• Automatic result visualization and analysis

Requirements

pip install torch torchvision
pip install segmentation-models-pytorch
pip install opencv-python
pip install scikit-learn
pip install pandas matplotlib
pip install tqdm pillow
pip install huggingface-hub

Dataset Structure

Organize your datasets in the following structure:

project_root/
├── mdsmg_experiment.py
├── AFM/
│   ├── images/
│   │   ├── image1.png
│   │   ├── image2.tif
│   │   └── ...
│   └── masks/
│       ├── image1.png
│       ├── image2.tif
│       └── ...
├── SEM/
│   ├── images/
│   └── masks/
└── ...

Supported image formats: PNG, TIF/TIFF, JPG/JPEG

Basic Usage

1. Single Dataset, Single Configuration

python mdsmg_experiment.py --datasets AFM --encoders resnet50 --decoders unet --pretrained imagenet

2. Multiple Configurations

python mdsmg_experiment.py --datasets AFM SEM --encoders resnet50 se_resnext101 --decoders unet deeplabv3 --pretrained imagenet micronet

3. Quick Test Run

python mdsmg_experiment.py --datasets AFM --epochs 10 --batch-size 1 --image-size 128

4. Full Experiment Suite

python mdsmg_experiment.py --datasets AFM SEM Carbon --encoders resnet50 se_resnext101 --decoders unet deeplabv3 deeplabv3plus --pretrained imagenet micronet --epochs 200 --visualize

Command Line Arguments

Argument	Type	Default	Description
--datasets	list	['AFM']	Dataset names to use
--encoders	list	['resnet50']	Encoder architectures
--decoders	list	['unet']	Decoder architectures
--pretrained	list	['imagenet']	Pretraining strategies
--epochs	int	50	Number of training epochs
--batch-size	int	2	Batch size for training
--image-size	int	256	Input image size (square)
--device	str	'cuda'	Device to use (cuda/cpu)
--output	str	'experiment_results.csv'	Output CSV filename
--visualize	flag	False	Create visualization plots

Available Options

Encoders

• resnet50: ResNet-50 backbone
• se_resnext101: SE-ResNeXt-101 backbone

Decoders

• unet: U-Net decoder
• deeplabv3: DeepLabV3 decoder
• deeplabv3plus: DeepLabV3+ decoder

Pretraining

• imagenet: ImageNet pretrained weights
• micronet: MicroNet pretrained weights (specialized for microscopy)

Example Commands

Research Experiment

# Full comparison study
python mdsmg_experiment.py \
    --datasets AFM SEM Carbon \
    --encoders resnet50 se_resnext101 \
    --decoders unet deeplabv3 deeplabv3plus \
    --pretrained imagenet micronet \
    --epochs 200 \
    --batch-size 2 \
    --image-size 256 \
    --visualize \
    --output full_study_results.csv

Quick Prototyping

# Fast testing with small configuration
python mdsmg_experiment.py \
    --datasets AFM \
    --encoders resnet50 \
    --decoders unet \
    --epochs 20 \
    --batch-size 4 \
    --image-size 128

CPU Training

# For systems without GPU
python mdsmg_experiment.py \
    --datasets AFM \
    --device cpu \
    --batch-size 1 \
    --epochs 10

Memory-Constrained Training

# For limited GPU memory
python mdsmg_experiment.py \
    --datasets AFM \
    --batch-size 1 \
    --image-size 128 \
    --epochs 50

Output Files

The script generates several output files:

1. CSV Results (experiment_results.csv): Detailed metrics for each configuration
2. Model Weights (best_*.pth): Best model weights for each configuration
3. Visualizations (experiment_results.png): Performance comparison plots (if --visualize is used)

Performance Monitoring

The script automatically displays:

• Training progress with loss values
• Validation metrics (IoU, Dice coefficient)
• Best model performance
• GPU memory usage
• Experiment completion status

Memory Management

For systems with limited GPU memory:

1. Reduce batch size: --batch-size 1
2. Reduce image size: --image-size 128
3. Use CPU: --device cpu
4. Close other GPU applications

Troubleshooting

Common Issues

1. CUDA Out of Memory

   # Solution: Reduce batch size and image size
   python mdsmg_experiment.py --batch-size 1 --image-size 128

2. Dataset Not Found

   # Ensure dataset folder structure is correct
   # Check that images/ and masks/ folders exist
   

3. MicroNet Weights Loading Failed

   # This is normal - will fall back to ImageNet weights
   # MicroNet weights are optional
   

Performance Tips

1. Faster Training: Use larger batch sizes if memory allows
2. Better Results: Use larger image sizes (512x512) if memory allows
3. Quick Testing: Start with 10-20 epochs for initial experiments
4. Production: Use 100-200 epochs for final results

Results Interpretation

The CSV output contains:

• best_val_iou: Best validation IoU score (higher is better)
• final_val_dice: Final validation Dice coefficient
• config: Model configuration string
• train_samples: Number of training samples
• val_samples: Number of validation samples

Typical IoU scores:

• > 0.8: Excellent
• 0.6-0.8: Good
• 0.4-0.6: Fair
• < 0.4: Poor

Citation

@software{Materials Data Segmentation Benchmark,
  title={Context Determines Optimal Architecture in Materials Segmentation},
  author={Lu, Mingjian and Tripathi, Pawan Kumar and Shteyn, Mark and Ganguly, Debargha and French, Roger H. and Chaudhary, Vipin and Wu, Yinghui},
  year={2026},
  url={https://openreview.net/forum?id=kYqWdpXCx8}
}