PBF-Layer-Seg

Laser Powder Bed Fusion Layer Instance Segmentation for Anomaly / Defect Identification

Requirements

• Python v3.12.10
• uv Python Package Manager
• CUDA Toolkit v12.6

My System

The following is a break-down of my system configuration:

• CPU: AMD Ryzen 7 5800X 8-Core Processor
• GPU: NVIDIA GeForce RTX 3090 w/ 24GB VRAM
• RAM: 128GB DDR4

Setup Instructions

uv Package Manager

1. Open a command prompt in the root directory, and run uv sync. This will configure the necessary Python packages to run the code.

Dataset Setup

1. Download the three Peregrine datasets:

• v2021-03
• v2022-10.1
• v2025-09

2. Move the data into a parent Peregrine folder:

Peregrine/                          <--- DATABASE_BASE_DIR
├── Peregrine Dataset v2021-03/
├── Peregrine Dataset v2022-10.1/
└── Peregrine Dataset v2025-09/

3. Download the model weights from here, and place in models/UNet/checkpoints/.

Configuration Setup

Configure the parameters of the repository, found within the configs/config.py file, such as:

1. NUM_WORKERS: Should be set to the number of cores available on your CPU.
2. DATASET_BASE_DIR: The full path to the root Peregrine folder, set in the previous section.

Dataset Standardization and Train/Val/Test Split

With the above configured, run the following command to standardize and combine the three datasets into a Unified_Unet_Dataset folder:

uv run python setup.py

Functionality

With the above configured, you can perform the following functionalities.

Training

To train your own model on the data, run the following:

uv run python train.py

Testing

To test on the held-out testing data using the trained model:

1. Set the TEST_DATA_DIR parameter in config.py to point to test/AddUp_FormUp_350_Maraging_Steel, test/EOS_M290_Inconel_718_2, or test/Concept_Laser_M2_Builds

2. Run the following command:

   uv run python test_suite_analysis.py

Inference

To conduct inference on your own data using the trained model, run the following:

uv run python inference.py <PATH_TO_IMAGE>

Overview and Analysis Files

For more metadata information, review the following write-ups:

• Dataset Overview
• Model Overview
• Fine-Tuned Model Evaluation

TODO Checklist

In future iterations, I plan to develop the following capabilities and features:

• Automatically export trained weights to the following to speed-up inference:
  • ONNX
  • TensorRT
• Compare performance metrics against larger pre-trained EfficientNet Encoders
• Quantify epistemic uncertainty (model confidence) by integrating Mote Carlo (MC) dropout as a toggle for testing and inference
• Conduct multi-class classification, and compare performance against binary classification