How ananyapselvam/mdc-net-steel-defect-captioning works
MDC-Net is a steel-surface defect captioning project. Its implemented workflow parses NEU-DET XML annotations into captions and image metadata, splits records by image, loads and augments samples into batches, then trains and validates an image encoder–caption decoder. The README also describes evaluation and caption generation, but the sampled runtime code shows validation loss rather than a separate evaluation or inference-generation implementation; those documented stages are therefore marked as unverified. The graph follows visible calls and README statements; unsampled modules are not assumed to connect.
Main entry points: dataloader.py Batch assembly · model.py MDC-Net · train.py Training and validation · trainepoch.py Epoch training variant
Components 18
Data preparation
- XML caption preprocessing
src/preprocess_neu.py - Annotation records
- Image-level split
src/split_data.py - Train/val/test records
Input pipeline
- Image transforms
src/augmentation.py - Caption vocabulary
src/vocabulary.py - NEU dataset
src/dataset.py - Batch assembly
src/dataloader.py - Runtime settings
src/config.py
Caption model
- MDC-Net
src/model.py - Image encoder
src/model.py - Caption decoder
src/model.py - Defect caption generation
src
Training and evaluation
- Training and validation
src/train.py - Epoch training variant
src/trainepoch.py - Model checkpoint
Other components
- NEU-DET data
- timm model library
Connections 23
- NEU-DET data → XML caption preprocessing: provides XML/images
src/preprocess_neu.py - XML caption preprocessing → Annotation records: writes records
src/preprocess_neu.py - Annotation records → Image-level split: supplies annotations
src/split_data.py - Image-level split → Train/val/test records: writes splits
src/split_data.py - Batch assembly → Train/val/test records: reads train/val
src/dataloader.py - NEU dataset → Image transforms: applies transforms
src/dataset.py - NEU dataset → Caption vocabulary: numericalizes captions
src/dataset.py - NEU dataset → Runtime settings: uses settings
src/dataset.py - Batch assembly → NEU dataset: creates datasets
src/dataloader.py - Batch assembly → Caption vocabulary: builds vocabulary
src/dataloader.py - Batch assembly → Runtime settings: sets batch options
src/dataloader.py - Training and validation → Batch assembly: loads batches
src/train.py - Training and validation → MDC-Net: trains and validates
src/train.py - Training and validation → Runtime settings: uses settings
src/train.py - Epoch training variant → Batch assembly: loads batches
src/trainepoch.py - Epoch training variant → MDC-Net: trains and validates
src/trainepoch.py - Epoch training variant → Runtime settings: uses settings
src/trainepoch.py - MDC-Net → Image encoder: runs encoder
src/model.py - MDC-Net → Caption decoder: runs decoder
src/model.py - Image encoder → timm model library: creates backbone
src/model.py - Training and validation → Model checkpoint: saves best weights
src/train.py - Epoch training variant → Model checkpoint: saves weights
src/trainepoch.py - MDC-Net → Defect caption generation: supports captioning
README.md