How rohanshad/cmr_transformer works
This repository provides pretrained cardiac-MRI video encoders plus a Lightning workflow for fine-tuning and evaluation. The main path is configured training or evaluation through the trainer, data module, dataset and video model, producing task predictions; a separate quickstart loads pretrained weights and returns embeddings. README-described report-supervised contrastive pretraining and downstream tasks are represented, but their unsampled internals are not expanded. Some links rely on README statements or the source index rather than inspected call sites.
Main entry points: model_factory.py Model factory · mri_trainer.py Trainer CLI · minimal_run.py Encoder quickstart · data_modules.py Data modules
Components 19
Training workflow
- Trainer CLI
mri_trainer.py - Task configs
configs - Encoder quickstart
minimal_run.py
MRI data
- Data modules
data/data_modules.py - MRI datasets
data/torch_datasets.py - MRI files
- Video transforms
data/data_modules.py
Encoders and tasks
- Model factory
model_factory.py - MViT encoder
arch/mvit.py - ResNet encoder
arch/resnet.py - Report text encoder
model_factory.py - Task heads
model_factory.py - Contrastive pretraining
model_factory.py - Pretrained weights
Embeddings and attention
- MRI embeddings
model_factory.py - Attention outputs
mri_trainer.py - Attention visualizer
visualize_attentions.py
Other components
- Researcher
- Hugging Face
Connections 20
- Researcher → Trainer CLI: runs task
README.md - Researcher → Encoder quickstart: runs encoder
README.md - Trainer CLI → Task configs: loads config
README.md - Trainer CLI → Data modules: uses data
mri_trainer.py - Data modules → MRI datasets: creates datasets
data/data_modules.py - MRI datasets → MRI files: reads scans
data/torch_datasets.py - Data modules → Video transforms: builds transforms
data/data_modules.py - Trainer CLI → Model factory: uses model
mri_trainer.py - Model factory → MViT encoder: constructs encoder
model_factory.py - Model factory → ResNet encoder: constructs encoder
model_factory.py - Model factory → Task heads: builds task head
model_factory.py - Model factory → Report text encoder: supports text model
README.md - Report text encoder → Contrastive pretraining: guides vision
README.md - Contrastive pretraining → MViT encoder: trains representations
README.md - Hugging Face → Pretrained weights: hosts checkpoint
README.md - Encoder quickstart → Pretrained weights: downloads weights
minimal_run.py - Encoder quickstart → Model factory: loads encoder
minimal_run.py - Model factory → MRI embeddings: returns embeddings
minimal_run.py - Trainer CLI → Attention outputs: writes maps
mri_trainer.py - Attention outputs → Attention visualizer: reads maps
visualize_attentions.py