How hao1635/langmamba works
LangMamba is a low-dose CT denoising framework with a Mayo-data training workflow and a documented two-stage setup: prepare candidate tokens and train a VQGAN-style autoencoder, then train the LEDA denoiser. The graph captures the sampled training calls and README-described model stages; inference is documented, but its test entry point is absent from the supplied tree, so its wiring is not shown.
Main entry points: train.py Training loop · langmamba_model.py LangMamba model · vqgan.py VQGAN autoencoder · cond_transformer.py Conditional transformer
Components 19
Training workflow
- Training loop
train.py - Run options
options/train_options.py
CT data
- Dataset factory
data/__init__.py - Mayo CT pairs
data/mayo16_dataset.py - Dataset transforms
data/base_dataset.py
LEDA denoiser
- Model factory
models/__init__.py - Model lifecycle
models/base_model.py - LangMamba model
models/langmamba_model.py - Network components
models/networks.py - Selective scan
models/emamba.py - LEDA loss
models/leda_loss.py
Token autoencoder
- Candidate tokens
process_words.py - VQGAN autoencoder
models/taming/models/vqgan.py - Image encoder-decoder
models/taming/modules/diffusionmodules/model.py - Vector quantizer
models/taming/modules/vqvae/quantize.py - Conditional transformer
models/taming/models/cond_transformer.py - Token transformer
models/taming/modules/transformer/mingpt.py
Other components
- Researcher
- Denoised CT
Connections 19
- Researcher → Training loop: starts training
- Training loop → Run options: parses settings
- Training loop → Dataset factory: creates dataset
- Dataset factory → Mayo CT pairs: selects dataset
- Mayo CT pairs → Dataset transforms: uses transforms
- Training loop → Model factory: creates model
- Model factory → LangMamba model: selects model
- Training loop → Model lifecycle: sets up and optimizes
- LangMamba model → Model lifecycle: extends lifecycle
- LangMamba model → Network components: builds network
- Network components → Selective scan: uses scan blocks
- LangMamba model → LEDA loss: computes loss
- LangMamba model → Denoised CT: predicts denoised CT
- Researcher → Candidate tokens: prepares tokens
- Candidate tokens → VQGAN autoencoder: supplies candidates
- VQGAN autoencoder → Image encoder-decoder: encodes and decodes
- VQGAN autoencoder → Vector quantizer: quantizes latents
- Conditional transformer → VQGAN autoencoder: uses first stage
- Conditional transformer → Token transformer: uses transformer