How leaplabthu/lineardiff works
LinearDiff is a PyTorch research implementation with two user-facing workflows: ImageNet classification using VCA-DeiT, and class-conditioned image generation using VCA-DiT or VCA-SiT, followed by sample evaluation. The graph preserves those separate training and inference paths and the SiT transport layer. Several links to generation internals rely on the documented workflow and repository layout; unsampled wiring is not inferred beyond those documented roles.
Main entry points: main.py Train / evaluate · dit/sample_ddp_nonpz.py Sample DiT · train.py Train SiT · sit/sample_ddp_nonpz.py Sample SiT
Components 24
Classification
- Train / evaluate
image_classification/main.py - Data pipeline
image_classification/data/build.py - Image dataset access
image_classification/data/cached_image_folder.py - VCA-DeiT
image_classification/models/vca_deit.py - Optimization schedule
image_classification/optimizer.py - Learning-rate schedule
image_classification/lr_scheduler.py - Classifier checkpoint
DiT Generation
- Train DiT
image_generation/models/dit/train.py - VCA-DiT model
image_generation/models/dit/models.py - Gaussian diffusion
image_generation/models/dit/diffusion/gaussian_diffusion.py - Sample DiT
image_generation/models/dit/sample_ddp_nonpz.py - DiT checkpoint
SiT Generation
- Train SiT
image_generation/models/sit/train.py - VCA-SiT model
image_generation/models/sit/models.py - Transport and sampler
image_generation/models/sit/transport/transport.py - Sample SiT
image_generation/models/sit/sample_ddp_nonpz.py - SiT checkpoint
Evaluation
- Build reference stats
image_generation/tools/create_ref_npz.py - FID evaluation
image_generation/tools/evaluator_read.py
Other components
- Researcher
- ImageNet data
- Generated samples
- Reference statistics
- Evaluation metrics
Connections 34
- Researcher → Train / evaluate: starts
- ImageNet data → Data pipeline: provides images
- Train / evaluate → Data pipeline: builds loaders
- Data pipeline → Image dataset access: loads images
- Train / evaluate → VCA-DeiT: builds model
- Train / evaluate → Optimization schedule: builds optimizer
- Train / evaluate → Learning-rate schedule: builds schedule
- Train / evaluate → Classifier checkpoint: saves checkpoint
- Researcher → Train DiT: starts
- ImageNet data → Train DiT: provides images
- Train DiT → VCA-DiT model: trains
- Train DiT → DiT checkpoint: saves checkpoint
- Researcher → Sample DiT: starts
- DiT checkpoint → Sample DiT: loads checkpoint
- Sample DiT → VCA-DiT model: samples with
- Sample DiT → Gaussian diffusion: uses diffusion
- Sample DiT → Generated samples: writes samples
- Researcher → Train SiT: starts
- ImageNet data → Train SiT: provides images
- Train SiT → VCA-SiT model: trains
- Train SiT → Transport and sampler: computes losses
- Train SiT → SiT checkpoint: saves checkpoint
- Researcher → Sample SiT: starts
- SiT checkpoint → Sample SiT: loads checkpoint
- Sample SiT → VCA-SiT model: samples with
- Sample SiT → Transport and sampler: uses sampler
- Sample SiT → Generated samples: writes samples
- ImageNet data → Build reference stats: provides images
- Researcher → Build reference stats: creates reference
- Build reference stats → Reference statistics: writes statistics
- Researcher → FID evaluation: runs evaluation
- Generated samples → FID evaluation: provides samples
- Reference statistics → FID evaluation: provides baseline
- FID evaluation → Evaluation metrics: computes metrics