How yuchuantian/dic works
Project Type and Purpose
• Type: Research-code repository implementing “DiC: Rethinking Conv3x3 Designs in Diffusion Models” in PyTorch.
• Purpose: Train, sample, and evaluate ConvUNet-based diffusion models (DiC) on image datasets or pre-computed VAE features.
Main System Components
A. Data Sources
• Raw ImageNet Dataset (folders train/, val/)
• Precomputed VAE Features (imagenet_feature/)
• External Model Hubs (HuggingFace, ModelScope) for VAE download
B. Core Code Modules
1. Model Definitions
• dic_models.py (ConvUNet architecture)
• models.py (additional network definitions)
2. Diffusion Engine (diffusion/)
• diffusion_utils.py (sampling loops, noise schedules)
• gaussian_diffusion.py (forward/backward diffusion math)
• respace.py (timestep rescaling)
• timestep_sampler.py (noise scheduling strategies)
3. Utilities
• utils/parser_setter.py (CLI argument parsing)
• utils/fastdit_kit.py (data loaders, VAE feature handling)
C. Orchestration Scripts
• Training
• train.py (fp32, torchrun)
• train_accelerate.py (fp16, accelerate)
• Sampling
• sample_ddp.py (distributed sampling)
• Evaluation
• evaluator.py (metrics on generated samples)
• Download Helper
• download.py (automate VAE / dataset pulls)
D. External Services & Artifacts
• HuggingFace / ModelScope (dataset and VAE hosting)
• Output Artifacts: model checkpoints, sampled images, evaluation logs
Relationships & Data Flow
1. Data Ingestion
• Optionally: Raw Images → VAE Encoder → Feature Cache
2. Training Pipeline
• CLI Script (train.py or train_accelerate.py)
↓
• Parser (parser_setter)
↓
• Model Instantiation (dic_models.py + diffusion modules)
↓
• Diffusion Process (gaussian_diffusion, timestep_sampler)
↓
• Optimizer / Backprop
↓
• Checkpoint Writer
3. Sampling Pipeline
• CLI Script (sample_ddp.py)
↓
• Load Checkpoint → Diffusion Engine → Generated Images
4. Evaluation Pipeline
• evaluator.py reads checkpoints & generated images → computes metrics
5. Download Utility
• download.py orchestrates pulls from HuggingFace/ModelScope
Architectural Patterns & Principles
• Modular separation: Data I/O, Model definition, Diffusion logic, Orchestration.
• Reusable components: diffusion/ folder encapsulates all scheduler/math.
• CLI-driven entrypoints for training, sampling, evaluation.
• Distributed/disaggregated execution via torch.distributed / accelerate.
Final Checklist
– Have you separated all 4 major layers (Data, Core Modules, Orchestrators, External)?
– Are all data flows directional and unambiguous?
– Is each code folder/file represented by a logical block?
– Did you highlight the VAE feature path vs. raw image path?