How lemonmindyes/minidream works
Project Type and Purpose
• MiniDream is a PyTorch‐based research toolkit rather than a client/server or web app.
• It implements “mini” variants of several state‐of‐the‐art models (BPE tokenizer, MobileNet, ResNet, U-Net, ViT, VQ-VAE).
• Each model lives in its own top‐level Python package and can be trained via one of the train_*.py scripts.
File Structure Analysis
• Top-level artifacts:
• modules: mini_bpe/, mini_mobilenet/, mini_resnet/, mini_unet/, mini_vit/, mini_vqvae/
• train scripts: train_mini_*.py
• requirements.txt, README.md, imgs/
• Each module contains:
• config.py — hyperparameter container
• model definition(s) (.py)
• utils.py or dataset.py where needed
• No server or database; dependencies are PyTorch and Lightning.
Insights from README
• Describes each model’s usage snippet and points to corresponding image in imgs/.
• Installation and training commands emphasize a CLI‐driven workflow.
• No external services; purely code modules and training pipelines.
a. Main Components
1. Model Modules (six packages)
• mini_bpe
• mini_mobilenet
• mini_resnet
• mini_unet
• mini_vit
• mini_vqvae
2. Training Controllers (train_mini_*.py scripts)
3. Shared Libraries
• PyTorch
• PyTorch Lightning
4. Auxiliary Artifacts
• Config objects (per‐model)
• Utils / dataset code
• Image assets (imgs/)
b. Relationships & Interactions
• Each train script:
• Imports Config and Model class from its module
• Instantiates Config → Model → Lightning Trainer → runs training
• Models call their own utils (e.g., tokenization, data transforms, visualization).
• Underlying ML framework (PyTorch, Lightning) provides tensor ops, training loop, checkpointing.
c. Architectural Patterns & Principles
• Package‐per‐model for modularity and extensibility.
• Config‐driven instantiation: separates hyperparameters from code.
• CLI entry points (train scripts) decouple experiments from module code.
d. Key Technologies
• Python packages (PEP 420)
• PyTorch as core DL library
• PyTorch Lightning for standardized training loops
• Standard Python utils for data loading/visualization
Detailed Capturing of Architectural Elements
• If desired, group the six model packages into a “Model Library” container.
• Directional flow:
1. User ▶ train_mini_resnet.py
2. ▶ Config (n_layers, num_class)
3. ▶ ResNet model graph
4. ▶ DataLoader / dataset utils
5. ▶ Lightning Trainer ▶ back to model & PyTorch
6. ▶ Outputs (checkpoints, logs)