How nimit/mech_synthesis works
This repository is a research codebase for generating and evaluating planar linkage mechanisms. Its workflow spans dataset preparation, learned image/latent representations, coordinate-sequence prediction, and retrieval or motion-generation experiments. The tree also shows separate continuous-latent and vision-transformer training tracks, plus embedding visualization and dataset inspection. The README is empty, and sampled excerpts cover only parts of the runtime; the graph therefore includes documented-by-name capability areas but omits uncertain connections rather than assuming a unified pipeline.
Main entry points: dataset.py Linkage datasets · llama_latent_model.py Latent sequence model · creating_dataset_files_V2.py Dataset creation · generating_walking_mechanisms.py Walking mechanisms
Components 18
Data preparation
- Dataset creation
dataset_generation/creating_dataset_files_V2.py - Dataset tools
dataset_generation - Linkage datasets
dataset.py - Continuous dataset
dataset_continuous.py
Learning models
- VAE encoder
vae.py - Vision transformer
vit/vit_model.py - Vision training
vit/vit_train_V1.py - Latent sequence model
llama_latent_model.py - Latent training
llama_latent_train.py - Continuous model
llama_latent_continuous.py - Continuous training
llama_latent_continuous_train.py
Retrieval and analysis
- Sequence inference
helpers/inference.py - Embedding analysis
helpers/embedding_TSNE.py - DTW retrieval
cd_dtw_serach.py - Continuous DTW
dtw_calc_for_knn_continuous.py
Motion generation
- Walking mechanisms
walking_mechanism_generation/generating_walking_mechanisms.py
Other components
- Researcher
- Pose service
Connections 18
- Researcher → Dataset creation: prepares data
- Dataset creation → Linkage datasets: provides files
- Dataset tools → Dataset creation: supports preparation
- Linkage datasets → Vision training: supplies samples
- Vision training → Vision transformer: trains
- Linkage datasets → Latent training: supplies samples
- Latent training → Latent sequence model: trains
- Continuous dataset → Continuous training: supplies samples
- Continuous training → Continuous model: trains
- Linkage datasets → Sequence inference: loads samples
- Latent sequence model → Sequence inference: predicts tokens
- Vision transformer → Embedding analysis: provides embeddings
- Continuous model → Continuous DTW: predicts curves
- Continuous dataset → Continuous DTW: loads latents
- Researcher → Walking mechanisms: requests generation
- Latent sequence model → Walking mechanisms: decodes coordinates
- Walking mechanisms → Pose service: requests poses
- Researcher → DTW retrieval: runs retrieval