How spaicom/semasia-datasets works
SEMASIA is a benchmark and analysis toolkit for large-scale pretrained vision-model embeddings. Its principal workflow encodes benchmark images with timm models, stores latent representations, and runs geometric, graph/topological, prototype, and alignment analyses; model metadata also supports regression studies. The graph keeps these domain stages distinct and includes the reusable latent-space and graph-analysis layers. README-described capabilities are included, but many script-to-module connections cannot be verified from the sampled source, so those links are omitted.
Main entry points: encode_dataset_all_timm.py Dataset encoding · latent.py Latent-space operations · graph.py Graph analysis · alignment.py Alignment evaluation
Components 25
Embedding generation
- Dataset encoding
scripts/encode_dataset_all_timm.py - Timm encoder
src/models/timm.py - Latent dataset
Latent analysis
- Metric extraction
scripts/compute_metrics.py - Latent-space operations
src/objects/latent.py - Prototype metrics
src/metrics/prototypes.py - Graph analysis
src/objects/graph.py - Graph metrics
src/metrics/graph.py - KNN construction
src/tsp/graph_inference.py - Spectral methods
src/tsp/laplace.py - TDA features
src/tda/features.py - TDA extraction
scripts/tda_extraction.py - Graph signatures
scripts/tsp_extraction.py - Latent visualizations
src/plotting/latent.py - TDA visualizations
src/plotting/tda.py
Space alignment
- Alignment evaluation
scripts/alignment.py - Alignment methods
src/objects/alignment.py - Alignment metrics
src/metrics/alignment.py - Prototype comparison
scripts/prototype_comparison.py
Statistical insights
- Metadata regression
scripts/stat_analysis.py - Model metadata
src/utils/timm_metadata.py - Metadata visualization
src/visualizations/parallel_coordinates.py
Other components
- Researcher
- Benchmark images
- Pretrained models
Connections 30
- Researcher → Dataset encoding: starts encoding
- Benchmark images → Dataset encoding: supplies images
- Pretrained models → Dataset encoding: provides encoders
- Dataset encoding → Timm encoder: uses model encoder
- Dataset encoding → Latent dataset: writes embeddings
- Researcher → Metric extraction: requests metrics
- Latent dataset → Metric extraction: supplies latents
- Researcher → TDA extraction: requests TDA
- Latent dataset → TDA extraction: supplies latents
- Researcher → Graph signatures: requests signatures
- Latent dataset → Graph signatures: supplies latents
- Researcher → Alignment evaluation: starts evaluation
- Latent dataset → Alignment evaluation: supplies paired spaces
- Researcher → Prototype comparison: requests comparison
- Latent dataset → Prototype comparison: supplies model spaces
- Researcher → Metadata regression: starts regression
- Model metadata → Metadata regression: provides metadata
- Metric extraction → Latent-space operations: analyzes spaces
- Latent-space operations → Prototype metrics: supports prototypes
- Latent-space operations → Graph analysis: provides embeddings
src/objects/latent.py - Graph analysis → KNN construction: builds neighborhood graph
src/objects/graph.py - Graph analysis → Spectral methods: computes spectrum
src/objects/graph.py - Graph analysis → Graph metrics: computes metrics
src/objects/graph.py - KNN construction → Spectral methods: passes adjacency
src/tsp/graph_inference.py - TDA extraction → TDA features: extracts features
- Alignment evaluation → Alignment methods: evaluates methods
- Alignment evaluation → Alignment metrics: scores results
- Latent-space operations → Latent visualizations: visualizes spaces
src/plotting/latent.py - Researcher → TDA visualizations: views TDA results
- Metadata regression → Metadata visualization: visualizes statistics