yewinhtet74/sdm-retrieval

How yewinhtet74/sdm-retrieval works

SDM-Retrieval is a research workflow for comparing training-free compression methods on frozen CLIP image and caption embeddings. Users configure a Flickr30k run, extract or load aligned features, optionally select a fixed training subset to fit a reducer, then evaluate bidirectional retrieval quality and FAISS latency/storage and save results. The graph preserves the documented baseline, PCA, SVD and SDM branches, plus diagnostics. Edges reflect the supplied excerpts, SOURCE INDEX references and README; unsampled data-loading and orchestration details are not inferred beyond those sources.

Main entry points: evaluate.py Evaluation orchestration · features.py Feature extraction · outputs Cached embeddings · reducers.py Reducer methods

Components 15

Data and configuration

Run configuration src/config.py
Flickr30k data
Dataset loading src/data.py

Embedding preparation

Frozen CLIP src/model.py
Feature extraction src/features.py
Cached embeddings outputs

Compression methods

Training subset src/subset.py
Reducer methods src/reducers.py
SDM cluster data clusters

Retrieval evaluation

Evaluation orchestration src/evaluate.py
Recall metrics src/metrics.py
FAISS benchmarks src/retrieval.py
I2T diagnostics src/diagnose_i2t.py
Evaluation results outputs/results

Other components

Researcher

Connections 16

Researcher → Run configuration: configures run README.md
Researcher → Flickr30k data: provides dataset README.md
Run configuration → Dataset loading: configures loading
Flickr30k data → Dataset loading: supplies records README.md
Dataset loading → Feature extraction: builds loaders src/features.py
Feature extraction → Frozen CLIP: loads encoder src/features.py
Feature extraction → Cached embeddings: writes embeddings src/features.py
Cached embeddings → Training subset: supplies training data
Training subset → Reducer methods: fits methods scripts/03_eval_reducers.py
Cached embeddings → Evaluation orchestration: supplies evaluation features src/evaluate.py
Reducer methods → Evaluation orchestration: transforms embeddings src/evaluate.py
Evaluation orchestration → Recall metrics: computes recall src/evaluate.py
Evaluation orchestration → FAISS benchmarks: benchmarks index src/evaluate.py
Evaluation orchestration → Evaluation results: saves result table src/evaluate.py
Researcher → I2T diagnostics: requests diagnosis
Reducer methods → SDM cluster data: uses SDM clusters

Drawn by GitDiagram on November 14, 2026 from the repository's file tree, README and source files. Read as Markdown ·