How asbj021/geneticpatchpruning works
GeneticPatchPruning explores Vision Transformer efficiency by selecting image patches, using genetic search to produce pruning records and training learned selectors that can be evaluated against CLIP baselines. The source index supports the main links among data collection, the genetic algorithm, CLIP scoring, selector training, and evaluation. The README gives only the product title, so broader evaluation and training branches are represented from the tree without assuming unshown call paths.
Main entry points: data_utils.py Dataset utilities · model.py Patch selectors · clip_model.py CLIP inference · runner.py Search runner
Components 22
Data and collection
- Dataset utilities
gpp/dataset/data_utils.py - Patch datasets
gpp/dataset/dataset_patch_selection.py - GA collection
gpp/data_collection/genetic_patch_pruning.py - Parallel collection
gpp/data_collection/parallel_gpp.py
Patch search
- Search runner
gpp/genetic_algo/runner.py - Genetic search
gpp/genetic_algo/ga.py - Fitness scoring
gpp/genetic_algo/scoring.py - CLIP inference
gpp/model/clip_model.py - Selection records
gpp/genetic_algo/io.py
Selector training
- Patch selectors
gpp/model/model.py - Selector training
gpp/train/train.py - Adaptive training
gpp/train/adaptive_train.py - GA distillation
gpp/train/ga_distill_train.py - Training workflow
gpp/train/new_train.py - Patch training
gpp/train/p_train.py
Evaluation
- Selector evaluation
gpp/eval/evaluate.py - Adaptive evaluation
gpp/eval/adaptive_eval.py - Distillation evaluation
gpp/eval/eval_ga_distill.py - Baseline comparison
gpp/eval/compare_og_ga.py - Benchmarks and analysis
gpp/eval
Other components
- Researcher
- Image datasets
Connections 36
- Researcher → GA collection: starts collection
- Researcher → Selector training: starts training
- Researcher → Selector evaluation: runs evaluation
- Image datasets → Dataset utilities: provides samples
- Dataset utilities → Patch datasets: uses
gpp/dataset/data_utils.py - GA collection → Dataset utilities: loads data
gpp/data_collection/genetic_patch_pruning.py - GA collection → Search runner: runs search
gpp/data_collection/genetic_patch_pruning.py - GA collection → CLIP inference: loads CLIP
gpp/data_collection/genetic_patch_pruning.py - Parallel collection → Dataset utilities: loads data
gpp/data_collection/parallel_gpp.py - Parallel collection → Search runner: runs search
gpp/data_collection/parallel_gpp.py - Parallel collection → CLIP inference: loads CLIP
gpp/data_collection/parallel_gpp.py - Search runner → Genetic search: invokes search
gpp/genetic_algo/runner.py - Search runner → CLIP inference: prepares inference
gpp/genetic_algo/runner.py - Search runner → Selection records: writes records
gpp/genetic_algo/runner.py - Genetic search → Fitness scoring: scores candidates
gpp/genetic_algo/ga.py - Fitness scoring → CLIP inference: evaluates patches
gpp/genetic_algo/scoring.py - Selector training → Dataset utilities: uses data
gpp/train/train.py - Selector training → Patch selectors: trains
gpp/train/train.py - Adaptive training → Dataset utilities: uses data
gpp/train/adaptive_train.py - Adaptive training → Patch selectors: trains
gpp/train/adaptive_train.py - GA distillation → Dataset utilities: uses data
gpp/train/ga_distill_train.py - GA distillation → Patch selectors: trains
gpp/train/ga_distill_train.py - Training workflow → Dataset utilities: uses data
gpp/train/new_train.py - Training workflow → Patch selectors: trains
gpp/train/new_train.py - Patch training → Dataset utilities: uses data
gpp/train/p_train.py - Patch training → Patch selectors: trains
gpp/train/p_train.py - Selector evaluation → Dataset utilities: loads data
gpp/eval/evaluate.py - Selector evaluation → Patch selectors: uses selector
gpp/eval/evaluate.py - Selector evaluation → CLIP inference: runs inference
gpp/eval/evaluate.py - Adaptive evaluation → Dataset utilities: loads data
gpp/eval/adaptive_eval.py - Adaptive evaluation → Patch selectors: uses selector
gpp/eval/adaptive_eval.py - Distillation evaluation → Dataset utilities: loads data
gpp/eval/eval_ga_distill.py - Distillation evaluation → Patch selectors: evaluates selector
gpp/eval/eval_ga_distill.py - Distillation evaluation → CLIP inference: runs inference
gpp/eval/eval_ga_distill.py - Baseline comparison → Dataset utilities: loads data
gpp/eval/compare_og_ga.py - Baseline comparison → CLIP inference: compares CLIP
gpp/eval/compare_og_ga.py