The current ATLAS v17 mobile AI/ML source is canonical under ATLAS/: Python source grouped by purpose and the complete mobile application with its bundled models.
Start current mobile work in ATLAS/python/ and ATLAS/mobile_app/slt_mobile_app/. Former Python paths and the sibling mobile-project path remain compatibility links. Read PROJECT_GROUND_TRUTH.md for current model state and training/evaluation gates. The older v16 and DS-GCN pipelines remain preserved for historical callers.
| Stage | Current path | Purpose |
|---|---|---|
| Stage 0 | active/v16/extract_v16.py | Apple Vision-style landmark extraction and v16 feature formatting |
| Stage 1 | active/v16/train_stage_1_v16.py | Isolated sign classification with Squeezeformer |
| Stage 2 | active/v16/train_stage_2_v16_fixed.py | Continuous recognition with CTC and anti-template-memorization fixes |
| Inference | active/v16/inference_v16.py | Unified video inference with trimming, sliding windows, TTA, and CTC beam search |
| Compatibility | src_v16/*.py | Thin wrappers for existing scripts and mobile export tools |
See CURRENT_PIPELINE.md for commands and cleanup notes.
src/ keeps the older DS-GCN/Transformer implementation and desktop/demo utilities.legacy/copies/ contains duplicate * copy.py files moved out of the active tree.docs/archive/ is for older planning and review notes that may describe pre-v16 architecture.artifacts/ is for generated metrics, reports, charts, and exported outputs.data/local/; root dataset names are compatibility symlinks.artifacts/model_assets/; models and weights are compatibility symlinks.scripts/extraction/; root extractor files are compatibility wrappers.# v16 inference through the canonical path
KMP_DUPLICATE_LIB_OK=TRUE python active/v16/inference_v16.py path/to/video.mp4
# v16 inference through the compatibility path
KMP_DUPLICATE_LIB_OK=TRUE python src_v16/inference_v16.py path/to/video.mp4
# v16 Stage 1 training
python active/v16/train_stage_1_v16.py \
--data_path src_v16/ASL_landmarks_v16 \
--save_dir models/output_v16_d384_aug \
--manifest models/manifest_v16.json
# v16 Stage 2 training
python active/v16/train_stage_2_v16_fixed.py \
--data_path src_v16/ASL_landmarks_v16 \
--phrase_data src_v16/ASL_phrases_v16 \
--stage1_ckpt src_v16/output_v16_d384/best_model.pth \
--manifest models/manifest_v16.json
active/v16/ for code changes.src_v16/ only when maintaining older scripts that already import it.data/local/, artifacts/generated/, and artifacts/reports/.The current ATLAS v17 mobile AI/ML source is canonical under ATLAS/: Python source grouped by purpose and the complete mobile application with its bundled models.
Start current mobile work in ATLAS/python/ and ATLAS/mobile_app/slt_mobile_app/. Former Python paths and the sibling mobile-project path remain compatibility links. Read PROJECT_GROUND_TRUTH.md for current model state and training/evaluation gates. The older v16 and DS-GCN pipelines remain preserved for historical callers.
| Stage | Current path | Purpose |
|---|---|---|
| Stage 0 | active/v16/extract_v16.py | Apple Vision-style landmark extraction and v16 feature formatting |
| Stage 1 | active/v16/train_stage_1_v16.py | Isolated sign classification with Squeezeformer |
| Stage 2 | active/v16/train_stage_2_v16_fixed.py | Continuous recognition with CTC and anti-template-memorization fixes |
| Inference | active/v16/inference_v16.py | Unified video inference with trimming, sliding windows, TTA, and CTC beam search |
| Compatibility | src_v16/*.py | Thin wrappers for existing scripts and mobile export tools |
See CURRENT_PIPELINE.md for commands and cleanup notes.
src/ keeps the older DS-GCN/Transformer implementation and desktop/demo utilities.legacy/copies/ contains duplicate * copy.py files moved out of the active tree.docs/archive/ is for older planning and review notes that may describe pre-v16 architecture.artifacts/ is for generated metrics, reports, charts, and exported outputs.data/local/; root dataset names are compatibility symlinks.artifacts/model_assets/; models and weights are compatibility symlinks.scripts/extraction/; root extractor files are compatibility wrappers.# v16 inference through the canonical path
KMP_DUPLICATE_LIB_OK=TRUE python active/v16/inference_v16.py path/to/video.mp4
# v16 inference through the compatibility path
KMP_DUPLICATE_LIB_OK=TRUE python src_v16/inference_v16.py path/to/video.mp4
# v16 Stage 1 training
python active/v16/train_stage_1_v16.py \
--data_path src_v16/ASL_landmarks_v16 \
--save_dir models/output_v16_d384_aug \
--manifest models/manifest_v16.json
# v16 Stage 2 training
python active/v16/train_stage_2_v16_fixed.py \
--data_path src_v16/ASL_landmarks_v16 \
--phrase_data src_v16/ASL_phrases_v16 \
--stage1_ckpt src_v16/output_v16_d384/best_model.pth \
--manifest models/manifest_v16.json
active/v16/ for code changes.src_v16/ only when maintaining older scripts that already import it.data/local/, artifacts/generated/, and artifacts/reports/.