Upload configs/ablations.yaml with huggingface_hub

configs/ablations.yaml

@@ -0,0 +1,142 @@
+# Ablation Study Configurations
+#
+# Each ablation isolates one variable to measure its contribution.
+# Run: python scripts/run_ablations.py --config configs/ablations.yaml
+#
+# Baseline: SCRFD-34G with default settings
+
+ablations:
+  # 1. Sample Redistribution (SCRFD's key contribution)
+  sample_redistribution:
+    description: "Impact of large-scale crops on small face detection"
+    experiments:
+      - name: "baseline_crops"
+        data.crop_scales: [0.3, 0.45, 0.6, 0.8, 1.0]
+        note: "Standard crops only (no SR)"
+      - name: "extended_crops"
+        data.crop_scales: [0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]
+        note: "Full Sample Redistribution (SCRFD default)"
+      - name: "extreme_crops"
+        data.crop_scales: [0.3, 0.6, 1.0, 1.5, 2.0, 2.5, 3.0]
+        note: "Even larger crops — test limits"
+
+  # 2. Loss Function Choice
+  loss_functions:
+    description: "GFL vs Focal Loss for classification"
+    experiments:
+      - name: "gfl_diou"
+        model.losses.cls: gfl
+        model.losses.reg: diou
+        note: "SCRFD default"
+      - name: "focal_diou"
+        model.losses.cls: focal
+        model.losses.reg: diou
+        note: "Standard focal loss"
+      - name: "gfl_smoothl1"
+        model.losses.cls: gfl
+        model.losses.reg: smooth_l1
+        note: "GFL + Smooth L1 regression"
+
+  # 3. Anchor Matching Strategy
+  matching:
+    description: "ATSS vs fixed IoU threshold matching"
+    experiments:
+      - name: "atss_k9"
+        model.matching.type: atss
+        model.matching.topk: 9
+        note: "ATSS with k=9 (SCRFD default)"
+      - name: "atss_k5"
+        model.matching.type: atss
+        model.matching.topk: 5
+        note: "ATSS with smaller k"
+      - name: "iou_035"
+        model.matching.type: iou
+        model.matching.pos_iou: 0.35
+        model.matching.neg_iou: 0.3
+        note: "Fixed IoU (TinaFace-style lower threshold)"
+      - name: "iou_050"
+        model.matching.type: iou
+        model.matching.pos_iou: 0.5
+        model.matching.neg_iou: 0.4
+        note: "Standard IoU threshold"
+
+  # 4. Robustness Augmentations
+  robustness:
+    description: "Impact of blur/compression/lighting augmentations"
+    experiments:
+      - name: "no_robustness"
+        data.robustness_augmentation: false
+        note: "Standard aug only"
+      - name: "blur_only"
+        data.robustness_augmentation: true
+        data.robustness_blur_prob: 0.3
+        data.robustness_jpeg_prob: 0.0
+        data.robustness_lowlight_prob: 0.0
+        note: "Blur augmentation only"
+      - name: "compression_only"
+        data.robustness_augmentation: true
+        data.robustness_blur_prob: 0.0
+        data.robustness_jpeg_prob: 0.3
+        data.robustness_lowlight_prob: 0.0
+        note: "JPEG compression only"
+      - name: "full_robustness"
+        data.robustness_augmentation: true
+        note: "All robustness augmentations"
+
+  # 5. GroupNorm vs BatchNorm
+  normalization:
+    description: "GN vs BN (TinaFace found GN better at small batch sizes)"
+    experiments:
+      - name: "group_norm"
+        model.neck.use_gn: true
+        model.head.use_gn: true
+        note: "GroupNorm (SCRFD/TinaFace default)"
+      - name: "batch_norm"
+        model.neck.use_gn: false
+        model.head.use_gn: false
+        note: "Standard BatchNorm"
+
+  # 6. Input Resolution
+  input_resolution:
+    description: "Trade-off between resolution and speed for tiny faces"
+    experiments:
+      - name: "res_320"
+        data.input_size: 320
+        inference.input_size: 320
+      - name: "res_480"
+        data.input_size: 480
+        inference.input_size: 480
+      - name: "res_640"
+        data.input_size: 640
+        inference.input_size: 640
+      - name: "res_960"
+        data.input_size: 960
+        inference.input_size: 960
+        note: "Higher res for tiny faces (+5-10% Hard AP, ~4x slower)"
+
+  # 7. Landmark Multi-Task Learning
+  landmarks:
+    description: "RetinaFace finding: auxiliary landmark loss improves detection"
+    experiments:
+      - name: "no_landmarks"
+        model.head.use_landmarks: false
+      - name: "with_landmarks"
+        model.head.use_landmarks: true
+        model.losses.lmk_weight: 0.1
+        note: "Requires RetinaFace landmark annotations"
+
+  # 8. Tracker Configuration
+  tracking:
+    description: "ByteTrack hyperparameter sensitivity"
+    experiments:
+      - name: "no_tracking"
+        engine.use_tracking: false
+      - name: "conservative"
+        engine.tracker.high_thresh: 0.6
+        engine.tracker.low_thresh: 0.2
+        engine.tracker.max_lost: 15
+      - name: "aggressive"
+        engine.tracker.high_thresh: 0.3
+        engine.tracker.low_thresh: 0.05
+        engine.tracker.max_lost: 60
+        note: "More tracks, more false positives"