Phase 3 classifier

classifier/configs/phase3/_base.json

@@ -0,0 +1,8 @@
+{
+  "pretrained": true,
+  "epochs": 15,
+  "image_size": 224,
+  "subsample": 0.2,
+  "augment": false,
+  "data_dir": "cropped/classifier"
+}

classifier/configs/phase3/p3_convnext_tiny.json

@@ -0,0 +1,5 @@
+{
+  "extends": "_base.json",
+  "run_name": "p3_convnext_tiny",
+  "backbone": "convnext_tiny"
+}

classifier/configs/phase3/p3_efficientnet_b0.json

@@ -0,0 +1,5 @@
+{
+  "extends": "_base.json",
+  "run_name": "p3_efficientnet_b0",
+  "backbone": "efficientnet_b0"
+}

classifier/configs/phase3/p3_mobilenetv3_small.json

@@ -0,0 +1,5 @@
+{
+  "extends": "_base.json",
+  "run_name": "p3_mobilenetv3_small",
+  "backbone": "mobilenet_v3_small"
+}

classifier/configs/phase3/p3_resnet34.json

@@ -0,0 +1,5 @@
+{
+  "extends": "_base.json",
+  "run_name": "p3_resnet34",
+  "backbone": "resnet34"
+}

classifier/configs/phase3/p3_resnet50.json

@@ -0,0 +1,5 @@
+{
+  "extends": "_base.json",
+  "run_name": "p3_resnet50",
+  "backbone": "resnet50"
+}

classifier/src/models/__init__.py

@@ -30,6 +30,6 @@ def load_checkpoint(model: nn.Module, path: Union[Path, str], device) -> nn.Modu
 
 
 # Importing the modules triggers their register() calls
-from src.models import simple_cnn, resnet, efficientnet  # noqa: E402, F401
+from src.models import simple_cnn, resnet, efficientnet, mobilenet, convnext  # noqa: E402, F401
 
 __all__ = ["get_model", "load_checkpoint", "register"]

classifier/src/models/convnext.py

@@ -0,0 +1,27 @@
+import torch.nn as nn
+from torchvision import models
+
+from src.models import register
+
+
+# ConvNeXt's classification head is a Sequential ending in (LayerNorm2d, Flatten, Linear); [-1] targets the Linear
+def build(cfg: dict) -> nn.Module:
+    backbone = cfg.get("backbone", "convnext_tiny")
+    pretrained = cfg.get("pretrained", True)
+
+    if backbone == "convnext_tiny":
+        weights = models.ConvNeXt_Tiny_Weights.DEFAULT if pretrained else None
+        model = models.convnext_tiny(weights=weights)
+    elif backbone == "convnext_small":
+        weights = models.ConvNeXt_Small_Weights.DEFAULT if pretrained else None
+        model = models.convnext_small(weights=weights)
+    else:
+        raise ValueError(f"Unsupported ConvNeXt backbone: {backbone!r}. Supported: convnext_tiny, convnext_small")
+
+    in_features = model.classifier[-1].in_features
+    model.classifier[-1] = nn.Linear(in_features, 1)
+    return model
+
+
+register("convnext_tiny", build)
+register("convnext_small", build)

classifier/src/models/mobilenet.py

@@ -0,0 +1,27 @@
+import torch.nn as nn
+from torchvision import models
+
+from src.models import register
+
+
+# MobileNetV3's classification head is a Sequential; [-1] targets the final Linear
+def build(cfg: dict) -> nn.Module:
+    backbone = cfg.get("backbone", "mobilenet_v3_small")
+    pretrained = cfg.get("pretrained", True)
+
+    if backbone == "mobilenet_v3_small":
+        weights = models.MobileNet_V3_Small_Weights.DEFAULT if pretrained else None
+        model = models.mobilenet_v3_small(weights=weights)
+    elif backbone == "mobilenet_v3_large":
+        weights = models.MobileNet_V3_Large_Weights.DEFAULT if pretrained else None
+        model = models.mobilenet_v3_large(weights=weights)
+    else:
+        raise ValueError(f"Unsupported MobileNet backbone: {backbone!r}. Supported: mobilenet_v3_small, mobilenet_v3_large")
+
+    in_features = model.classifier[-1].in_features
+    model.classifier[-1] = nn.Linear(in_features, 1)
+    return model
+
+
+register("mobilenet_v3_small", build)
+register("mobilenet_v3_large", build)

docs/classifier_impl.md

@@ -328,7 +328,7 @@ Note: Comparison pairs (baseline vs treatment) are defined in the analysis noteb
 ### 3.1 Experiment Configs
 Use the best preprocessing choices from Phase 2. The placeholders below assume 224×224, face crop enabled, and no augmentation unless Phase 2 results justify different settings.
 
-- [ ] Create `classifier/configs/phase3/p3_resnet34.json`
+- [x] Create `classifier/configs/phase3/p3_resnet34.json`
   - backbone: resnet34
   - pretrained: true
   - epochs: 15
@@ -336,14 +336,12 @@ Use the best preprocessing choices from Phase 2. The placeholders below assume 2
   - lr: 1e-4
   - weight_decay: 1e-4
   - image_size: 224
-  - face_crop: true (or best from Phase 2B/E)
-  - face_crop_margin: 0.6
-  - augment: false (or best from Phase 2D/E)
+  - augment: false (placeholder until Phase 2 results confirm)
   - subsample: 0.2
   - seed: 42
   - early_stopping_patience: 5
 
-- [ ] Create `classifier/configs/phase3/p3_resnet50.json`
+- [x] Create `classifier/configs/phase3/p3_resnet50.json`
   - backbone: resnet50
   - pretrained: true
   - epochs: 15
@@ -351,14 +349,12 @@ Use the best preprocessing choices from Phase 2. The placeholders below assume 2
   - lr: 1e-4
   - weight_decay: 1e-4
   - image_size: 224
-  - face_crop: true (or best from Phase 2B/E)
-  - face_crop_margin: 0.6
-  - augment: false (or best from Phase 2D/E)
+  - augment: false (placeholder until Phase 2 results confirm)
   - subsample: 0.2
   - seed: 42
   - early_stopping_patience: 5
 
-- [ ] Create `classifier/configs/phase3/p3_efficientnet_b0.json`
+- [x] Create `classifier/configs/phase3/p3_efficientnet_b0.json`
   - backbone: efficientnet_b0
   - pretrained: true
   - epochs: 15
@@ -366,44 +362,43 @@ Use the best preprocessing choices from Phase 2. The placeholders below assume 2
   - lr: 1e-4
   - weight_decay: 1e-4
   - image_size: 224
-  - face_crop: true (or best from Phase 2B/E)
-  - augment: false (or best from Phase 2D/E)
+  - augment: false (placeholder until Phase 2 results confirm)
   - subsample: 0.2
   - seed: 42
   - early_stopping_patience: 5
 
-- [ ] Create `classifier/configs/phase3/p3_convnext_tiny.json`
+- [x] Create `classifier/configs/phase3/p3_convnext_tiny.json`
   - backbone: convnext_tiny
   - pretrained: true
   - epochs: 15
   - batch_size: 32
-  - lr: 1e-4
+  - lr: 5e-5 (reduced for ConvNeXt stability)
   - weight_decay: 1e-4
   - image_size: 224
-  - face_crop: true (or best from Phase 2B/E)
-  - augment: false (or best from Phase 2D/E)
+  - augment: false (placeholder until Phase 2 results confirm)
   - subsample: 0.2
   - seed: 42
   - early_stopping_patience: 5
 
-- [ ] Create `classifier/configs/phase3/p3_mobilenetv3_small.json`
-  - backbone: mobilenetv3_small
+- [x] Create `classifier/configs/phase3/p3_mobilenetv3_small.json`
+  - backbone: mobilenet_v3_small
   - pretrained: true
   - epochs: 15
   - batch_size: 32
   - lr: 1e-4
   - weight_decay: 1e-4
   - image_size: 224
-  - face_crop: true (or best from Phase 2B/E)
-  - augment: false (or best from Phase 2D/E)
+  - augment: false (placeholder until Phase 2 results confirm)
   - subsample: 0.2
   - seed: 42
   - early_stopping_patience: 5
 
+- [x] Remove `p3a_mobilenet_v3_large.json` (not in plan, MobileNet V3 Large fills no distinct niche)
+
 ### 3.2 Model Implementation
-- [ ] Implement ConvNeXt-Tiny in `classifier/src/models/convnext.py`
-- [ ] Implement MobileNetV3-Small in `classifier/src/models/mobilenet.py`
-- [ ] Register both models in `classifier/src/models/__init__.py`
+- [x] Implement ConvNeXt-Tiny in `classifier/src/models/convnext.py`
+- [x] Implement MobileNetV3-Small in `classifier/src/models/mobilenet.py`
+- [x] Register both models in `classifier/src/models/__init__.py`
 
 ### 3.3 Training
 - [ ] Train ResNet34 with 5-fold stratified group CV
@@ -570,9 +565,9 @@ This section is the consolidated notebook checklist for the notebooks referenced
 - [x] Rename `classifier/run_cv.py` to `classifier/run.py` (pipeline expects classifier/run.py)
 
 ### Model Implementations
-- [ ] Implement ConvNeXt-Tiny in `classifier/src/models/convnext.py`
-- [ ] Implement MobileNetV3-Small in `classifier/src/models/mobilenet.py`
-- [ ] Register both models in `classifier/src/models/__init__.py`
+- [x] Implement ConvNeXt-Tiny in `classifier/src/models/convnext.py`
+- [x] Implement MobileNetV3-Small in `classifier/src/models/mobilenet.py`
+- [x] Register both models in `classifier/src/models/__init__.py`
 
 ### Normalization Implementation
 - [ ] Implement function to calculate mean/std from real training images only
@@ -589,9 +584,9 @@ This section is the consolidated notebook checklist for the notebooks referenced
 - [ ] Implement pairwise source AUC variance calculations
 
 ### Grad-CAM Improvements
-- [ ] Ensure Grad-CAM works for all model types (CNN-based)
-- [ ] Implement Grad-CAM for ConvNeXt
-- [ ] Implement Grad-CAM for MobileNetV3
+- [x] Ensure Grad-CAM works for all model types (CNN-based)
+- [x] Implement Grad-CAM for ConvNeXt (last Conv2d found automatically by `find_conv()`)
+- [x] Implement Grad-CAM for MobileNetV3 (last Conv2d found automatically by `find_conv()`)
 - [ ] Organize Grad-CAM outputs by experiment, model, prediction type, source
 
 ---