Upload 23 files

milk10k_effb2_metadata/MILK10K_EFFB2_METADATA_CLI.md

@@ -86,7 +86,7 @@ python train_milk10k_effb2_dual_metadata.py \
   --output-dir milk10k_effb2_focal_sampler_p05
 ```
 
-## 6. MILK Long-Tail Loss
+## 6. LDAM + DRW Loss
 
 Recommended first run:
 
@@ -94,10 +94,10 @@ Recommended first run:
 python train_milk10k_effb2_dual_metadata.py \
   --clinical-checkpoint best_effnetb2_ufes_clinical.pth \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
-  --loss milk_lt \
+  --loss ldam \
   --weighted-sampler \
   --sampler-power 0.5 \
-  --output-dir milk10k_effb2_milk_lt_sampler_p05
+  --output-dir milk10k_effb2_ldam_sampler_p05
 ```
 
 Without sampler:
@@ -106,25 +106,25 @@ Without sampler:
 python train_milk10k_effb2_dual_metadata.py \
   --clinical-checkpoint best_effnetb2_ufes_clinical.pth \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
-  --loss milk_lt \
-  --output-dir milk10k_effb2_milk_lt
+  --loss ldam \
+  --output-dir milk10k_effb2_ldam
 ```
 
-More conservative prior correction:
+More conservative LDAM margin with delayed DRW:
 
 ```bash
 python train_milk10k_effb2_dual_metadata.py \
   --clinical-checkpoint best_effnetb2_ufes_clinical.pth \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
-  --loss milk_lt \
-  --lt-logit-tau 0.5 \
-  --lt-max-margin 0.3 \
+  --loss ldam \
+  --ldam-max-margin 0.3 \
+  --ldam-drw-start-epoch 8 \
   --weighted-sampler \
   --sampler-power 0.5 \
-  --output-dir milk10k_effb2_milk_lt_conservative
+  --output-dir milk10k_effb2_ldam_conservative
 ```
 
-Note: do not add `--class-weight` with `--loss milk_lt`; `milk_lt` already uses effective-number alpha.
+Note: do not add `--class-weight` with `--loss ldam`; LDAM+DRW already uses effective-number alpha.
 
 ## 7. K-Fold
 
@@ -134,17 +134,17 @@ Note: do not add `--class-weight` with `--loss milk_lt`; `milk_lt` already uses
 python train_milk10k_effb2_dual_metadata.py \
   --clinical-checkpoint best_effnetb2_ufes_clinical.pth \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
-  --loss milk_lt \
+  --loss ldam \
   --weighted-sampler \
   --sampler-power 0.5 \
   --k-folds 5 \
-  --output-dir milk10k_effb2_milk_lt_kfold5
+  --output-dir milk10k_effb2_ldam_kfold5
 ```
 
 Outputs:
 
 ```text
-milk10k_effb2_milk_lt_kfold5/
+milk10k_effb2_ldam_kfold5/
   fold_00/
   fold_01/
   fold_02/
@@ -174,11 +174,11 @@ Example with AMP:
 python train_milk10k_effb2_dual_metadata.py \
   --clinical-checkpoint best_effnetb2_ufes_clinical.pth \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
-  --loss milk_lt \
+  --loss ldam \
   --weighted-sampler \
   --sampler-power 0.5 \
   --amp \
-  --output-dir milk10k_effb2_milk_lt_amp
+  --output-dir milk10k_effb2_ldam_amp
 ```
 
 ## 9. Smoke Checks
@@ -197,7 +197,7 @@ python train_milk10k_effb2_dual_metadata.py \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
   --freeze-epochs 0 \
   --finetune-epochs 0 \
-  --loss milk_lt \
+  --loss ldam \
   --output-dir /tmp/milk10k_effb2_smoke_single
 ```
 
@@ -209,7 +209,7 @@ python train_milk10k_effb2_dual_metadata.py \
   --dermoscopic-checkpoint efficientnet_b2_best_dermoscopic.pt \
   --freeze-epochs 0 \
   --finetune-epochs 0 \
-  --loss milk_lt \
+  --loss ldam \
   --k-folds 2 \
   --output-dir /tmp/milk10k_effb2_smoke_kfold
 ```
@@ -243,7 +243,7 @@ Use the saved checkpoint directly. You do not need to pass the original branch c
 
 ```bash
 python predict_milk10k_effb2_dual_metadata.py \
-  --checkpoint milk10k_effb2_milk_lt_sampler_p05/best.pt \
+  --checkpoint milk10k_effb2_ldam_sampler_p05/best.pt \
   --data-dir /marimo/milk10k \
   --output milk10k_effb2_test_predictions.csv \
   --batch-size 16 \
@@ -261,7 +261,7 @@ For an unlabeled test set, pass image root and metadata CSV explicitly:
 
 ```bash
 python predict_milk10k_effb2_dual_metadata.py \
-  --checkpoint milk10k_effb2_milk_lt_sampler_p05/best.pt \
+  --checkpoint milk10k_effb2_ldam_sampler_p05/best.pt \
   --input-dir /path/to/MILK10k_Test_Input \
   --metadata-csv /path/to/MILK10k_Test_Metadata.csv \
   --output milk10k_effb2_test_predictions.csv \

milk10k_effb2_metadata/cli.py

@@ -76,13 +76,12 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--class-weight", action="store_true")
     parser.add_argument("--weighted-sampler", action="store_true")
     parser.add_argument("--sampler-power", type=float, default=1.0)
-    parser.add_argument("--loss", choices=["ce", "focal", "milk_lt"], default="ce")
+    parser.add_argument("--loss", choices=["ce", "focal", "ldam"], default="ce")
     parser.add_argument("--focal-gamma", type=float, default=2.0)
-    parser.add_argument("--lt-beta", type=float, default=0.9999)
-    parser.add_argument("--lt-max-margin", type=float, default=0.5)
-    parser.add_argument("--lt-logit-tau", type=float, default=1.0)
-    parser.add_argument("--lt-draw-start-epoch", "--lt-drw-start-epoch", dest="lt_draw_start_epoch", type=int, default=0)
-    parser.add_argument("--lt-alpha-max", type=float, default=10.0)
+    parser.add_argument("--ldam-beta", type=float, default=0.9999)
+    parser.add_argument("--ldam-max-margin", type=float, default=0.5)
+    parser.add_argument("--ldam-drw-start-epoch", type=int, default=0)
+    parser.add_argument("--ldam-alpha-max", type=float, default=10.0)
     parser.add_argument("--k-folds", type=int, default=1)
     parser.add_argument("--amp", action="store_true")
     parser.add_argument(

milk10k_effb2_metadata/losses.py

@@ -27,15 +27,14 @@ class FocalLoss(nn.Module):
         return loss.mean()
 
 
-class MILKLongTailLoss(nn.Module):
-    """LDAM + balanced-softmax/logit adjustment + deferred effective-number alpha."""
+class LDAMLoss(nn.Module):
+    """LDAM with deferred effective-number reweighting."""
 
     def __init__(
         self,
         class_counts: torch.Tensor,
         beta: float = 0.9999,
         max_margin: float = 0.5,
-        logit_tau: float = 1.0,
         deferred_start_epoch: int = 0,
         alpha_max: float = 10.0,
     ) -> None:
@@ -43,15 +42,12 @@ class MILKLongTailLoss(nn.Module):
         counts = class_counts.float().clamp_min(1.0)
         margins = 1.0 / torch.sqrt(torch.sqrt(counts))
         margins = margins * (max_margin / margins.max().clamp_min(1e-12))
-        priors = counts / counts.sum()
         alpha = effective_number_alpha(counts, beta)
         alpha = alpha.clamp(max=alpha_max)
         alpha = alpha * (counts.numel() / alpha.sum().clamp_min(1e-12))
 
         self.register_buffer("margins", margins)
-        self.register_buffer("log_priors", priors.log())
         self.register_buffer("alpha", alpha)
-        self.logit_tau = logit_tau
         self.deferred_start_epoch = deferred_start_epoch
         self.current_epoch = 0
 
@@ -60,12 +56,10 @@ class MILKLongTailLoss(nn.Module):
 
     def forward(self, logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
         margins = self.margins.to(device=logits.device, dtype=logits.dtype)
-        log_priors = self.log_priors.to(device=logits.device, dtype=logits.dtype)
         alpha = self.alpha.to(device=logits.device, dtype=logits.dtype)
         adjusted_logits = logits.clone()
         rows = torch.arange(labels.size(0), device=labels.device)
         adjusted_logits[rows, labels] = adjusted_logits[rows, labels] - margins[labels]
-        adjusted_logits = adjusted_logits + self.logit_tau * log_priors
         loss = F.cross_entropy(adjusted_logits, labels, reduction="none")
         if self.current_epoch >= self.deferred_start_epoch:
             loss = loss * alpha[labels]
@@ -76,7 +70,7 @@ def effective_number_alpha(counts: torch.Tensor, beta: float) -> torch.Tensor:
     if beta <= 0.0:
         return torch.ones_like(counts)
     if beta >= 1.0:
-        raise ValueError("--lt-beta must be less than 1.0")
+        raise ValueError("--ldam-beta must be less than 1.0")
     beta_tensor = torch.tensor(beta, dtype=counts.dtype, device=counts.device)
     effective_num = 1.0 - torch.pow(beta_tensor, counts)
     alpha = (1.0 - beta_tensor) / effective_num.clamp_min(1e-12)
@@ -93,15 +87,14 @@ def class_count_tensor(train_df: pd.DataFrame, label_to_idx: dict[str, int], dev
 
 
 def build_loss(train_df: pd.DataFrame, label_to_idx: dict[str, int], args: argparse.Namespace, device: torch.device) -> nn.Module:
-    if args.loss == "milk_lt":
+    if args.loss == "ldam":
         counts = class_count_tensor(train_df, label_to_idx, device)
-        return MILKLongTailLoss(
+        return LDAMLoss(
             class_counts=counts,
-            beta=args.lt_beta,
-            max_margin=args.lt_max_margin,
-            logit_tau=args.lt_logit_tau,
-            deferred_start_epoch=args.lt_draw_start_epoch,
-            alpha_max=args.lt_alpha_max,
+            beta=args.ldam_beta,
+            max_margin=args.ldam_max_margin,
+            deferred_start_epoch=args.ldam_drw_start_epoch,
+            alpha_max=args.ldam_alpha_max,
         )
 
     weight = None

milk10k_effb2_metadata/training.py

@@ -416,8 +416,8 @@ def run_training_split(
         f"freeze_metadata_head={args.freeze_metadata_head}, metadata_lr={args.metadata_lr}"
     )
     print(f"Loss: {args.loss}, class_weight={args.class_weight}, weighted_sampler={args.weighted_sampler}")
-    if args.loss == "milk_lt" and args.class_weight:
-        print("Note: --class-weight is ignored for --loss milk_lt because milk_lt uses effective-number alpha.")
+    if args.loss == "ldam" and args.class_weight:
+        print("Note: --class-weight is ignored for --loss ldam because LDAM+DRW uses effective-number alpha.")
 
     history: list[dict[str, Any]] = []
     history_path = output_dir / "history.csv"