Upload 172 files

milk10k_effb2_metadata/.cache/huggingface/download/.gitattributes.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 bed0738c7eeb449bca98b5d2f33c89a1ee56349a
-1781874436.5302918
+1781597425.7859669

milk10k_effb2_metadata/.cache/huggingface/download/datasets.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 745bbba7691dd8a1caf1f1450fe9cd267979fb72
-1781874436.9742603
+1781597425.9924324

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/MILK10K_EFFB2_METADATA_CLI.md.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-5ce67edf48fa89591db61470df8123f51bd41b3c
-1781874440.1852803
+c48b211a983203f99f058ce6a5ae649db2561a43
+9c467a783f5db02fc2938b239aafbb79ec441940
+1781597425.6549404

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__init__.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 4b177b7643716beb9197ff1a6f13ddc082bcf679
-1781874440.2582364
+1781597425.585372

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/__init__.cpython-310.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 0896243fe721016e63775ed3d25d17e47ecb9d29
-1781874440.2790968
+1781597425.813415

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/__init__.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-f749dc60d0e290f7836d8c2c8f07bf1a4b3f8092
-1781874440.3494253
+c48b211a983203f99f058ce6a5ae649db2561a43
+08bf4b46a5615b9ea4d9a1fafc10abaf6fbcdf77
+1781597426.1462557

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/checkpoints.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-887412cccb05670924444266c84da488c3cb550f
-1781874440.3751688
+c48b211a983203f99f058ce6a5ae649db2561a43
+e48728b832ad1f61f864094af5ae7197e1beded2
+1781597425.511972

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/cli.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-f3ae0db61747159a686668094eb2b5655957eff9
-1781874440.4572845
+c48b211a983203f99f058ce6a5ae649db2561a43
+152dc30657add84c181fa4f9c068d86349fba1a6
+1781597426.539304

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/data.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-050a1b828cbf226f94edcb1c25a85f16d8efbfd3
-1781874440.495286
+c48b211a983203f99f058ce6a5ae649db2561a43
+34f4712b6ab226459b0dc4bce8824d994070021a
+1781597426.559293

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/engine.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-d12567f62ac7a71d398da2f6d7b3aa4b45ff26ee
-1781874440.6045096
+c48b211a983203f99f058ce6a5ae649db2561a43
+94d085df27494f40c7b4cfbc6a4d426985cfcba9
+1781597426.480496

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/inference.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-1a50445c233a49e4ef322096c77cbeef9e69da0d
-1781874440.663752
+c48b211a983203f99f058ce6a5ae649db2561a43
+70b1a231d6c0bcffa94631deb7b8cbb85d9bf224
+1781597426.6514366

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/losses.cpython-310.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 dc43b0d3f3a6c7e77371155a67a730881f96634c
-1781874440.7430017
+1781597426.7428267

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/losses.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-bf1619a2288cc0ee7b5d2b249d53d909ed360af1
-1781874440.824899
+c48b211a983203f99f058ce6a5ae649db2561a43
+e67f1957628509f5390aa9041e8f4b14ca19e6ed
+1781597426.6966894

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/metrics.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-b5170ee786353b800dde0d44d791970638f4636a
-1781874440.7838733
+c48b211a983203f99f058ce6a5ae649db2561a43
+57e036ee2530157cec8597a2194d0f699e2d98d5
+1781597427.0974405

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/model_setup.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-71aa62c2e0b8759071ecd4f34514840262e45116
-1781874440.86317
+c48b211a983203f99f058ce6a5ae649db2561a43
+5a90a95188fd6ccadf70579cd1ad768f675f5258
+1781597427.0447693

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/models.cpython-310.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 d38b14cad04e5a63a1fd0ac34908ce69f8bcd187
-1781874440.908546
+1781597427.375944

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/models.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-410d47b19afe6aef34e5d0406a0825c99db998fa
-1781874440.980258
+c48b211a983203f99f058ce6a5ae649db2561a43
+85b1b6b5686a45d2a47d5df2e20b712b10aeaa55
+1781597427.6779335

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/runner.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-5a933741bffa31ff27fe68c1ecdcb5c41aa33d3d
-1781874441.1509938
+c48b211a983203f99f058ce6a5ae649db2561a43
+9efb15b1ccf8afef6145cc53221067fe920b75b4
+1781597427.760574

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/train_milk10k_effb2_dual_metadata.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-a40d0d3068febb913fda09370729e9b28c82eaf6
-1781874441.1871524
+c48b211a983203f99f058ce6a5ae649db2561a43
+9a7cb0f115b95ac83a1bda2d29654b6b44b80f8c
+1781597427.551546

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/training.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-29eac2bda4df380fbc6bb09637a3e2c809c792d7
-1781874441.2221453
+c48b211a983203f99f058ce6a5ae649db2561a43
+4f979947db3b11e272935ae1768a693b54a56630
+1781597427.717837

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/__pycache__/training_utils.cpython-314.pyc.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-0e993de2817fd59125f75837d139510f66594925
-1781874441.2654667
+c48b211a983203f99f058ce6a5ae649db2561a43
+9b6237f6b9bedbe8465352e8617f62719de01c27
+1781597427.864535

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/checkpoints.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 ca041bd29958b95380013b4c1ab77a377241a5b6
-1781874441.3055234
+1781597428.0874379

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/cli.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-8a5798cb0338c2ceb3060276d12d5c0157599933
-1781874441.3974206
+c48b211a983203f99f058ce6a5ae649db2561a43
+f2afe680f9be4c32d0b764cec33e954b9405c36a
+1781597428.2136478

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/data.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-eec459f9a2ebb5248657e8717c82b8bcd736cf04
-1781874441.4294744
+c48b211a983203f99f058ce6a5ae649db2561a43
+18280e8278d5c3309d25807787869b5399200cfa
+1781597428.513635

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/engine.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-729093ac2a6ba64edcad71867cc8192da3bf5622
-1781874441.6402843
+c48b211a983203f99f058ce6a5ae649db2561a43
+a99ff221829611a452ec800c5ffec8ef45a46721
+1781597428.426173

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/inference.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-e5b2253f1ac4f797a4860934dded4ec70a10f096
-1781874441.594284
+c48b211a983203f99f058ce6a5ae649db2561a43
+5161a1a84c4904356be64f2d8a41b4d450a05860
+1781597428.82443

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/losses.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-a9f264a556310e3492c808a55e3091fda8b37a01
-1781874441.681009
+c48b211a983203f99f058ce6a5ae649db2561a43
+8b6d627c534d71d99d855b29c65bc294b87c6303
+1781597428.79422

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/metrics.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-350b8c2c12a305b4d908d8bd904a8fc484f83bfd
-1781874441.9017274
+c48b211a983203f99f058ce6a5ae649db2561a43
+e9c3f3d079aa08915f21fa758bf05891d861c1bd
+1781597428.8403232

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/model_setup.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-6d556c9f707647fa2f747bb44640b2d63f94e750
-1781874443.7257292
+c48b211a983203f99f058ce6a5ae649db2561a43
+adaed11bb482bc9312d8a9d2b5bcd555936bea9e
+1781597428.9916527

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/models.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-6ee30370cbba7b01fca972db8e8bebb302cc1087
-1781874443.4865744
+c48b211a983203f99f058ce6a5ae649db2561a43
+f6d126347f16f6904eae70a9055005db53a23c81
+1781597428.9753091

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/runner.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-9d4afea6ec9edfb4e65c5c54135ab748d4e6f866
-1781874443.812626
+c48b211a983203f99f058ce6a5ae649db2561a43
+4b39d6e160bc3dab5624ef6da0e572de33b72980
+1781597429.1600146

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/train_milk10k_effb2_dual_metadata.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 da1072ea523e7c2a806284dfafacbb6dba6a1293
-1781874443.932739
+1781597429.3428674

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/training.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-c04d11f5a74450e1e61c75fc9d97985a0606c7a3
-1781874443.9688227
+c48b211a983203f99f058ce6a5ae649db2561a43
+72e7f7b777bddfa3c6e4ed06cf40cf230d8696ee
+1781597429.706166

milk10k_effb2_metadata/.cache/huggingface/download/milk10k_effb2_metadata/training_utils.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
-096cf0ef6f52826e0233290e2d03532b28d99d3b
-1781874443.9975698
+c48b211a983203f99f058ce6a5ae649db2561a43
+96df26e95c0a775a4d5d85ced5c23db69a0e3687
+1781597429.9209738

milk10k_effb2_metadata/.cache/huggingface/download/predict_milk10k_effb2_dual_metadata.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 15f35146cc23e29e3348f85ea82d19bcfbc4f1f0
-1781874444.1152453
+1781597430.34601

milk10k_effb2_metadata/.cache/huggingface/download/train_milk10k_effb2_dual_metadata.py.metadata

@@ -1,3 +1,3 @@
-f50543d6bf12766c60a351d928232a984930a1db
+c48b211a983203f99f058ce6a5ae649db2561a43
 da1072ea523e7c2a806284dfafacbb6dba6a1293
-1781874444.2369857
+1781597429.9241464

milk10k_effb2_metadata/cli.py

@@ -243,4 +243,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--calibration-step", type=float, default=0.25)
     parser.add_argument("--calibration-passes", type=int, default=3)
     parser.add_argument("--patience", type=int, default=6)
+    parser.add_argument("--tau", type=float, default=0.0, help="Logit adjustment parameter for Generalized Balanced Softmax")
+    parser.add_argument("--lws-epochs", type=int, default=0, help="Number of epochs to train Learnable Weight Scaling (LWS) post-training")
+    parser.add_argument("--ema", action="store_true", help="Enable Exponential Moving Average (EMA) for model weights")
+    parser.add_argument("--ema-decay", type=float, default=0.999, help="Decay rate for EMA")
     return parser.parse_args()

milk10k_effb2_metadata/engine.py

@@ -35,6 +35,7 @@ def run_epoch(
     use_amp: bool = False,
     tail_class_indices: list[int] | None = None,
     class_names: list[str] | None = None,
+    ema_model: nn.Module | None = None,
 ) -> dict[str, float]:
     training = optimizer is not None
     model.train(training)
@@ -65,6 +66,8 @@ def run_epoch(
                     loss.backward()
                     torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
                     optimizer.step()
+                if ema_model is not None:
+                    ema_model.update_parameters(model)
 
         batch_size = labels.size(0)
         total_loss += float(loss.detach().item()) * batch_size
@@ -161,6 +164,7 @@ def save_checkpoint(
     metadata_spec: dict[str, Any],
     args: argparse.Namespace,
     extra: dict[str, Any] | None = None,
+    ema_model: nn.Module | None = None,
 ) -> None:
     payload = {
         "epoch": epoch,
@@ -176,6 +180,8 @@ def save_checkpoint(
         "metadata_spec": metadata_spec,
         "args": json_safe(vars(args)),
     }
+    if ema_model is not None:
+        payload["ema_model_state"] = ema_model.state_dict()
     if extra:
         payload.update(json_safe(extra))
     torch.save(payload, path)
@@ -202,6 +208,7 @@ def train_phase(
     tail_class_names: list[str] | None = None,
     train_class_counts: dict[str, int] | None = None,
     best_val_tail_recall: float = float("-inf"),
+    ema_model: nn.Module | None = None,
 ) -> tuple[int, float, float]:
     if num_epochs <= 0:
         return start_epoch, best_val_f1, best_val_tail_recall
@@ -237,9 +244,11 @@ def train_phase(
             use_amp,
             tail_class_indices,
             class_names,
+            ema_model=ema_model,
         )
+        eval_model = ema_model if ema_model is not None else model
         val_stats = run_epoch(
-            model,
+            eval_model,
             val_loader,
             criterion,
             device,
@@ -288,6 +297,7 @@ def train_phase(
                 label_to_idx,
                 metadata_spec,
                 args,
+                ema_model=ema_model,
             )
             print(
                 f"Saved best checkpoint: phase={phase} epoch={epoch:03d} "
@@ -316,6 +326,7 @@ def train_phase(
                     "train_class_counts": train_class_counts or {},
                     "selection_metric": "val_tail_recall_macro",
                 },
+                ema_model=ema_model,
             )
             print(
                 f"Saved tail checkpoint: phase={phase} epoch={epoch:03d} "
@@ -337,6 +348,7 @@ def train_phase(
                 "last_selection_metric": float(val_stats[selection_metric]),
                 "last_val_stats": val_stats,
             },
+            ema_model=ema_model,
         )
         print(
             f"Saved last checkpoint: phase={phase} epoch={epoch:03d} "

milk10k_effb2_metadata/losses.py

@@ -27,6 +27,28 @@ class FocalLoss(nn.Module):
         return loss.mean()
 
 
+class GeneralizedBalancedSoftmaxLoss(nn.Module):
+    def __init__(
+        self,
+        class_counts: torch.Tensor,
+        tau: float = 1.0,
+        weight: torch.Tensor | None = None,
+    ) -> None:
+        super().__init__()
+        self.tau = tau
+        self.weight = weight
+        counts = class_counts.float().clamp_min(1.0)
+        self.register_buffer("log_counts", torch.log(counts))
+
+    def forward(self, logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+        if self.training and self.tau > 0.0:
+            log_counts = self.log_counts.to(device=logits.device, dtype=logits.dtype)
+            adjusted_logits = logits + self.tau * log_counts
+        else:
+            adjusted_logits = logits
+        return F.cross_entropy(adjusted_logits, labels, weight=self.weight)
+
+
 class LDAMLoss(nn.Module):
     """LDAM with deferred effective-number reweighting."""
 
@@ -179,7 +201,12 @@ def build_loss(train_df: pd.DataFrame, label_to_idx: dict[str, int], args: argpa
         y = np.array([label_to_idx[label] for label in train_df["label"]])
         weights = compute_class_weight(class_weight="balanced", classes=np.arange(len(label_to_idx)), y=y)
         weight = torch.tensor(weights, dtype=torch.float32, device=device)
-    ce_loss: nn.Module = nn.CrossEntropyLoss(weight=weight)
+        
+    if getattr(args, "tau", 0.0) > 0.0:
+        counts = class_count_tensor(train_df, label_to_idx, device)
+        ce_loss: nn.Module = GeneralizedBalancedSoftmaxLoss(counts, tau=args.tau, weight=weight)
+    else:
+        ce_loss: nn.Module = nn.CrossEntropyLoss(weight=weight)
     if args.loss == "focal":
         return FocalLoss(weight=weight, gamma=args.focal_gamma)
     if args.loss == "ce_dice":

milk10k_effb2_metadata/milk10k_effb2_metadata/MILK10K_EFFB2_METADATA_CLI.md

@@ -20,16 +20,11 @@ Training code is split by responsibility:
 ```text
 training.py         Thin entry facade: normalize args, load dataframe, choose single run vs k-fold.
 runner.py           Full split runner: split CSVs, loaders, loss, train phases, final metrics/files.
-engine.py           Epoch/phase loop: run_epoch, train_phase, save best/last checkpoints.
+engine.py           Epoch/phase loop: run_epoch, train_phase, save best checkpoint.
 model_setup.py      Backend detection, model construction, resume checkpoint, optimizer param groups.
 training_utils.py   JSON-safe serialization, run_config.json, kfold_summary.csv/json.
 ```
 
-Training always writes both checkpoint types to the run directory:
-
-- `best.pt`: epoch with the best configured validation selection metric.
-- `last.pt`: most recently completed epoch, updated after every epoch and also available when early stopping triggers.
-
 Common places to edit:
 
 ```text
@@ -145,22 +140,6 @@ Metadata fusion can be combined with every image fusion mode:
 
 Normal online image transforms keep the original metadata vector. If you materialize offline transform augmentations, duplicate the original row metadata unchanged. Do not invent metadata for generated synthetic lesions in this trainer; use real-row metadata or disable metadata for synthetic-only experiments.
 
-For synthetic paired augmentations, prefer appending a small train-only subset instead of replacing the whole training CSV:
-
-```bash
-python train_milk10k_effb2_dual_metadata.py \
-  --data-dir /marimo/milk10k \
-  --augmented-data-dir /marimo/milk10k_augmented \
-  --augmented-classes INF BEN_OTH DF VASC \
-  --augmented-max-per-class 25 \
-  --zero-augmented-metadata \
-  --loss ce_f1 \
-  --f1-weight 0.3 \
-  --f1-ignore-classes MAL_OTH
-```
-
-`--zero-augmented-metadata` only zeros metadata vectors for appended augmented rows; real rows still use metadata normally.
-
 ## 3. Class Weight Only
 
 ```bash
@@ -184,28 +163,6 @@ python train_milk10k_effb2_dual_metadata.py \
   --output-dir milk10k_effb2_sampler_p05
 ```
 
-## Hybrid epoch balancing
-
-Use hybrid balancing when the largest class dominates training but inverse-frequency
-loss weights make predictions too soft:
-
-```bash
-python train_milk10k_effb2_dual_metadata.py \
-  --data-dir /path/to/milk10k \
-  --balance-mode hybrid \
-  --balance-head-ratio 2.0 \
-  --balance-tail-floor 100 \
-  --balance-min-source-count 20 \
-  --loss ce \
-  --output-dir milk10k_effb2_hybrid_balance
-```
-
-The largest class is sampled without replacement up to twice the second-largest
-class. Eligible classes below 100 rows are sampled with replacement to 100 rows
-and receive the stronger train transform. Classes with fewer than 20 source rows
-are left unchanged. Sampling changes by epoch and is reproducible from `--seed`.
-Validation is never resampled. Do not combine this mode with `--weighted-sampler`.
-
 Stronger sampling:
 
 ```bash