roboflow · Borda · Mar 17, 2026 · Mar 17, 2026 · Mar 17, 2026 · Mar 17, 2026
@@ -276,6 +276,21 @@ def training_step(self, batch: Tuple, batch_idx: int) -> torch.Tensor:
             sync_dist=train_log_sync_dist,
             batch_size=batch_size,
         )
+        optimizer = self.optimizers()
+        if isinstance(optimizer, list):
+            optimizer = optimizer[0]
+        # Optimizer may have multiple param groups with different LRs (e.g., backbone/decoder).
+        # Preserve the first group's LR for backward compatibility, but also log the
+        # min/max across all groups so the progress bar reflects the full schedule.
+        group_lrs = [pg["lr"] for pg in optimizer.param_groups if "lr" in pg]
+        if group_lrs:
+            base_lr = group_lrs[0]
+            min_lr = min(group_lrs)
+            max_lr = max(group_lrs)
+            # Keep LR visible in the live progress bar every step.
+            self.log("train/lr", base_lr, prog_bar=True, on_step=True, on_epoch=False)
+            self.log("train/lr_min", min_lr, prog_bar=True, on_step=True, on_epoch=False)
+            self.log("train/lr_max", max_lr, prog_bar=True, on_step=True, on_epoch=False)
         return loss_scaled
 
     def validation_step(self, batch: Tuple, batch_idx: int) -> Dict[str, Any]:

@@ -75,6 +75,7 @@ def _fit_and_read_csv(mc: RFDETRBaseConfig, tc: TrainConfig, criterion=None) ->
 _REQUIRED_DETECTION = frozenset(
     {
         "train/loss",
+        "train/lr",
         "val/loss",
         "val/mAP_50",
         "val/mAP_50_95",

@@ -586,6 +586,10 @@ def _run_step(self, tmp_path, loss_dict=None, weight_dict=None, accumulate_grad_
         fake_criterion.weight_dict = weight_dict or {"loss_ce": 1.0}
         module.log = MagicMock()
         module.log_dict = MagicMock()
+        # Provide a real optimizer so param_groups carries a real "lr" key.
+        real_param = nn.Parameter(torch.randn(4))
+        real_optimizer = torch.optim.SGD([real_param], lr=1e-3)
+        module.optimizers = MagicMock(return_value=real_optimizer)
         trainer = MagicMock()
         trainer.accumulate_grad_batches = accumulate_grad_batches
         module._trainer = trainer
@@ -622,6 +626,18 @@ def test_logs_train_loss_to_prog_bar(self, tmp_path):
         assert len(train_loss_calls) == 1
         assert train_loss_calls[0].kwargs.get("prog_bar") is True
 
+    def test_logs_learning_rate_to_prog_bar(self, tmp_path):
+        """Current learning rate must be logged as train/lr with prog_bar=True for monitoring."""
+        module, samples, targets, _, _ = self._run_step(tmp_path)
+
+        module.training_step((samples, targets), batch_idx=0)
+
+        lr_calls = [c for c in module.log.call_args_list if c[0][0] == "train/lr"]
+        assert len(lr_calls) == 1
+        assert lr_calls[0].kwargs.get("prog_bar") is True
+        assert lr_calls[0].kwargs.get("on_step") is True
+        assert lr_calls[0].kwargs.get("on_epoch") is False
+
     def test_logs_individual_losses_as_dict(self, tmp_path):
         """Each component loss must be logged separately under train/ prefix."""
         loss_dict = {"loss_ce": torch.tensor(0.5), "loss_bbox": torch.tensor(0.3)}