claude_refactor_v3: Add and passed test on splited gin-configs loading but without weigts

src/training/trainer_new.py

@@ -230,7 +230,7 @@ class Trainer:
     def train(self) -> None:
         """Full pipeline: setup → build → train → evaluate → cleanup."""
         self._validate_backbone()
-        clear_vram()
+        #! clear_vram()
         set_seed(self.pipeline_cfg.seed)
         self._setup_output_dir()
         self._setup_tracker()
@@ -256,6 +256,7 @@ class Trainer:
 
     def _validate_backbone(self) -> None:
         """Reject unsupported backbones up front with a helpful message."""
+        LOGGER.info("⚙️  Validate backbone")
         backbone = self.models_common_cfg.backbone
         if backbone not in _SUPPORTED_BACKBONES:
             raise NotImplementedError(
@@ -271,6 +272,7 @@ class Trainer:
 
     def _setup_output_dir(self) -> None:
         """Create output_dir, save config.json, init csv_logger."""
+        LOGGER.info("⚙️  Setup out dir")
         self.output_dir = Path(self.pipeline_cfg.output_dir)
         self.output_dir.mkdir(parents=True, exist_ok=True)
 
@@ -290,6 +292,7 @@ class Trainer:
 
     def _setup_tracker(self) -> None:
         """W&B + TensorBoard tracker."""
+        LOGGER.info("⚙️ Setup tracker...")
         assert self.output_dir is not None and self.full_config is not None
         self.tracker = ExperimentTracker(
             output_dir=self.output_dir,
@@ -303,6 +306,7 @@ class Trainer:
 
     def _build_model(self) -> None:
         """Build (or load) the encoder model based on the active backbone."""
+        LOGGER.info("⚙️ Build loss...")
         backbone = self.models_common_cfg.backbone
 
         if self.pipeline_cfg.resume_from is not None:
@@ -327,6 +331,7 @@ class Trainer:
 
     def _build_model_from_resume(self, backbone: str) -> None:
         """Resume model from checkpoint. Sets self.model, self.resume_ckpt, self.start_epoch."""
+        LOGGER.info("⚙️ Build model from resume...")
         LOGGER.info("Resuming from %s", self.pipeline_cfg.resume_from)
         # Both DINOv3 and StripNet go through AsymmetricEncoder.load_checkpoint.
         # Note: load_checkpoint doesn't support StripNet — known existing limitation.
@@ -348,6 +353,7 @@ class Trainer:
 
     def _build_stripnet_model(self) -> nn.Module:
         """Construct AsymmetricEncoder configured for StripNet."""
+        LOGGER.info("⚙️ Build StripNet model...")
         assert isinstance(self.models_cfg, StripNetModelsConfig)
         m = self.models_cfg
         # DINO paths passed but ignored at runtime when backbone='stripnet'.
@@ -370,6 +376,7 @@ class Trainer:
         ).to(self.hardware_cfg.device)
 
     def _build_dinov3_model(self) -> nn.Module:
+        LOGGER.info("⚙️  Build DINOv3 model...")
         """Construct AsymmetricEncoder configured for DINOv3."""
         assert isinstance(self.models_cfg, DINOv3ModelsConfig)
         m = self.models_cfg
@@ -393,6 +400,7 @@ class Trainer:
 
     def _configure_gradient_checkpointing(self) -> None:
         """Enable gradient checkpointing on encoders that support it."""
+        LOGGER.info("⚙️  Configure gradient checkpointing...")
         assert self.model is not None
         backbone = self.models_common_cfg.backbone
         if not self.hardware_cfg.gradient_checkpointing:
@@ -406,11 +414,11 @@ class Trainer:
                 self.model.sat_encoder.set_gradient_checkpointing(True)
             if self.model.text_encoder is not None:
                 self.model.text_encoder.transformer.gradient_checkpointing = True
-            LOGGER.info("Gradient checkpointing enabled (DINOv3 + DGTRS)")
+            LOGGER.info("✅ Gradient checkpointing enabled (DINOv3 + DGTRS)")
         elif backbone == "stripnet":
             if self.model.text_encoder is not None:
                 self.model.text_encoder.transformer.gradient_checkpointing = True
-            LOGGER.info("Gradient checkpointing enabled (DGTRS only; StripNet doesn't support it)")
+            LOGGER.info("✅ Gradient checkpointing enabled (DGTRS only; StripNet doesn't support it)")
 
     def _log_model_summary(self) -> None:
         """Log trainable param count, save model_summary.txt, hook W&B."""
@@ -428,6 +436,7 @@ class Trainer:
 
     def _build_loss(self) -> None:
         """Build InfoNCELoss or WeightedInfoNCELoss based on training_cfg.loss_type."""
+        LOGGER.info("⚙️  Build loss...")
         t = self.training_cfg
         if t.loss_type == "symmetric":
             self.loss_fn = InfoNCELoss(
@@ -465,6 +474,7 @@ class Trainer:
 
     def _build_neg_bank(self) -> None:
         """Optional NegativeMemoryBank for hard-negative mining."""
+        LOGGER.info("⚙️  Build negative bank...")
         assert self.model is not None
         if self.training_cfg.neg_bank_size > 0:
             self.neg_bank = NegativeMemoryBank(
@@ -478,6 +488,7 @@ class Trainer:
 
     def _build_data_loaders(self) -> None:
         """Build train/test/train_eval datasets, samplers, loaders."""
+        LOGGER.info("⚙️  Build dataloaders...")
         drone_train_tf = get_drone_train_transform(image_size=256)
         sat_train_tf = get_satellite_train_transform(image_size=256)
         eval_tf = get_dino_transform(image_size=256)
@@ -594,6 +605,7 @@ class Trainer:
 
     def _build_optimizer_and_scheduler(self) -> None:
         """Build AdamW with per-group LR + cosine-warmup scheduler + GradScaler."""
+        LOGGER.info("⚙️  Build optimizer & scheduler...")
         assert self.model is not None and self.loss_fn is not None and self.train_loader is not None
         t = self.training_cfg
 
@@ -637,6 +649,7 @@ class Trainer:
 
     def _restore_from_resume(self) -> None:
         """Restore optimizer/scheduler/loss state on resume."""
+        LOGGER.info("⚙️  Restore from resume...")
         if self.resume_ckpt is None:
             return
         assert self.optimizer is not None and self.loss_fn is not None and self.scheduler is not None
@@ -653,6 +666,7 @@ class Trainer:
 
     def _setup_profiler(self) -> None:
         """Optional PyTorch profiler (only if start_epoch == 0)."""
+        LOGGER.info("⚙️  Setup profiler...")
         if self.tracking_cfg.use_profiler and self.start_epoch == 0:
             assert self.output_dir is not None
             self.profiler = TrainingProfiler(