weightslab 1.2.3__tar.gz → 1.2.5__tar.gz

weightslab-1.2.5/CHANGELOG.md

@@ -0,0 +1 @@
+# Changelog - 2026-06-17 v1.2.5 (1)

{weightslab-1.2.3/weightslab.egg-info → weightslab-1.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: weightslab
-Version: 1.2.3
+Version: 1.2.5
 Summary: Paving the way between black-box and white-box modeling.
 Author-email: Alexandru-Andrei Rotaru <alexandru@graybx.com>, Guillaue Pelluet <guillaue@graybx.com>
 License: BSD 2-clause

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '1.2.3'
-__version_tuple__ = version_tuple = (1, 2, 3)
+__version__ = version = '1.2.5'
+__version_tuple__ = version_tuple = (1, 2, 5)
 
-__commit_id__ = commit_id = 'g27ecce886'
+__commit_id__ = commit_id = 'g935f5b0b4'

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/components/checkpoint_manager.py

@@ -833,8 +833,8 @@ class CheckpointManager:
         weights_model = None
 
         if 'model' in changed_components and model is not None and dump_model_architecture:
-            logger.info("Dumping model architecture...")
-            self.save_model_architecture(model)
+            # logger.info("Dumping model architecture...")
+            # self.save_model_architecture(model)
             should_save_weights = True
             weights_model = model
 

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/components/global_monitoring.py

@@ -219,9 +219,10 @@ class GuardContext:
         Executed upon entering the 'with' block. Sets the model to training mode.
         """
         self._maybe_pause_at_step()
-        if f:
-            pause_controller.resume(force=f)
-        pause_controller.wait_if_paused()
+        if not is_in_evaluation():
+            if f:
+                pause_controller.resume(force=f)
+            pause_controller.wait_if_paused()
         self.architecture_guard.__enter__()
 
         # Set the current context for this execution

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/integrations/ultralytics/trainer.py

@@ -28,7 +28,6 @@ from torch.nn import Identity
 from ultralytics.models.yolo.detect import DetectionTrainer
 
 import weightslab as wl
-from weightslab.components.global_monitoring import pause_controller
 from weightslab.backend import ledgers
 
 from .collate import wl_ul_dict_collate
@@ -71,23 +70,26 @@ class WLAwareTrainer(DetectionTrainer):
             @wl.eval_fn
             def _validate(loader):
                 # Clean pause ctrl callbacks
+                import copy
                 raised_exc = None
+                val_loader = trainer.validator.dataloader
                 try:
-                    on_val_batch_start_callbacks = trainer.validator.callbacks.pop('on_val_batch_start')  # Remove pause ctrl deps. We can use it as resume pause ctrl will trigger the training in //.
-                    on_val_batch_end_callbacks = trainer.validator.callbacks.pop('on_val_batch_end')  # Remove pause ctrl deps. We can use it as resume pause ctrl will trigger the training in //.
-                    val_loader = trainer.validator.dataloader
                     trainer.validator.dataloader = loader
-                    trainer.validator(model=(trainer.ema.ema if trainer.ema else trainer.model))
+                    # validator(model=...) takes UL's *standalone* path, which
+                    # wraps the model in AutoBackend and FUSES it in place
+                    # (conv+bn -> conv with bias, bn deleted). Fusing the live
+                    # EMA/model rewrites its state_dict keys, so the next
+                    # trainer.ema.update(model) raises KeyError 'model.*.conv.bias'.
+                    # Validate on a throwaway deep copy so the live EMA / training
+                    # model are never mutated. (`underlying` is the unwrapped model
+                    # captured above; avoids deep-copying the WL wrapper.)
+                    src = trainer.ema.ema if trainer.ema else underlying
+                    eval_model = copy.deepcopy(src).eval()
+                    trainer.validator(model=eval_model)
                 except Exception as e:
                     raised_exc = e
-                    pass
-
-                # Set Val loader
-                trainer.validator.dataloader = val_loader  # Reset val trainer
-
-                # Set val callbacks
-                trainer.validator.callbacks["on_val_batch_start"].extend(on_val_batch_start_callbacks)
-                trainer.validator.callbacks["on_val_batch_end"].extend(on_val_batch_end_callbacks)
+                finally:
+                    trainer.validator.dataloader = val_loader  # Reset val loader
 
                 # Finally raise exc.
                 if raised_exc is not None:

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/src.py

@@ -1729,14 +1729,6 @@ def save_signals(
     preds_np     = expand_dim(preds_np)
     preds_raw_np = expand_dim(preds_raw_np)
 
-    # During evaluation mode we must not mutate dataframe state.
-    try:
-        from weightslab.components.evaluation_controller import eval_controller
-        if eval_controller.is_running():
-            return
-    except Exception:
-        pass
-
     # Enqueue to dataframe manager buffer for efficiency
     DATAFRAME_M.enqueue_batch(
         sample_ids=batch_ids_np,
@@ -2666,12 +2658,13 @@ def run_pending_evaluation(
     controlled_loader = _EvalManagedLoader(loader_if, split_name, total_batches, max_batches=max_steps)
     eval_error = None
 
-    # Set evaluation context (exempt from watchdog timeouts)
+    # Set evaluation context (exempt from watchdog timeouts) and guarding
     from weightslab.components.global_monitoring import set_in_evaluation, reset_in_evaluation
     eval_context_token = set_in_evaluation(True)
 
     try:
-        _eval_fn(controlled_loader)
+        with th.no_grad():
+            _eval_fn(controlled_loader)
     except _EvalCanceled as exc:
         logger_obj.warning("[wl.run_pending_evaluation] canceled: %s", exc)
         eval_controller.mark_canceled(str(exc))

{weightslab-1.2.3 → weightslab-1.2.5}/weightslab/trainer/trainer_services.py

@@ -466,7 +466,7 @@ def grpc_serve(
     watchdog_exit_on_stuck = str(os.getenv("GRPC_WATCHDOG_EXIT_ON_STUCK", "0")).strip().lower() in {"1", "true", "yes", "on"}
     watchdog_restart_threshold = int(os.getenv("GRPC_WATCHDOG_RESTART_THRESHOLD", "3"))  # Restart after 3 unhealthy checks
     watchdog_details_limit = int(os.getenv("GRPC_WATCHDOG_INFLIGHT_DETAILS_LIMIT", "10"))
-    watchdog_disabled = str(os.getenv("WEIGHTSLAB_DISABLE_WATCHDOGS", "0")).strip().lower() in {"1", "true", "yes", "on"}
+    watchdog_disabled = str(os.getenv("WEIGHTSLAB_DISABLE_WATCHDOGS", "1")).strip().lower() in {"1", "true", "yes", "on"}  # Default state: disabled
     config = get_hyperparams()
     grpc_tls_enabled = _resolve_bool_setting(config, "grpc_tls_enabled", "GRPC_TLS_ENABLED", "0")
     grpc_tls_key_file = _resolve_grpc_tls_path(

{weightslab-1.2.3 → weightslab-1.2.5/weightslab.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: weightslab
-Version: 1.2.3
+Version: 1.2.5
 Summary: Paving the way between black-box and white-box modeling.
 Author-email: Alexandru-Andrei Rotaru <alexandru@graybx.com>, Guillaue Pelluet <guillaue@graybx.com>
 License: BSD 2-clause

weightslab-1.2.3/CHANGELOG.md

@@ -1 +0,0 @@
-# Changelog - 2026-06-12 v1.2.3 (1)