nshtrainer 1.0.0b20__tar.gz → 1.0.0b22__tar.gz

{nshtrainer-1.0.0b20 → nshtrainer-1.0.0b22}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: nshtrainer
-Version: 1.0.0b20
+Version: 1.0.0b22
 Summary: 
 Author: Nima Shoghi
 Author-email: nimashoghi@gmail.com

{nshtrainer-1.0.0b20 → nshtrainer-1.0.0b22}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "nshtrainer"
-version = "1.0.0-beta20"
+version = "1.0.0-beta22"
 description = ""
 authors = ["Nima Shoghi <nimashoghi@gmail.com>"]
 readme = "README.md"

{nshtrainer-1.0.0b20 → nshtrainer-1.0.0b22}/src/nshtrainer/trainer/_config.py

@@ -602,6 +602,10 @@ class TrainerConfig(C.Config):
     :meth:`~lightning.pytorch.core.LightningModule.log_dict`.
     """
 
+    experimental_barebones_profiler_enabled: bool = False
+    """If enabled, will enable the profiler in barebones mode.
+    The implementation of this is very hacky and may not work as expected."""
+
     precision: (
         Literal[
             "64-true",

{nshtrainer-1.0.0b20 → nshtrainer-1.0.0b22}/src/nshtrainer/trainer/trainer.py

@@ -249,8 +249,6 @@ class Trainer(LightningTrainer):
                 if logger_config is not None
                 and (logger := logger_config.create_logger(hparams)) is not None
             ]
-            if not hparams.barebones
-            else None
         )
 
         if hparams.auto_determine_num_nodes:
@@ -280,6 +278,12 @@ class Trainer(LightningTrainer):
         _update_kwargs(**hparams.lightning_kwargs)
         _update_kwargs(**kwargs_ctor)
 
+        # Handle barebones mode
+        if kwargs.get("barebones"):
+            # Remove the logger if it's an empty list
+            if (logger := kwargs.get("logger")) is not None and not logger:
+                kwargs["logger"] = None
+
         return kwargs
 
     if TYPE_CHECKING:
@@ -312,8 +316,17 @@ class Trainer(LightningTrainer):
         self._hparams = hparams
         self.debug = self.hparams.debug
 
+        experimental_profiler = None
+        if hparams.experimental_barebones_profiler_enabled:
+            experimental_profiler = kwargs.pop("profiler", None)
         super().__init__(**kwargs)
 
+        # Set up the profiler again
+        if experimental_profiler is not None:
+            from lightning.pytorch.trainer import setup
+
+            setup._init_profiler(self, experimental_profiler)
+
         # Add our own start time callback to measure the start time.
         self.callbacks.append(RuntimeTrackerCallback())