fix(trainer): 修复检查点回调参数顺序和权重保存选项

khaosz/config/param_config.py

@@ -52,10 +52,11 @@ class BaseModelIO:
         self.config.load(str(paths["config"]))
         self.tokenizer.load(str(paths["tokenizer"]))
         
+        if self.model is None:
+            self.model = Transformer(self.config)
+        
         if paths["model"].exists():
             state_dict = st.load_file(str(paths["model"]))
-            if self.model is None:
-                self.model = Transformer(self.config)
             self.model.load_state_dict(state_dict)
         
         return self

khaosz/trainer/checkpoint.py

@@ -2,16 +2,15 @@ import os
 import pickle as pkl
 import matplotlib.pyplot as plt
 
-from torch.optim import Optimizer
-from torch.optim.lr_scheduler import LRScheduler
+from torch import Tensor
 from typing import Dict, Optional
 
 
 class Checkpoint:
     def __init__(
         self,
-        optimizer_state: Optimizer,
-        scheduler_state: LRScheduler,
+        optimizer_state: Dict[str, Tensor],
+        scheduler_state: Dict[str, Tensor],
         epoch: int = 0,
         iteration: int = 0,
         metrics: Optional[Dict[str, list]] = None,
@@ -36,7 +35,7 @@ class Checkpoint:
             pkl.dump(train_state, f)
             
         if save_metric_plot and self.metrics:
-            self._plot_metrics()
+            self._plot_metrics(save_dir)
     
     @classmethod
     def load(cls, save_dir: str) -> "Checkpoint":
@@ -56,7 +55,7 @@ class Checkpoint:
             metrics=train_state["metrics"]
         )
     
-    def _plot_metrics(self):
+    def _plot_metrics(self, save_dir: str):
         for metric_name, metric_value in self.metrics.items():
             plt.figure(figsize=(10, 6))
             plt.plot(metric_value, label=metric_name)
@@ -65,5 +64,6 @@ class Checkpoint:
             plt.legend()
             plt.grid(True, alpha=0.3)
             
-            plt.savefig(f'{metric_name}.png', dpi=150, bbox_inches='tight')
+            save_path = os.path.join(save_dir, f"{metric_name}.png")
+            plt.savefig(save_path, dpi=150, bbox_inches='tight')
             plt.close()

khaosz/trainer/train_callback.py

@@ -72,8 +72,8 @@ class SchedulerCallback(TrainCallback):
     """
     Scheduler callback for trainer.
     """
-    def __init__(self, scheduler: LRScheduler):
-        self.scheduler: LRScheduler = scheduler
+    def __init__(self):
+        self.scheduler: LRScheduler = None
     
     def on_train_begin(self, context: 'TrainContext'):
         for group in context.optimizer.param_groups:
@@ -92,9 +92,16 @@ class CheckpointCallback(TrainCallback):
     """ 
     Checkpoint callback for trainer.
     """
-    def __init__(self, interval: int, save_dir: str):
-        self.interval = interval
+    def __init__(
+        self, 
+        save_dir: str,  
+        interval: int,
+        weight_only: bool = False
+    ):
         self.save_dir = save_dir
+        self.interval = interval
+        self.weight_only = weight_only
+        
         self.last_ckpt_iter = 0
     
     @only_on_rank(0)

khaosz/trainer/trainer.py

@@ -28,9 +28,9 @@ class Trainer:
         train_config = self.train_config
         return [
             ProgressBarCallback(train_config.n_epoch),
-            CheckpointCallback(train_config.checkpoint_interval, train_config.checkpoint_dir),
+            CheckpointCallback(train_config.checkpoint_dir, train_config.checkpoint_interval),
             GradientClippingCallback(train_config.max_grad_norm),
-            SchedulerCallback(train_config.scheduler),
+            SchedulerCallback(),
         ]
     
     def _build_context(self, checkpoint: Optional[Checkpoint]) -> TrainContext: