feat(train): 添加warmup学习率调度器

roboimi/demos/vla_scripts/train_vla.py

@@ -9,6 +9,7 @@ from tqdm import tqdm
 from omegaconf import DictConfig, OmegaConf
 from torch.utils.data import DataLoader, random_split
 from torch.optim import AdamW
+from torch.optim.lr_scheduler import LambdaLR
 from pathlib import Path
 
 # Ensure correct import path
@@ -43,6 +44,43 @@ def recursive_to_device(data, device):
     return data
 
 
+def get_lr_schedule_with_warmup(optimizer, warmup_steps, max_steps, scheduler_type='cosine', min_lr=0):
+    """
+    Create a learning rate scheduler with warmup.
+
+    Args:
+        optimizer: PyTorch optimizer
+        warmup_steps: Number of warmup steps
+        max_steps: Total training steps
+        scheduler_type: Type of scheduler after warmup ('cosine' or 'constant')
+        min_lr: Minimum learning rate (for cosine decay)
+
+    Returns:
+        LambdaLR scheduler
+    """
+    import math
+    # Capture initial lr before LambdaLR modifies it
+    base_lr = optimizer.param_groups[0]['lr']
+    min_lr_ratio = min_lr / base_lr if base_lr > 0 else 0.0
+
+    def lr_lambda(step):
+        # Warmup phase: linear increase from 0 to 1
+        if step < warmup_steps:
+            return float(step) / float(max(1, warmup_steps))
+
+        # Post-warmup phase
+        if scheduler_type == 'cosine':
+            # Cosine annealing from 1 to min_lr_ratio
+            progress = float(step - warmup_steps) / float(max(1, max_steps - warmup_steps))
+            cosine_decay = 0.5 * (1.0 + math.cos(math.pi * progress))
+            return max(min_lr_ratio, cosine_decay)
+        else:
+            # Constant learning rate
+            return 1.0
+
+    return LambdaLR(optimizer, lr_lambda)
+
+
 @hydra.main(version_base=None, config_path="../../vla/conf", config_name="config")
 def main(cfg: DictConfig):
     """
@@ -173,11 +211,25 @@ def main(cfg: DictConfig):
         log.warning("⚠️  Training will continue, but inference may not work correctly")
 
     # =========================================================================
-    # 3. Setup Optimizer
+    # 3. Setup Optimizer & LR Scheduler
     # =========================================================================
     optimizer = AdamW(agent.parameters(), lr=cfg.train.lr, weight_decay=1e-5)
     log.info(f"🔧 Optimizer: AdamW (lr={cfg.train.lr})")
 
+    # Setup learning rate scheduler with warmup
+    warmup_steps = int(cfg.train.get('warmup_steps', 500))
+    scheduler_type = cfg.train.get('scheduler_type', 'cosine')
+    min_lr = float(cfg.train.get('min_lr', 1e-6))
+
+    scheduler = get_lr_schedule_with_warmup(
+        optimizer,
+        warmup_steps=warmup_steps,
+        max_steps=cfg.train.max_steps,
+        scheduler_type=scheduler_type,
+        min_lr=min_lr
+    )
+    log.info(f"📈 LR Scheduler: {scheduler_type} with {warmup_steps} warmup steps (min_lr={min_lr})")
+
     # =========================================================================
     # 4. Training Loop
     # =========================================================================
@@ -265,16 +317,19 @@ def main(cfg: DictConfig):
         torch.nn.utils.clip_grad_norm_(agent.parameters(), max_norm=1.0)
 
         optimizer.step()
+        scheduler.step()
 
         # =====================================================================
         # Logging
         # =====================================================================
         if step % cfg.train.log_freq == 0:
+            current_lr = optimizer.param_groups[0]['lr']
             pbar.set_postfix({
                 "loss": f"{loss.item():.4f}",
+                "lr": f"{current_lr:.2e}",
                 "best_loss": f"{best_loss:.4f}"
             })
-            log.info(f"Step {step}/{cfg.train.max_steps} | Loss: {loss.item():.4f}")
+            log.info(f"Step {step}/{cfg.train.max_steps} | Loss: {loss.item():.4f} | LR: {current_lr:.2e}")
 
         # =====================================================================
         # Checkpoint saving & Validation
@@ -290,9 +345,11 @@ def main(cfg: DictConfig):
                 'step': step,
                 'model_state_dict': agent.state_dict(),
                 'optimizer_state_dict': optimizer.state_dict(),
+                'scheduler_state_dict': scheduler.state_dict(),
                 'loss': loss.item(),
                 'val_loss': val_loss,
                 'dataset_stats': dataset_stats,
+                'current_lr': optimizer.param_groups[0]['lr'],
             }, checkpoint_path)
             log.info(f"💾 Checkpoint saved: {checkpoint_path}")
 
@@ -305,9 +362,11 @@ def main(cfg: DictConfig):
                     'step': step,
                     'model_state_dict': agent.state_dict(),
                     'optimizer_state_dict': optimizer.state_dict(),
+                    'scheduler_state_dict': scheduler.state_dict(),
                     'loss': loss.item(),
                     'val_loss': val_loss,
                     'dataset_stats': dataset_stats,
+                    'current_lr': optimizer.param_groups[0]['lr'],
                 }, best_model_path)
                 log.info(f"🌟 Best model updated: {best_model_path} (val_loss: {best_loss:.4f})")
 
@@ -319,8 +378,10 @@ def main(cfg: DictConfig):
         'step': cfg.train.max_steps,
         'model_state_dict': agent.state_dict(),
         'optimizer_state_dict': optimizer.state_dict(),
+        'scheduler_state_dict': scheduler.state_dict(),
         'loss': loss.item(),
         'dataset_stats': dataset_stats,
+        'current_lr': optimizer.param_groups[0]['lr'],
     }, final_model_path)
     log.info(f"💾 Final model saved: {final_model_path}")
 

roboimi/vla/conf/config.yaml

@@ -11,4 +11,9 @@ train:
   log_freq: 100       # Log frequency (steps)
   save_freq: 2000     # Save checkpoint frequency (steps)
   device: "cuda"      # Device: "cuda" or "cpu"
-  num_workers: 8      # DataLoader workers (set to 0 for debugging, 8 for production)
+  num_workers: 8      # DataLoader workers (set to 0 for debugging, 8 for production)
+
+  # Learning rate scheduler with warmup
+  warmup_steps: 500   # Number of warmup steps
+  scheduler_type: "cosine"  # Scheduler after warmup: "constant" or "cosine"
+  min_lr: 1e-6        # Minimum learning rate (for cosine decay)