modified: src/config.py

modified:   src/train.py

src/config.py

@@ -24,6 +24,9 @@ EPOCHS = 50
 WEIGHT_DECAY = 1e-3
 ENABLE_EARLY_STOP = False
 EARLY_STOP_PATIENCE = 15
+SPECTRAL_LOSS_WEIGHT = 0.6
+CORR_LOSS_WEIGHT = 0.3
+CORR_LOSS_EPS = 1e-8
 
 # Device
 import torch

src/train.py

@@ -1,5 +1,4 @@
 import torch
-import torch.nn as nn
 import torch.optim as optim
 from config import *
 from dataset import get_dataloaders
@@ -10,6 +9,47 @@ def masked_l1_loss(pred, target, mask):
     denom = mask.sum().clamp(min=1.0)
     return diff.sum() / denom
 
+def masked_spectral_mag_loss(pred, target, mask):
+    # Apply mask in time domain first so missing labels do not pollute spectrum.
+    pred_masked = pred * mask
+    target_masked = target * mask
+
+    pred_fft = torch.fft.rfft(pred_masked, dim=1)
+    target_fft = torch.fft.rfft(target_masked, dim=1)
+    pred_mag = torch.abs(pred_fft)
+    target_mag = torch.abs(target_fft)
+
+    # Weight each sample/channel by its valid-label ratio.
+    valid_ratio = (mask.sum(dim=1) / mask.shape[1]).clamp(min=0.0, max=1.0)
+    freq_weight = valid_ratio.unsqueeze(1).expand_as(pred_mag)
+
+    diff = torch.abs(pred_mag - target_mag) * freq_weight
+    denom = freq_weight.sum().clamp(min=1.0)
+    return diff.sum() / denom
+
+def masked_corr_loss(pred, target, mask, eps=1e-8):
+    # Compute per-sample/per-channel Pearson correlation with masked timesteps.
+    count = mask.sum(dim=1)  # (batch, channels)
+    valid = count > 1.0
+    count_safe = count.clamp(min=1.0)
+
+    pred_mean = (pred * mask).sum(dim=1) / count_safe
+    target_mean = (target * mask).sum(dim=1) / count_safe
+
+    pred_centered = (pred - pred_mean.unsqueeze(1)) * mask
+    target_centered = (target - target_mean.unsqueeze(1)) * mask
+
+    cov = (pred_centered * target_centered).sum(dim=1)
+    pred_var = (pred_centered ** 2).sum(dim=1)
+    target_var = (target_centered ** 2).sum(dim=1)
+
+    corr = cov / (torch.sqrt(pred_var * target_var) + eps)
+    corr = torch.where(valid, corr, torch.zeros_like(corr))
+
+    valid_float = valid.float()
+    denom = valid_float.sum().clamp(min=1.0)
+    return ((1.0 - corr) * valid_float).sum() / denom
+
 def train_model():
     train_loader, val_loader, _, _, _ = get_dataloaders()
     
@@ -27,6 +67,9 @@ def train_model():
     for epoch in range(EPOCHS):
         model.train()
         train_loss = 0.0
+        train_time_loss = 0.0
+        train_spec_loss = 0.0
+        train_corr_loss = 0.0
         
         for inputs, targets, masks in train_loader:
             inputs = inputs.to(DEVICE)
@@ -35,32 +78,57 @@ def train_model():
             
             optimizer.zero_grad()
             outputs = model(inputs)
-            loss = masked_l1_loss(outputs, targets, masks)
+            time_loss = masked_l1_loss(outputs, targets, masks)
+            spec_loss = masked_spectral_mag_loss(outputs, targets, masks)
+            corr_loss = masked_corr_loss(outputs, targets, masks, eps=CORR_LOSS_EPS)
+            loss = time_loss + SPECTRAL_LOSS_WEIGHT * spec_loss + CORR_LOSS_WEIGHT * corr_loss
             
             loss.backward()
             torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
             optimizer.step()
             
             train_loss += loss.item()
+            train_time_loss += time_loss.item()
+            train_spec_loss += spec_loss.item()
+            train_corr_loss += corr_loss.item()
             
         # 验证
         model.eval()
         val_loss = 0.0
+        val_time_loss = 0.0
+        val_spec_loss = 0.0
+        val_corr_loss = 0.0
         with torch.no_grad():
             for inputs, targets, masks in val_loader:
                 inputs = inputs.to(DEVICE)
                 targets = targets.to(DEVICE)
                 masks = masks.to(DEVICE)
                 outputs = model(inputs)
-                loss = masked_l1_loss(outputs, targets, masks)
+                time_loss = masked_l1_loss(outputs, targets, masks)
+                spec_loss = masked_spectral_mag_loss(outputs, targets, masks)
+                corr_loss = masked_corr_loss(outputs, targets, masks, eps=CORR_LOSS_EPS)
+                loss = time_loss + SPECTRAL_LOSS_WEIGHT * spec_loss + CORR_LOSS_WEIGHT * corr_loss
                 val_loss += loss.item()
+                val_time_loss += time_loss.item()
+                val_spec_loss += spec_loss.item()
+                val_corr_loss += corr_loss.item()
                 
         train_loss /= len(train_loader)
+        train_time_loss /= len(train_loader)
+        train_spec_loss /= len(train_loader)
+        train_corr_loss /= len(train_loader)
         val_loss /= len(val_loader)
+        val_time_loss /= len(val_loader)
+        val_spec_loss /= len(val_loader)
+        val_corr_loss /= len(val_loader)
         
         scheduler.step(val_loss)
         
-        print(f"Epoch {epoch+1}/{EPOCHS} | Train Loss: {train_loss:.4f} | Val Loss: {val_loss:.4f}")
+        print(
+            f"Epoch {epoch+1}/{EPOCHS} | "
+            f"Train Loss: {train_loss:.4f} (L1={train_time_loss:.4f}, Spec={train_spec_loss:.4f}, Corr={train_corr_loss:.4f}) | "
+            f"Val Loss: {val_loss:.4f} (L1={val_time_loss:.4f}, Spec={val_spec_loss:.4f}, Corr={val_corr_loss:.4f})"
+        )
         
         if val_loss < best_val_loss:
             best_val_loss = val_loss