SURE-tools 2.4.5__py3-none-any.whl → 2.4.13__py3-none-any.whl

SURE/TranscriptomeDecoder.py

@@ -1,28 +1,28 @@
 import torch
 import torch.nn as nn
-import torch.optim as optim
 import torch.nn.functional as F
+import torch.optim as optim
 from torch.utils.data import Dataset, DataLoader
 import numpy as np
-from typing import Dict, List, Tuple, Optional
-import matplotlib.pyplot as plt
-from tqdm import tqdm
+from typing import Dict, Optional
 import warnings
 warnings.filterwarnings('ignore')
 
 class TranscriptomeDecoder:
+    """Transcriptome decoder"""
+    
     def __init__(self, 
                  latent_dim: int = 100,
                  gene_dim: int = 60000,
-                 hidden_dim: int = 512,  # Reduced for memory efficiency
+                 hidden_dim: int = 512,
                  device: str = None):
         """
-        Whole-transcriptome decoder
+        Simple but powerful decoder for latent to transcriptome mapping
         
         Args:
             latent_dim: Latent variable dimension (typically 50-100)
             gene_dim: Number of genes (full transcriptome ~60,000)
-            hidden_dim: Hidden dimension (reduced for memory efficiency)
+            hidden_dim: Hidden dimension optimized
             device: Computation device
         """
         self.latent_dim = latent_dim
@@ -30,10 +30,6 @@ class TranscriptomeDecoder:
         self.hidden_dim = hidden_dim
         self.device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
         
-        # Memory optimization settings
-        self.gradient_checkpointing = True
-        self.mixed_precision = True
-        
         # Initialize model
         self.model = self._build_model()
         self.model.to(self.device)
@@ -43,259 +39,163 @@ class TranscriptomeDecoder:
         self.training_history = None
         self.best_val_loss = float('inf')
         
-        print(f"🚀 TranscriptomeDecoder Initialized:")
+        print(f"🚀 SimpleTranscriptomeDecoder Initialized:")
         print(f"   - Latent Dimension: {latent_dim}")
         print(f"   - Gene Dimension: {gene_dim}")
         print(f"   - Hidden Dimension: {hidden_dim}")
         print(f"   - Device: {self.device}")
         print(f"   - Parameters: {sum(p.numel() for p in self.model.parameters()):,}")
     
-    class MemoryEfficientBlock(nn.Module):
-        """Memory-efficient building block with gradient checkpointing"""
-        def __init__(self, input_dim, output_dim, use_checkpointing=True):
-            super().__init__()
-            self.use_checkpointing = use_checkpointing
-            self.net = nn.Sequential(
-                nn.Linear(input_dim, output_dim),
-                nn.BatchNorm1d(output_dim),
-                nn.GELU(),
-                nn.Dropout(0.1)
-            )
-        
-        def forward(self, x):
-            if self.use_checkpointing and self.training:
-                return torch.utils.checkpoint.checkpoint(self.net, x)
-            return self.net(x)
-    
-    class SparseGeneProjection(nn.Module):
-        """Sparse gene projection to reduce memory usage"""
-        def __init__(self, latent_dim, gene_dim, projection_dim=256):
-            super().__init__()
-            self.projection_dim = projection_dim
-            self.gene_embeddings = nn.Parameter(torch.randn(gene_dim, projection_dim) * 0.02)
-            self.latent_projection = nn.Linear(latent_dim, projection_dim)
-            self.activation = nn.GELU()
-            
-        def forward(self, latent):
-            # Project latent to gene space efficiently
-            batch_size = latent.shape[0]
-            latent_proj = self.latent_projection(latent)  # [batch, projection_dim]
-            
-            # Efficient matrix multiplication
-            gene_embeds = self.gene_embeddings.T  # [projection_dim, gene_dim]
-            output = torch.matmul(latent_proj, gene_embeds)  # [batch, gene_dim]
-            
-            return self.activation(output)
-        
-    class ChunkedTransformer(nn.Module):    
-        def __init__(self, gene_dim, hidden_dim=512, chunk_size=2000, num_layers=3):
-            super().__init__()
-            self.chunk_size = chunk_size
-            self.hidden_dim = hidden_dim
-            self.num_chunks = (gene_dim + chunk_size - 1) // chunk_size
-        
-            # 共享的Transformer层
-            self.transformer_layers = nn.ModuleList([
-                nn.Sequential(
-                    nn.Linear(hidden_dim, hidden_dim),
-                    nn.GELU(),
-                    nn.Dropout(0.1),
-                    nn.Linear(hidden_dim, hidden_dim),
-                ) for _ in range(num_layers)
-            ])
-        
-            # 每个chunk独立的投影层
-            self.input_projections = nn.ModuleList([
-                nn.Linear(min(chunk_size, gene_dim - i * chunk_size), hidden_dim)
-                for i in range(self.num_chunks)
-            ])
-            self.output_projections = nn.ModuleList([
-                nn.Linear(hidden_dim, min(chunk_size, gene_dim - i * chunk_size))
-                for i in range(self.num_chunks)
-            ])
-        
-        def forward(self, x):
-            batch_size, gene_dim = x.shape
-            output = torch.zeros_like(x)
-        
-            for i in range(self.num_chunks):
-                start_idx = i * self.chunk_size
-                end_idx = min((i + 1) * self.chunk_size, gene_dim)
-                current_chunk_size = end_idx - start_idx
-            
-                chunk = x[:, start_idx:end_idx]  # [batch_size, current_chunk_size]
-            
-                # 投影到hidden_dim
-                chunk_proj = self.input_projections[i](chunk)  # [batch_size, hidden_dim]
-            
-                # Transformer处理
-                for layer in self.transformer_layers:
-                    chunk_proj = layer(chunk_proj) + chunk_proj
-            
-                # 投影回原始维度
-                chunk_out = self.output_projections[i](chunk_proj)  # [batch_size, current_chunk_size]
-            
-                output[:, start_idx:end_idx] = chunk_out
-        
-            return output
-    
     class Decoder(nn.Module):
-        """Decoder model"""
-        def __init__(self, latent_dim, gene_dim, hidden_dim):
+        """Memory-efficient decoder architecture with dimension handling"""
+        
+        def __init__(self, latent_dim: int, gene_dim: int, hidden_dim: int):
             super().__init__()
             self.latent_dim = latent_dim
             self.gene_dim = gene_dim
             self.hidden_dim = hidden_dim
             
-            # Stage 1: Latent expansion (memory efficient)
+            # Stage 1: Latent variable expansion
             self.latent_expansion = nn.Sequential(
                 nn.Linear(latent_dim, hidden_dim * 2),
+                nn.BatchNorm1d(hidden_dim * 2),
                 nn.GELU(),
                 nn.Dropout(0.1),
                 nn.Linear(hidden_dim * 2, hidden_dim),
+                nn.BatchNorm1d(hidden_dim),
+                nn.GELU(),
             )
             
-            # Stage 2: Sparse gene projection
-            self.gene_projection = TranscriptomeDecoder.SparseGeneProjection(
-                latent_dim, gene_dim, hidden_dim
-            )
-            
-            # Stage 3: Chunked processing
-            self.chunked_processor = TranscriptomeDecoder.ChunkedTransformer(
-                gene_dim, hidden_dim, chunk_size=2000, num_layers=3
+            # Stage 2: Direct projection to gene dimension (simpler approach)
+            self.gene_projector = nn.Sequential(
+                nn.Linear(hidden_dim, hidden_dim * 2),
+                nn.GELU(),
+                nn.Dropout(0.1),
+                nn.Linear(hidden_dim * 2, gene_dim),  # Direct projection to gene_dim
             )
             
-            # Stage 4: Multi-head output with memory efficiency
-            self.output_heads = nn.ModuleList([
-                nn.Sequential(
-                    nn.Linear(hidden_dim, hidden_dim // 2),
-                    nn.GELU(),
-                    nn.Linear(hidden_dim // 2, 1)
-                ) for _ in range(2)  # Reduced from 3 to 2 heads
-            ])
-            
-            # Adaptive fusion
-            self.fusion_gate = nn.Sequential(
-                nn.Linear(hidden_dim, hidden_dim // 4),
+            # Stage 3: Lightweight gene interaction
+            self.gene_interaction = nn.Sequential(
+                nn.Conv1d(1, 32, kernel_size=3, padding=1),
                 nn.GELU(),
-                nn.Linear(hidden_dim // 4, len(self.output_heads)),
-                nn.Softmax(dim=-1)
+                nn.Dropout1d(0.1),
+                nn.Conv1d(32, 1, kernel_size=3, padding=1),
             )
             
             # Output scaling
             self.output_scale = nn.Parameter(torch.ones(1))
             self.output_bias = nn.Parameter(torch.zeros(1))
             
-            self.latent_to_gene = nn.Linear(hidden_dim, gene_dim)
-            
             self._init_weights()
         
         def _init_weights(self):
+            """Weight initialization"""
             for module in self.modules():
                 if isinstance(module, nn.Linear):
                     nn.init.xavier_uniform_(module.weight)
                     if module.bias is not None:
                         nn.init.zeros_(module.bias)
+                elif isinstance(module, nn.Conv1d):
+                    nn.init.kaiming_uniform_(module.weight)
+                    if module.bias is not None:
+                        nn.init.zeros_(module.bias)
         
-        def forward(self, latent):
+        def forward(self, latent: torch.Tensor) -> torch.Tensor:
             batch_size = latent.shape[0]
             
-            # 1. Latent expansion
-            latent_expanded = self.latent_expansion(latent)
-            
-            # 2. Gene projection (memory efficient)
-            gene_features = self.gene_projection(latent)
+            # 1. Expand latent variables
+            latent_features = self.latent_expansion(latent)  # [batch_size, hidden_dim]
             
-            # 3. Add latent information
-            latent_gene_injection = self.latent_to_gene(latent_expanded)
-            gene_features = gene_features + latent_gene_injection
+            # 2. Direct projection to gene dimension
+            gene_output = self.gene_projector(latent_features)  # [batch_size, gene_dim]
             
-            # 4. Chunked processing (memory efficient)
-            gene_features = self.chunked_processor(gene_features)
+            # 3. Gene interaction with dimension safety
+            if self.gene_dim > 1:  # Only apply if gene_dim > 1
+                gene_output = gene_output.unsqueeze(1)  # [batch_size, 1, gene_dim]
+                interaction_output = self.gene_interaction(gene_output)  # [batch_size, 1, gene_dim]
+                gene_output = gene_output + interaction_output  # Residual connection
+                gene_output = gene_output.squeeze(1)  # [batch_size, gene_dim]
             
-            # 5. Multi-head output with chunking
-            final_output = torch.zeros(batch_size, self.gene_dim, device=latent.device)
+            # 4. Final activation (ensure non-negative)
+            gene_output = F.softplus(gene_output * self.output_scale + self.output_bias)
             
-            # Process output in chunks
-            chunk_size = 5000
-            for i in range(0, self.gene_dim, chunk_size):
-                end_idx = min(i + chunk_size, self.gene_dim)
-                chunk = gene_features[:, i:end_idx]
-                
-                head_outputs = []
-                for head in self.output_heads:
-                    head_out = head(chunk).squeeze(-1)
-                    head_outputs.append(head_out)
-                
-                # Adaptive fusion
-                gate_weights = self.fusion_gate(chunk.mean(dim=1, keepdim=True))
-                gate_weights = gate_weights.unsqueeze(1)
-                
-                # Weighted fusion
-                chunk_output = torch.zeros_like(head_outputs[0])
-                for j, head_out in enumerate(head_outputs):
-                    chunk_output = chunk_output + gate_weights[:, :, j] * head_out
-                
-                final_output[:, i:end_idx] = chunk_output
-            
-            # Final activation
-            final_output = F.softplus(final_output * self.output_scale + self.output_bias)
-            
-            return final_output
+            return gene_output
     
     def _build_model(self):
-        """Build model"""
+        """Build the decoder model"""
         return self.Decoder(self.latent_dim, self.gene_dim, self.hidden_dim)
     
+    def _create_dataset(self, latent_data, expression_data):
+        """Create dataset with dimension validation"""
+        class SimpleDataset(Dataset):
+            def __init__(self, latent, expression):
+                # Ensure dimensions match
+                assert latent.shape[0] == expression.shape[0], "Sample count mismatch"
+                assert latent.shape[1] == self.latent_dim, f"Latent dim mismatch: expected {self.latent_dim}, got {latent.shape[1]}"
+                assert expression.shape[1] == self.gene_dim, f"Gene dim mismatch: expected {self.gene_dim}, got {expression.shape[1]}"
+                
+                self.latent = torch.FloatTensor(latent)
+                self.expression = torch.FloatTensor(expression)
+            
+            def __len__(self):
+                return len(self.latent)
+            
+            def __getitem__(self, idx):
+                return self.latent[idx], self.expression[idx]
+        
+        return SimpleDataset(latent_data, expression_data)
+    
     def train(self,
               train_latent: np.ndarray,
               train_expression: np.ndarray,
               val_latent: np.ndarray = None,
               val_expression: np.ndarray = None,
-              batch_size: int = 16,  # Reduced batch size for memory
+              batch_size: int = 32,
               num_epochs: int = 100,
               learning_rate: float = 1e-4,
               checkpoint_path: str = 'transcriptome_decoder.pth'):
         """
-        Memory-efficient training with optimizations
+        Train the decoder model with dimension safety
         
         Args:
-            train_latent: Training latent variables
-            train_expression: Training expression data
-            val_latent: Validation latent variables
-            val_expression: Validation expression data
-            batch_size: Reduced batch size for memory constraints
+            train_latent: Training latent variables [n_samples, latent_dim]
+            train_expression: Training expression data [n_samples, gene_dim]
+            val_latent: Validation latent variables (optional)
+            val_expression: Validation expression data (optional)
+            batch_size: Batch size optimized for memory
             num_epochs: Number of training epochs
             learning_rate: Learning rate
-            checkpoint_path: Model save path
+            checkpoint_path: Path to save the best model
         """
-        print("🚀 Starting Training...")
-        print(f"📊 Batch size: {batch_size}")
+        print("🚀 Starting training...")
         
-        # Enable memory optimizations
-        torch.backends.cudnn.benchmark = True
-        if self.mixed_precision:
-            scaler = torch.cuda.amp.GradScaler()
+        # Dimension validation
+        self._validate_data_dimensions(train_latent, train_expression, "Training")
+        if val_latent is not None and val_expression is not None:
+            self._validate_data_dimensions(val_latent, val_expression, "Validation")
         
         # Data preparation
-        train_dataset = self._create_dataset(train_latent, train_expression)
+        train_dataset = self._create_safe_dataset(train_latent, train_expression)
         
         if val_latent is not None and val_expression is not None:
-            val_dataset = self._create_dataset(val_latent, val_expression)
+            val_dataset = self._create_safe_dataset(val_latent, val_expression)
+            train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=2)
+            val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=2)
+            print(f"📈 Using provided validation data: {len(val_dataset)} samples")
         else:
             # Auto split
             train_size = int(0.9 * len(train_dataset))
             val_size = len(train_dataset) - train_size
-            train_dataset, val_dataset = torch.utils.data.random_split(
-                train_dataset, [train_size, val_size]
-            )
+            train_subset, val_subset = torch.utils.data.random_split(train_dataset, [train_size, val_size])
+            train_loader = DataLoader(train_subset, batch_size=batch_size, shuffle=True, num_workers=2)
+            val_loader = DataLoader(val_subset, batch_size=batch_size, shuffle=False, num_workers=2)
+            print(f"📈 Auto-split validation: {val_size} samples")
         
-        train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, 
-                                 pin_memory=True, num_workers=2)
-        val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False,
-                               pin_memory=True, num_workers=2)
+        print(f"📊 Training samples: {len(train_loader.dataset)}")
+        print(f"📊 Validation samples: {len(val_loader.dataset)}")
+        print(f"📊 Batch size: {batch_size}")
         
-        # Optimizer with memory-friendly settings
+        # Optimizer configuration
         optimizer = optim.AdamW(
             self.model.parameters(),
             lr=learning_rate,
@@ -306,130 +206,181 @@ class TranscriptomeDecoder:
         # Learning rate scheduler
         scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_epochs)
         
-        # Loss function
-        criterion = nn.MSELoss()
+        # Loss function with dimension safety
+        def safe_loss(pred, target):
+            # Ensure dimensions match
+            if pred.shape != target.shape:
+                print(f"⚠️ Dimension mismatch: pred {pred.shape}, target {target.shape}")
+                # Truncate to minimum dimension (safety measure)
+                min_dim = min(pred.shape[1], target.shape[1])
+                pred = pred[:, :min_dim]
+                target = target[:, :min_dim]
+                
+            def correlation_loss(pred, target):
+                pred_centered = pred - pred.mean(dim=1, keepdim=True)
+                target_centered = target - target.mean(dim=1, keepdim=True)
+        
+                correlation = (pred_centered * target_centered).sum(dim=1) / (
+                    torch.sqrt(torch.sum(pred_centered ** 2, dim=1)) * 
+                    torch.sqrt(torch.sum(target_centered ** 2, dim=1)) + 1e-8
+                )
+        
+                return 1 - correlation.mean()
+            
+            mse_loss = F.mse_loss(pred, target)
+            poisson_loss = (pred - target * torch.log(pred + 1e-8)).mean()
+            corr_loss = correlation_loss(pred, target)
+            return mse_loss + 0.5 * poisson_loss + 0.3 * corr_loss
         
         # Training history
         history = {
-            'train_loss': [], 'val_loss': [],
-            'learning_rate': [], 'memory_usage': []
+            'train_loss': [],
+            'val_loss': [],
+            'learning_rate': []
         }
         
         best_val_loss = float('inf')
+        patience = 15
+        patience_counter = 0
         
+        print("\n📈 Starting training loop...")
         for epoch in range(1, num_epochs + 1):
-            print(f"\n📍 Epoch {epoch}/{num_epochs}")
-            
-            # Training phase with memory monitoring
-            train_loss = self._train_epoch(
-                train_loader, optimizer, criterion, scaler if self.mixed_precision else None
-            )
+            # Training phase
+            train_loss = self._train_epoch(train_loader, optimizer, safe_loss)
             
             # Validation phase
-            val_loss = self._validate_epoch(val_loader, criterion)
+            val_loss = self._validate_epoch(val_loader, safe_loss)
             
             # Update scheduler
             scheduler.step()
+            current_lr = scheduler.get_last_lr()[0]
             
             # Record history
             history['train_loss'].append(train_loss)
             history['val_loss'].append(val_loss)
-            history['learning_rate'].append(optimizer.param_groups[0]['lr'])
-            
-            # Memory usage tracking
-            if torch.cuda.is_available():
-                memory_used = torch.cuda.memory_allocated() / 1024**3  # GB
-                history['memory_usage'].append(memory_used)
-                print(f"💾 GPU Memory: {memory_used:.1f}GB / 20GB")
+            history['learning_rate'].append(current_lr)
             
-            print(f"📊 Train Loss: {train_loss:.4f} | Val Loss: {val_loss:.4f}")
-            print(f"⚡ Learning Rate: {optimizer.param_groups[0]['lr']:.2e}")
+            # Print progress
+            if epoch % 5 == 0 or epoch == 1:
+                print(f"📍 Epoch {epoch:3d}/{num_epochs} | "
+                      f"Train Loss: {train_loss:.4f} | "
+                      f"Val Loss: {val_loss:.4f} | "
+                      f"LR: {current_lr:.2e}")
             
-            # Save best model
+            # Early stopping and model saving
             if val_loss < best_val_loss:
                 best_val_loss = val_loss
+                patience_counter = 0
                 self._save_checkpoint(epoch, optimizer, scheduler, best_val_loss, history, checkpoint_path)
-                print("💾 Best model saved!")
+                if epoch % 10 == 0:
+                    print(f"💾 Best model saved (Val Loss: {best_val_loss:.4f})")
+            else:
+                patience_counter += 1
+                if patience_counter >= patience:
+                    print(f"🛑 Early stopping at epoch {epoch}")
+                    break
         
+        # Training completed
         self.is_trained = True
         self.training_history = history
         self.best_val_loss = best_val_loss
         
-        print(f"\n🎉 Training completed! Best validation loss: {best_val_loss:.4f}")
+        print(f"\n🎉 Training completed!")
+        print(f"🏆 Best validation loss: {best_val_loss:.4f}")
+        print(f"📊 Final training loss: {history['train_loss'][-1]:.4f}")
+        
         return history
     
-    def _train_epoch(self, train_loader, optimizer, criterion, scaler=None):
-        """Training epoch"""
+    def _validate_data_dimensions(self, latent_data, expression_data, data_type):
+        """Validate input data dimensions"""
+        assert latent_data.shape[1] == self.latent_dim, (
+            f"{data_type} latent dimension mismatch: expected {self.latent_dim}, got {latent_data.shape[1]}")
+        assert expression_data.shape[1] == self.gene_dim, (
+            f"{data_type} gene dimension mismatch: expected {self.gene_dim}, got {expression_data.shape[1]}")
+        assert latent_data.shape[0] == expression_data.shape[0], (
+            f"{data_type} sample count mismatch: latent {latent_data.shape[0]}, expression {expression_data.shape[0]}")
+        print(f"✅ {data_type} data dimensions validated")
+    
+    def _create_safe_dataset(self, latent_data, expression_data):
+        """Create dataset with safety checks"""
+        class SafeDataset(Dataset):
+            def __init__(self, latent, expression):
+                self.latent = torch.FloatTensor(latent)
+                self.expression = torch.FloatTensor(expression)
+                
+                # Safety check
+                if self.latent.shape[0] != self.expression.shape[0]:
+                    raise ValueError(f"Sample count mismatch: latent {self.latent.shape[0]}, expression {self.expression.shape[0]}")
+            
+            def __len__(self):
+                return len(self.latent)
+            
+            def __getitem__(self, idx):
+                return self.latent[idx], self.expression[idx]
+        
+        return SafeDataset(latent_data, expression_data)
+    
+    def _train_epoch(self, train_loader, optimizer, loss_fn):
+        """Train for one epoch with dimension safety"""
         self.model.train()
         total_loss = 0
         
-        pbar = tqdm(train_loader, desc='Training')
-        for latent, target in pbar:
-            latent = latent.to(self.device, non_blocking=True)
-            target = target.to(self.device, non_blocking=True)
-            
-            optimizer.zero_grad(set_to_none=True)  # Memory optimization
+        for batch_idx, (latent, target) in enumerate(train_loader):
+            latent = latent.to(self.device)
+            target = target.to(self.device)
             
-            if scaler:  # Mixed precision training
-                with torch.cuda.amp.autocast():
-                    pred = self.model(latent)
-                    loss = criterion(pred, target)
+            # Dimension check
+            if latent.shape[1] != self.latent_dim:
+                print(f"⚠️ Batch {batch_idx}: Latent dim mismatch {latent.shape[1]} != {self.latent_dim}")
+                continue
                 
-                scaler.scale(loss).backward()
-                scaler.step(optimizer)
-                scaler.update()
-            else:
-                pred = self.model(latent)
-                loss = criterion(pred, target)
-                loss.backward()
-                optimizer.step()
+            optimizer.zero_grad()
+            pred = self.model(latent)
             
-            total_loss += loss.item()
-            pbar.set_postfix({'Loss': f'{loss.item():.4f}'})
+            # Final dimension check before loss calculation
+            if pred.shape[1] != target.shape[1]:
+                min_dim = min(pred.shape[1], target.shape[1])
+                pred = pred[:, :min_dim]
+                target = target[:, :min_dim]
+                if batch_idx == 0:  # Only warn once
+                    print(f"⚠️ Truncating to min dimension: {min_dim}")
             
-            # Clear memory
-            del pred, loss
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
+            loss = loss_fn(pred, target)
+            loss.backward()
+            
+            # Gradient clipping for stability
+            torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
+            optimizer.step()
+            
+            total_loss += loss.item()
         
         return total_loss / len(train_loader)
     
-    def _validate_epoch(self, val_loader, criterion):
-        """Validation"""
+    def _validate_epoch(self, val_loader, loss_fn):
+        """Validate for one epoch with dimension safety"""
         self.model.eval()
         total_loss = 0
         
         with torch.no_grad():
-            for latent, target in val_loader:
-                latent = latent.to(self.device, non_blocking=True)
-                target = target.to(self.device, non_blocking=True)
+            for batch_idx, (latent, target) in enumerate(val_loader):
+                latent = latent.to(self.device)
+                target = target.to(self.device)
                 
                 pred = self.model(latent)
-                loss = criterion(pred, target)
-                total_loss += loss.item()
                 
-                # Clear memory
-                del pred, loss
+                # Dimension safety
+                if pred.shape[1] != target.shape[1]:
+                    min_dim = min(pred.shape[1], target.shape[1])
+                    pred = pred[:, :min_dim]
+                    target = target[:, :min_dim]
+                
+                loss = loss_fn(pred, target)
+                total_loss += loss.item()
         
         return total_loss / len(val_loader)
     
-    def _create_dataset(self, latent_data, expression_data):
-        """Create dataset"""
-        class EfficientDataset(Dataset):
-            def __init__(self, latent, expression):
-                self.latent = torch.FloatTensor(latent)
-                self.expression = torch.FloatTensor(expression)
-            
-            def __len__(self):
-                return len(self.latent)
-            
-            def __getitem__(self, idx):
-                return self.latent[idx], self.expression[idx]
-        
-        return EfficientDataset(latent_data, expression_data)
-    
     def _save_checkpoint(self, epoch, optimizer, scheduler, best_loss, history, path):
-        """Save checkpoint"""
+        """Save model checkpoint"""
         torch.save({
             'epoch': epoch,
             'model_state_dict': self.model.state_dict(),
@@ -444,22 +395,26 @@ class TranscriptomeDecoder:
             }
         }, path)
     
-    def predict(self, latent_data: np.ndarray, batch_size: int = 8) -> np.ndarray:
+    def predict(self, latent_data: np.ndarray, batch_size: int = 32) -> np.ndarray:
         """
-        Prediction
+        Predict gene expression from latent variables
         
         Args:
             latent_data: Latent variables [n_samples, latent_dim]
-            batch_size: Prediction batch size for memory control
+            batch_size: Prediction batch size
         
         Returns:
             expression: Predicted expression [n_samples, gene_dim]
         """
         if not self.is_trained:
-            warnings.warn("Model not trained. Predictions may be inaccurate.")
+            warnings.warn("⚠️ Model not trained. Predictions may be inaccurate.")
         
         self.model.eval()
         
+        # Input validation
+        if latent_data.shape[1] != self.latent_dim:
+            raise ValueError(f"Latent dimension mismatch: expected {self.latent_dim}, got {latent_data.shape[1]}")
+        
         if isinstance(latent_data, np.ndarray):
             latent_data = torch.FloatTensor(latent_data)
         
@@ -470,76 +425,83 @@ class TranscriptomeDecoder:
                 batch_latent = latent_data[i:i+batch_size].to(self.device)
                 batch_pred = self.model(batch_latent)
                 predictions.append(batch_pred.cpu())
-                
-                # Clear memory
-                del batch_pred
-                if torch.cuda.is_available():
-                    torch.cuda.empty_cache()
         
         return torch.cat(predictions).numpy()
     
     def load_model(self, model_path: str):
         """Load pre-trained model"""
         checkpoint = torch.load(model_path, map_location=self.device)
+        
+        # Check model configuration
+        if 'model_config' in checkpoint:
+            config = checkpoint['model_config']
+            if (config['latent_dim'] != self.latent_dim or 
+                config['gene_dim'] != self.gene_dim):
+                print("⚠️ Model configuration mismatch. Reinitializing model.")
+                self.model = self._build_model()
+                self.model.to(self.device)
+        
         self.model.load_state_dict(checkpoint['model_state_dict'])
         self.is_trained = True
         self.training_history = checkpoint.get('training_history')
         self.best_val_loss = checkpoint.get('best_val_loss', float('inf'))
-        print(f"✅ Model loaded! Best validation loss: {self.best_val_loss:.4f}")
-    
-    def get_memory_info(self) -> Dict:
-        """Get memory usage information"""
-        if torch.cuda.is_available():
-            memory_allocated = torch.cuda.memory_allocated() / 1024**3
-            memory_reserved = torch.cuda.memory_reserved() / 1024**3
-            return {
-                'allocated_gb': memory_allocated,
-                'reserved_gb': memory_reserved,
-                'available_gb': 20 - memory_allocated,
-                'utilization_percent': (memory_allocated / 20) * 100
-            }
-        return {'available_gb': 'N/A (CPU mode)'}
+        
+        print(f"✅ Model loaded successfully!")
+        print(f"🏆 Best validation loss: {self.best_val_loss:.4f}")
     
+    def get_model_info(self) -> Dict:
+        """Get model information"""
+        return {
+            'is_trained': self.is_trained,
+            'best_val_loss': self.best_val_loss,
+            'parameters': sum(p.numel() for p in self.model.parameters()),
+            'latent_dim': self.latent_dim,
+            'gene_dim': self.gene_dim,
+            'hidden_dim': self.hidden_dim,
+            'device': str(self.device)
+        }
 '''
-# Example usage with memory monitoring
+# Example usage
 def example_usage():
-    """Memory-efficient example"""
+    """Example demonstration with dimension safety"""
     
-    # 1. Initialize memory-efficient decoder
-    decoder = TranscriptomeDecoder(
+    # 1. Initialize decoder
+    decoder = SimpleTranscriptomeDecoder(
         latent_dim=100,
         gene_dim=2000,  # Reduced for example
-        hidden_dim=256   # Reduced for memory
+        hidden_dim=256
     )
     
-    # Check memory info
-    memory_info = decoder.get_memory_info()
-    print(f"📊 Memory Info: {memory_info}")
-    
-    # 2. Generate example data
-    n_samples = 500  # Reduced for memory
+    # 2. Generate example data with correct dimensions
+    n_samples = 1000
     latent_data = np.random.randn(n_samples, 100).astype(np.float32)
-    expression_data = np.random.randn(n_samples, 2000).astype(np.float32)
+    
+    # Create simulated expression data
+    weights = np.random.randn(100, 2000) * 0.1
+    expression_data = np.tanh(latent_data.dot(weights))
     expression_data = np.maximum(expression_data, 0)  # Non-negative
     
-    print(f"📈 Data shapes: Latent {latent_data.shape}, Expression {expression_data.shape}")
+    print(f"📊 Data shapes: Latent {latent_data.shape}, Expression {expression_data.shape}")
     
-    # 3. Train with memory monitoring
+    # 3. Train the model
     history = decoder.train(
         train_latent=latent_data,
         train_expression=expression_data,
-        batch_size=8,  # Small batch for memory
-        num_epochs=20   # Reduced for example
+        batch_size=32,
+        num_epochs=50,
+        learning_rate=1e-4
     )
     
-    # 4. Memory-efficient prediction
-    test_latent = np.random.randn(5, 100).astype(np.float32)
-    predictions = decoder.predict(test_latent, batch_size=2)
+    # 4. Make predictions
+    test_latent = np.random.randn(10, 100).astype(np.float32)
+    predictions = decoder.predict(test_latent)
     print(f"🔮 Prediction shape: {predictions.shape}")
     
-    # 5. Final memory check
-    final_memory = decoder.get_memory_info()
-    print(f"💾 Final memory usage: {final_memory}")
+    # 5. Get model info
+    info = decoder.get_model_info()
+    print(f"\n📋 Model Info:")
+    for key, value in info.items():
+        print(f"   {key}: {value}")
     
     return decoder