README.md

@@ -29,7 +29,7 @@ from aiia.model import AIIAConfig
 from aiia.pretrain import Pretrainer
 
 # Create your model
-config = AIIAConfig(model_name="AIIA-Base-512x20k")
+config = AIIAConfig(model_type="AIIA-Base-512x20k")
 model = AIIABase(config)
 
 # Initialize pretrainer with the model

example.py

@@ -1,24 +1,22 @@
-from aiia.model import AIIABase
+from src.aiia.model import AIIAmoe
-from aiia.model import AIIAConfig
+from src.aiia.model import AIIAConfig
-from aiia.pretrain import Pretrainer
+from src.aiia.pretrain import Pretrainer
 
 # Create your model
-config = AIIAConfig(model_name="AIIA-Base-512x20k")
+config = AIIAConfig(num_experts=5)
-model = AIIABase(config)
+model = AIIAmoe(config)
+model.save_pretrained("test")
+model = AIIAmoe.from_pretrained("test")
 
-# Initialize pretrainer with the model
 pretrainer = Pretrainer(model, learning_rate=1e-4, config=config)
 
-# List of dataset paths
+# Set checkpoint directory
-dataset_paths = [
+checkpoint_dir = "checkpoints/my_model"
-    "/path/to/dataset1.parquet",
-    "/path/to/dataset2.parquet"
-]
 
-# Start training with multiple datasets
+# Start training (will automatically load checkpoint if available)
 pretrainer.train(
-    dataset_paths=dataset_paths,
+    dataset_paths=["path/to/dataset1.parquet", "path/to/dataset2.parquet"],
-    num_epochs=10,
+    output_path="trained_models/my_model",
-    batch_size=2,
+    checkpoint_dir=checkpoint_dir,
-    sample_size=10000
+    num_epochs=10
 )

pyproject.toml

@@ -10,7 +10,7 @@ include = '\.pyi?$'
 
 [project]
 name = "aiia"
-version = "0.2.1"
+version = "0.3.1"
 description = "AIIA Deep Learning Model Implementation"
 readme = "README.md"
 authors = [

requirements.txt

@@ -6,3 +6,4 @@ pillow
 pandas
 torchvision
 pyarrow
+transformers>=4.48.0

setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = aiia
-version = 0.2.1
+version = 0.3.1
 author = Falko Habel
 author_email = falko.habel@gmx.de
 description = AIIA deep learning model implementation

src/aiia/__init__.py

@@ -1,7 +1,7 @@
-from .model.Model import AIIABase, AIIABaseShared, AIIAchunked, AIIAmoe, AIIASparseMoe, AIIArecursive
+from .model.Model import AIIABase, AIIABaseShared, AIIAmoe, AIIASparseMoe
 from .model.config import AIIAConfig
 from .data.DataLoader import DataLoader
 from .pretrain.pretrainer import Pretrainer, ProjectionHead
 
 
-__version__ = "0.2.1"
+__version__ = "0.3.1"

src/aiia/model/Model.py

@@ -1,119 +1,48 @@
 from .config import AIIAConfig
 from torch import nn
+from transformers import PreTrainedModel
 import torch
-import os
 import copy
-import warnings
 
 
-class AIIA(nn.Module):
+class AIIABase(PreTrainedModel):
-    def __init__(self, config: AIIAConfig, **kwargs):
+    config_class = AIIAConfig
-        super(AIIA, self).__init__()
+    base_model_prefix = "AIIA"
-        # Create a deep copy of the configuration to avoid sharing
-        self.config = copy.deepcopy(config)
     
-        # Update the config with any additional keyword arguments
+    def __init__(self, config: AIIAConfig):
-        for key, value in kwargs.items():
+        super().__init__(config)
-            setattr(self.config, key, value)
-
-    def save(self, path: str):
-        if not os.path.exists(path):
-            os.makedirs(path, exist_ok=True)
-        torch.save(self.state_dict(), f"{path}/model.pth")
-        self.config.save(path)
-
-    @classmethod
-    def load(cls, path, precision: str = None, strict: bool = True, **kwargs):
-        config = AIIAConfig.load(path)
-        device = 'cuda' if torch.cuda.is_available() else 'cpu'
-        
-        # Load the state dict to analyze structure
-        model_dict = torch.load(f"{path}/model.pth", map_location=device)
-        
-        # Special handling for AIIAmoe - detect number of experts from state_dict
-        if cls.__name__ == "AIIAmoe" and "num_experts" not in kwargs:
-            # Find maximum expert index
-            max_expert_idx = -1
-            for key in model_dict.keys():
-                if key.startswith("experts."):
-                    parts = key.split(".")
-                    if len(parts) > 1:
-                        try:
-                            expert_idx = int(parts[1])
-                            max_expert_idx = max(max_expert_idx, expert_idx)
-                        except ValueError:
-                            pass
-            
-            if max_expert_idx >= 0:
-                # experts.X keys found, use max_expert_idx + 1 as num_experts
-                kwargs["num_experts"] = max_expert_idx + 1
-        
-        # Create model with detected structural parameters
-        model = cls(config, **kwargs)
-        
-        # Handle precision conversion
-        dtype = None
-        if precision is not None:
-            if precision.lower() == 'fp16':
-                dtype = torch.float16
-            elif precision.lower() == 'bf16':
-                if device == 'cuda' and not torch.cuda.is_bf16_supported():
-                    warnings.warn("BF16 is not supported on this GPU. Falling back to FP16.")
-                    dtype = torch.float16
-                else:
-                    dtype = torch.bfloat16
-            else:
-                raise ValueError("Unsupported precision. Use 'fp16', 'bf16', or leave as None.")
-        
-        if dtype is not None:
-            for key, param in model_dict.items():
-                if torch.is_tensor(param):
-                    model_dict[key] = param.to(dtype)
-        
-        # Load state dict with strict parameter for flexibility
-        model.load_state_dict(model_dict, strict=strict)
-        return model
-
-
-class AIIABase(AIIA):
-    def __init__(self, config: AIIAConfig, **kwargs):
-        super().__init__(config=config, **kwargs)  
-        self.config = self.config
         
         # Initialize layers based on configuration
         layers = []
-        in_channels = self.config.num_channels
+        in_channels = config.num_channels
         
-        for _ in range(self.config.num_hidden_layers):
+        for _ in range(config.num_hidden_layers):
             layers.extend([
-                nn.Conv2d(in_channels, self.config.hidden_size,
+                nn.Conv2d(in_channels, config.hidden_size,
-                          kernel_size=self.config.kernel_size, padding=1),
+                          kernel_size=config.kernel_size, padding=1),
-                getattr(nn, self.config.activation_function)(),
+                getattr(nn, config.activation_function)(),
                 nn.MaxPool2d(kernel_size=1, stride=1)
              ])
-            in_channels = self.config.hidden_size
+            in_channels = config.hidden_size
         
         self.cnn = nn.Sequential(*layers)
 
     def forward(self, x):
         return self.cnn(x)
     
-class AIIABaseShared(AIIA):
+class AIIABaseShared(PreTrainedModel):
-    def __init__(self, config: AIIAConfig, **kwargs):
+    config_class = AIIAConfig
+    base_model_prefix = "AIIA"
+    
+    def __init__(self, config: AIIAConfig):
+        super().__init__(config)
         """
         Initialize the AIIABaseShared model.
 
         Args:
             config (AIIAConfig): Configuration object containing model parameters.
-            **kwargs: Additional keyword arguments to override configuration settings.
         """
-        super().__init__(config=config, **kwargs)
+        super().__init__(config=config)
-        
-        # Update configuration with new parameters if provided
-        self. config = copy.deepcopy(config)
-        
-        for key, value in kwargs.items():
-            setattr(self.config, key, value)
                     
         # Initialize the network components
         self._initialize_network()
@@ -172,16 +101,17 @@ class AIIABaseShared(AIIA):
             
         return out
 
-class AIIAExpert(AIIA):
+class AIIAExpert(PreTrainedModel):
-    def __init__(self, config: AIIAConfig, base_class=AIIABase, **kwargs):
+    config_class = AIIAConfig
-        super().__init__(config=config, **kwargs)
+    base_model_prefix = "AIIA"
-        self.config = self.config
+    def __init__(self, config: AIIAConfig, base_class=AIIABase):
+        super().__init__(config=config)
         
         # Initialize base CNN with configuration and chosen base class
         if issubclass(base_class, AIIABase):
-            self.base_cnn = AIIABase(self.config, **kwargs)
+            self.base_cnn = AIIABase(self.config)
         elif issubclass(base_class, AIIABaseShared):
-            self.base_cnn = AIIABaseShared(self.config, **kwargs)
+            self.base_cnn = AIIABaseShared(self.config)
         else:
             raise ValueError("Invalid base class")
     
@@ -198,26 +128,26 @@ class AIIAExpert(AIIA):
         # Process input through the base CNN
         return self.base_cnn(x)
 
-class AIIAmoe(AIIA):
+class AIIAmoe(PreTrainedModel):
-    def __init__(self, config: AIIAConfig, num_experts: int = 3, base_class=AIIABase, **kwargs):
+    config_class = AIIAConfig
-        super().__init__(config=config, **kwargs)
+    base_model_prefix = "AIIA"
+    
+    def __init__(self, config: AIIAConfig, base_class=AIIABase):
+        super().__init__(config=config)
         self.config = config
 
-        # Update the config to include the number of experts.
+        # Get num_experts directly from config instead of parameter
-        self.config.num_experts = num_experts
+        num_experts = getattr(config, "num_experts", 3)  # default to 3 if not in config
         
-        # Initialize multiple experts from the chosen base class.
+        # Initialize multiple experts from the chosen base class
         self.experts = nn.ModuleList([
-            AIIAExpert(self.config, base_class=base_class, **kwargs)
+            AIIAExpert(self.config, base_class=base_class)
             for _ in range(num_experts)
         ])
 
-        # To generate gating weights, we first need to determine the feature dimension.
+        gate_in_features = self.config.hidden_size
-        # Each expert is assumed to return an output of shape (B, C, H, W); after averaging over H and W, 
-        # we obtain a tensor of shape (B, C) where C is the number of channels (here assumed to be 224).
-        gate_in_features = 512  # Adjust this if your expert output changes.
 
-        # Create a gating network that maps the aggregated features to num_experts weights.
+        # Create a gating network that maps the aggregated features to num_experts weights
         self.gate = nn.Sequential(
             nn.Linear(gate_in_features, num_experts),
             nn.Softmax(dim=1)
@@ -261,9 +191,10 @@ class AIIAmoe(AIIA):
 
 
 class AIIASparseMoe(AIIAmoe):
-    def __init__(self, config: AIIAConfig, num_experts: int = 3, top_k: int = 2, base_class=AIIABase, **kwargs):
+    config_class = AIIAConfig
-        super().__init__(config=config, num_experts=num_experts, base_class=base_class, **kwargs)
+    base_model_prefix = "AIIA"
-        self.top_k = top_k
+    def __init__(self, config: AIIAConfig, base_class=AIIABase):
+        super().__init__(config=config, base_class=base_class)
     
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         # Compute the gate_weights similar to standard moe.
@@ -273,7 +204,7 @@ class AIIASparseMoe(AIIAmoe):
         gate_weights = self.gate(gate_input)
 
         # Select the top-k experts for each input based on gating weights.
-        _, top_k_indices = gate_weights.topk(self.top_k, dim=-1)
+        _, top_k_indices = gate_weights.topk(self.config.top_k, dim=-1)
         
         # Initialize a list to store outputs from selected experts.
         merged_outputs = []
@@ -294,64 +225,7 @@ class AIIASparseMoe(AIIAmoe):
         return torch.cat(merged_outputs, dim=0)
 
 
-class AIIAchunked(AIIA):
-    def __init__(self, config: AIIAConfig, patch_size: int = 16, base_class=AIIABase, **kwargs):
-        super().__init__(config=config, **kwargs)  
-        self.config = self.config
-        
-        # Update config with new parameters if provided
-        self.config.patch_size = patch_size
-        
-        # Initialize base CNN for processing each patch using the specified base class
-        if issubclass(base_class, AIIABase):
-            self.base_cnn = AIIABase(self.config, **kwargs)
-        elif issubclass(base_class, AIIABaseShared):  # Add support for AIIABaseShared
-            self.base_cnn = AIIABaseShared(self.config, **kwargs)
-        else:
-            raise ValueError("Invalid base class")
-
-    def forward(self, x):
-        patches = x.unfold(2, self.patch_size, self.patch_size).unfold(3, self.patch_size, self.patch_size)
-        patches = patches.contiguous().view(patches.size(0), patches.size(1), -1, self.patch_size, self.patch_size)
-        patch_outputs = []
-        
-        for p in torch.split(patches, 1, dim=2):
-            p = p.squeeze(2)
-            po = self.base_cnn(p)
-            patch_outputs.append(po)
-            
-        combined_output = torch.mean(torch.stack(patch_outputs, dim=0), dim=0)
-        return combined_output
-
-class AIIArecursive(AIIA):
-    def __init__(self, config: AIIAConfig, recursion_depth: int = 3, base_class=AIIABase, **kwargs):
-        
-        super().__init__(config=config, **kwargs)
-        self.config = self.config
-        
-        # Pass recursion_depth as a kwarg to the config
-        self.config.recursion_depth = recursion_depth
-            
-        # Initialize chunked CNN with updated config
-        self.chunked_cnn = AIIAchunked(self.config, base_class, **kwargs)
-        
-    def forward(self, x, depth=0):
-        if depth == self.recursion_depth:
-            return self.chunked_cnn(x)
-        else:
-            patches = x.unfold(2, 16, 16).unfold(3, 16, 16)
-            patches = patches.contiguous().view(patches.size(0), patches.size(1), -1, 16, 16)
-            processed_patches = []
-            
-            for p in torch.split(patches, 1, dim=2):
-                p = p.squeeze(2)
-                pp = self.forward(p, depth + 1)
-                processed_patches.append(pp)
-                
-            combined_output = torch.mean(torch.stack(processed_patches, dim=0), dim=0)
-            return combined_output
-
 if __name__ =="__main__":            
     config = AIIAConfig()
     model = AIIAmoe(config, num_experts=5)
-    model.save("test")
+    model.save_pretrained("test")

src/aiia/model/__init__.py

@@ -1,20 +1,15 @@
 from .Model import (
     AIIABase,
     AIIABaseShared,
-    AIIAchunked,
     AIIAmoe,
     AIIASparseMoe,
-    AIIArecursive
 )
 from .config import AIIAConfig
 
 __all__ = [
     "AIIABase",
     "AIIABaseShared",
-    "AIIAchunked",
     "AIIAmoe",
     "AIIASparseMoe",
-    "AIIArecursive",
     "AIIAConfig",
-
 ]

src/aiia/model/config.py

@@ -1,28 +1,24 @@
-import torch
+from transformers import PretrainedConfig
 import torch.nn as nn
-import json
-import os 
 
+class AIIAConfig(PretrainedConfig):
+    model_type = "AIIA"  # Add this class attribute
 
-class AIIAConfig:
     def __init__(
         self,
-        model_name: str = "AIIA",
         kernel_size: int = 3,
         activation_function: str = "GELU",
         hidden_size: int = 512,
         num_hidden_layers: int = 12,
         num_channels: int = 3,
-        learning_rate: float = 5e-5,
         **kwargs
     ):
-        self.model_name = model_name
+        super().__init__(**kwargs)
         self.kernel_size = kernel_size
         self.activation_function = activation_function
         self.hidden_size = hidden_size
         self.num_hidden_layers = num_hidden_layers
         self.num_channels = num_channels
-        self.learning_rate = learning_rate
 
         # Store additional keyword arguments as attributes
         for key, value in kwargs.items():
@@ -51,16 +47,3 @@ class AIIAConfig:
                 return {k: serialize(v) for k, v in value.items()}
             return value
         return {k: serialize(v) for k, v in self.__dict__.items()}
-
-    def save(self, file_path):
-        if not os.path.exists(file_path):
-            os.makedirs(file_path, exist_ok=True)
-        with open(os.path.join(file_path, "config.json"), "w") as f:
-            # Save the recursively converted dictionary.
-            json.dump(self.to_dict(), f, indent=4)
-
-    @classmethod
-    def load(cls, file_path):
-        with open(os.path.join(file_path, "config.json"), "r") as f:
-            config_dict = json.load(f)
-        return cls(**config_dict)

src/aiia/pretrain/pretrainer.py

@@ -1,9 +1,11 @@
 import torch
 from torch import nn
 import csv
+import datetime
+import time
 import pandas as pd
 from tqdm import tqdm
-from ..model.Model import AIIA
+from transformers import PreTrainedModel
 from ..model.config import AIIAConfig
 from ..data.DataLoader import AIIADataLoader
 import os
@@ -21,7 +23,7 @@ class ProjectionHead(nn.Module):
             return self.conv_rotate(x).mean(dim=(2, 3))  # Global average pooling for rotation task
 
 class Pretrainer:
-    def __init__(self, model: AIIA, learning_rate=1e-4, config: AIIAConfig=None):
+    def __init__(self, model: PreTrainedModel, learning_rate=1e-4, config: AIIAConfig=None):
         """
         Initialize the pretrainer with a model.
         
@@ -112,20 +114,169 @@ class Pretrainer:
             
         return batch_loss
 
-    def train(self, dataset_paths,output_path:str="AIIA", column="image_bytes",  num_epochs=3, batch_size=2, sample_size=10000):
+    def _save_checkpoint(self, checkpoint_dir, epoch, batch_count, checkpoint_name):
-        """
+        """Save a model checkpoint.
-        Train the model using multiple specified datasets.
         
         Args:
-            dataset_paths (list): List of paths to parquet datasets
+            checkpoint_dir (str): Directory to save the checkpoint
-            num_epochs (int): Number of training epochs
+            epoch (int): Current epoch number
-            batch_size (int): Batch size for training
+            batch_count (int): Current batch count
-            sample_size (int): Number of samples to use from each dataset
+            checkpoint_name (str): Name for the checkpoint file
+            
+        Returns:
+            str: Path to the saved checkpoint
         """
+        checkpoint_path = os.path.join(checkpoint_dir, checkpoint_name)
+        checkpoint_data = {
+            'epoch': epoch + 1,
+            'batch': batch_count,
+            'model_state_dict': self.model.state_dict(),
+            'projection_head_state_dict': self.projection_head.state_dict(),
+            'optimizer_state_dict': self.optimizer.state_dict(),
+            'train_losses': self.train_losses,
+            'val_losses': self.val_losses,
+        }
+        torch.save(checkpoint_data, checkpoint_path)
+        return checkpoint_path
+
+    def load_checkpoint(self, checkpoint_dir, specific_checkpoint=None):
+        """
+        Check for checkpoints and load if available.
+        
+        Args:
+            checkpoint_dir (str): Directory where checkpoints are stored
+            specific_checkpoint (str, optional): Specific checkpoint file to load. If None, loads the most recent.
+            
+        Returns:
+            tuple: (loaded_epoch, loaded_batch) if checkpoint was loaded, None otherwise
+        """
+        # Create checkpoint directory if it doesn't exist
+        os.makedirs(checkpoint_dir, exist_ok=True)
+        
+        # If a specific checkpoint is requested
+        if specific_checkpoint:
+            checkpoint_path = os.path.join(checkpoint_dir, specific_checkpoint)
+            if os.path.exists(checkpoint_path):
+                return self._load_checkpoint_file(checkpoint_path)
+            else:
+                print(f"Specified checkpoint {specific_checkpoint} not found.")
+                return None
+        
+        # Find all checkpoint files
+        checkpoint_files = [f for f in os.listdir(checkpoint_dir) if f.startswith("checkpoint_") and f.endswith(".pt")]
+        
+        if not checkpoint_files:
+            print("No checkpoints found in directory.")
+            return None
+        
+        # Find the most recent checkpoint
+        checkpoint_files.sort(key=lambda x: os.path.getmtime(os.path.join(checkpoint_dir, x)), reverse=True)
+        most_recent = checkpoint_files[0]
+        checkpoint_path = os.path.join(checkpoint_dir, most_recent)
+        
+        return self._load_checkpoint_file(checkpoint_path)
+
+    def _load_checkpoint_file(self, checkpoint_path):
+        """
+        Load a specific checkpoint file.
+        
+        Args:
+            checkpoint_path (str): Path to the checkpoint file
+            
+        Returns:
+            tuple: (loaded_epoch, loaded_batch) if checkpoint was loaded, None otherwise
+        """
+        try:
+            checkpoint = torch.load(checkpoint_path, map_location=self.device)
+            
+            # Load model state
+            self.model.load_state_dict(checkpoint['model_state_dict'])
+            
+            # Load projection head state
+            self.projection_head.load_state_dict(checkpoint['projection_head_state_dict'])
+            
+            # Load optimizer state
+            self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
+            
+            # Load loss history
+            self.train_losses = checkpoint.get('train_losses', [])
+            self.val_losses = checkpoint.get('val_losses', [])
+            
+            loaded_epoch = checkpoint['epoch']
+            loaded_batch = checkpoint['batch']
+            
+            print(f"Checkpoint loaded from {checkpoint_path}")
+            print(f"Resuming from epoch {loaded_epoch}, batch {loaded_batch}")
+            
+            return loaded_epoch, loaded_batch
+            
+        except Exception as e:
+            print(f"Error loading checkpoint: {e}")
+            return None
+
+
+    def train(self, dataset_paths, output_path="AIIA", column="image_bytes",
+              num_epochs=3, batch_size=2, sample_size=10000, checkpoint_dir=None):
+        """Train the model using multiple specified datasets with checkpoint resumption support."""
         if not dataset_paths:
             raise ValueError("No dataset paths provided")
 
-        # Read and merge all datasets
+        self._initialize_checkpoint_variables()
+        start_epoch, start_batch, resume_training = self._load_checkpoints(checkpoint_dir)
+
+        dataframes = self._load_and_merge_datasets(dataset_paths, sample_size)
+        aiia_loader = self._initialize_data_loader(dataframes, column, batch_size)
+
+        criterion_denoise, criterion_rotate, best_val_loss = self._initialize_loss_functions()
+
+        for epoch in range(start_epoch, num_epochs):
+            print(f"\nEpoch {epoch+1}/{num_epochs}")
+            print("-" * 20)
+            total_train_loss, batch_count = self._training_phase(aiia_loader.train_loader,
+                                                                 start_batch if (epoch == start_epoch and resume_training) else 0,
+                                                                 criterion_denoise,
+                                                                 criterion_rotate)
+
+            avg_train_loss = total_train_loss / max(batch_count, 1)
+            self.train_losses.append(avg_train_loss)
+            print(f"Training Loss: {avg_train_loss:.4f}")
+
+            val_loss = self._validation_phase(aiia_loader.val_loader, criterion_denoise, criterion_rotate)
+
+            if val_loss < best_val_loss:
+                best_val_loss = val_loss
+                self.model.save(output_path)
+                print("Best model saved!")
+
+        losses_path = os.path.join(os.path.dirname(output_path), 'losses.csv')
+        self.save_losses(losses_path)
+
+    def _initialize_checkpoint_variables(self):
+        """Initialize checkpoint tracking variables."""
+        self.last_checkpoint_time = time.time()
+        self.checkpoint_interval = 2 * 60 * 60  # 2 hours in seconds
+        self.last_22_date = None
+        self.recent_checkpoints = []
+
+    def _load_checkpoints(self, checkpoint_dir):
+        """Load checkpoints and return start epoch, batch, and resumption flag."""
+        start_epoch = 0
+        start_batch = 0
+        resume_training = False
+
+        if checkpoint_dir is not None:
+            os.makedirs(checkpoint_dir, exist_ok=True)
+            checkpoint_info = self.load_checkpoint(checkpoint_dir)
+            if checkpoint_info:
+                start_epoch, start_batch = checkpoint_info
+                resume_training = True
+                # Adjust epoch to be 0-indexed for the loop
+                start_epoch -= 1
+
+        return start_epoch, start_batch, resume_training
+
+    def _load_and_merge_datasets(self, dataset_paths, sample_size):
+        """Load and merge datasets."""
         dataframes = []
         for path in dataset_paths:
             try:
@@ -137,10 +288,11 @@ class Pretrainer:
         if not dataframes:
             raise ValueError("No valid datasets could be loaded")
 
-        merged_df = pd.concat(dataframes, ignore_index=True)
+        return pd.concat(dataframes, ignore_index=True)
 
-        # Initialize data loader
+    def _initialize_data_loader(self, merged_df, column, batch_size):
-        aiia_loader = AIIADataLoader(
+        """Initialize the data loader."""
+        return AIIADataLoader(
             merged_df,
             column=column,
             batch_size=batch_size,
@@ -148,24 +300,30 @@ class Pretrainer:
             collate_fn=self.safe_collate
         )
 
+    def _initialize_loss_functions(self):
+        """Initialize loss functions and tracking variables."""
         criterion_denoise = nn.MSELoss()
         criterion_rotate = nn.CrossEntropyLoss()
         best_val_loss = float('inf')
+        return criterion_denoise, criterion_rotate, best_val_loss
 
-        for epoch in range(num_epochs):
+    def _training_phase(self, train_loader, skip_batches, criterion_denoise, criterion_rotate):
-            print(f"\nEpoch {epoch+1}/{num_epochs}")
+        """Handle the training phase."""
-            print("-" * 20)
-            
-            # Training phase
         self.model.train()
         self.projection_head.train()
         total_train_loss = 0.0
         batch_count = 0
 
-            for batch_data in tqdm(aiia_loader.train_loader):
+        train_batches = list(enumerate(train_loader))
+        for i, batch_data in tqdm(train_batches[skip_batches:],
+                                  initial=skip_batches,
+                                  total=len(train_batches)):
             if batch_data is None:
                 continue
 
+            current_batch = i + 1
+            self._handle_checkpoints(current_batch)
+
             self.optimizer.zero_grad()
             batch_loss = self._process_batch(batch_data, criterion_denoise, criterion_rotate)
 
@@ -175,22 +333,42 @@ class Pretrainer:
                 total_train_loss += batch_loss.item()
                 batch_count += 1
 
-            avg_train_loss = total_train_loss / max(batch_count, 1)
+        return total_train_loss, batch_count
-            self.train_losses.append(avg_train_loss)
-            print(f"Training Loss: {avg_train_loss:.4f}")
 
-            # Validation phase
+    def _handle_checkpoints(self, current_batch):
+        """Handle checkpoint saving logic."""
+        current_time = time.time()
+        current_dt = datetime.datetime.now(datetime.timezone(datetime.timedelta(hours=2)))  # German time
+        today = current_dt.date()
+
+        if self.checkpoint_dir and (current_time - self.last_checkpoint_time) >= self.checkpoint_interval:
+            checkpoint_name = f"checkpoint_epoch{self.current_epoch+1}_batch{current_batch}.pt"
+            checkpoint_path = self._save_checkpoint(self.checkpoint_dir, self.current_epoch, current_batch, checkpoint_name)
+
+            # Track and maintain only 3 recent checkpoints
+            self.recent_checkpoints.append(checkpoint_path)
+            if len(self.recent_checkpoints) > 3:
+                oldest = self.recent_checkpoints.pop(0)
+                if os.path.exists(oldest):
+                    os.remove(oldest)
+
+            self.last_checkpoint_time = current_time
+            print(f"Checkpoint saved at {checkpoint_path}")
+
+        # Special 22:00 checkpoint (considering it's currently 10:15 PM)
+        is_22_oclock = current_dt.hour == 22 and current_dt.minute < 15
+
+        if self.checkpoint_dir and is_22_oclock and self.last_22_date != today:
+            checkpoint_name = f"checkpoint_22h_{today.strftime('%Y%m%d')}.pt"
+            checkpoint_path = self._save_checkpoint(self.checkpoint_dir, self.current_epoch, current_batch, checkpoint_name)
+            self.last_22_date = today
+            print(f"22:00 Checkpoint saved at {checkpoint_path}")
+
+    def _validation_phase(self, val_loader, criterion_denoise, criterion_rotate):
+        """Handle the validation phase."""
         self.model.eval()
         self.projection_head.eval()
-            val_loss = self._validate(aiia_loader.val_loader, criterion_denoise, criterion_rotate)
+        return self._validate(val_loader, criterion_denoise, criterion_rotate)
-            
-            if val_loss < best_val_loss:
-                best_val_loss = val_loss
-                self.model.save(output_path)
-                print("Best model saved!")
-
-        losses_path = os.path.join(os.path.dirname(output_path), 'losses.csv')
-        self.save_losses(losses_path)
 
     def _validate(self, val_loader, criterion_denoise, criterion_rotate):
         """Perform validation and return average validation loss."""

tests/model/test_aiia.py

@@ -1,159 +1,133 @@
 import os
 import torch
-from aiia.model.Model import AIIABase, AIIABaseShared, AIIAExpert, AIIAmoe, AIIAchunked, AIIAConfig, AIIASparseMoe
+from aiia.model.Model import AIIABase, AIIABaseShared, AIIAExpert, AIIAmoe, AIIAConfig, AIIASparseMoe
 
 def test_aiiabase_creation():
     config = AIIAConfig()
     model = AIIABase(config)
     assert isinstance(model, AIIABase)
 
-def test_aiiabase_save_load():
+def test_aiiabase_save_pretrained_from_pretrained():
     config = AIIAConfig()
     model = AIIABase(config)
-    save_path = "test_aiiabase_save_load"
+    save_pretrained_path = "test_aiiabase_save_pretrained_load"
 
     # Save the model
-    model.save(save_path)
+    model.save_pretrained(save_pretrained_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "model.safetensors"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "config.json"))
 
     # Load the model
-    loaded_model = AIIABase.load(save_path)
+    loaded_model = AIIABase.from_pretrained(save_pretrained_path)
 
     # Check if the loaded model is an instance of AIIABase
     assert isinstance(loaded_model, AIIABase)
 
     # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
+    os.remove(os.path.join(save_pretrained_path, "model.safetensors"))
-    os.remove(os.path.join(save_path, "config.json"))
+    os.remove(os.path.join(save_pretrained_path, "config.json"))
-    os.rmdir(save_path)
+    os.rmdir(save_pretrained_path)
 
 def test_aiiabase_shared_creation():
     config = AIIAConfig()
     model = AIIABaseShared(config)
     assert isinstance(model, AIIABaseShared)
 
-def test_aiiabase_shared_save_load():
+def test_aiiabase_shared_save_pretrained_from_pretrained():
     config = AIIAConfig()
     model = AIIABaseShared(config)
-    save_path = "test_aiiabase_shared_save_load"
+    save_pretrained_path = "test_aiiabase_shared_save_pretrained_load"
 
     # Save the model
-    model.save(save_path)
+    model.save_pretrained(save_pretrained_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "model.safetensors"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "config.json"))
 
     # Load the model
-    loaded_model = AIIABaseShared.load(save_path)
+    loaded_model = AIIABaseShared.from_pretrained(save_pretrained_path)
 
     # Check if the loaded model is an instance of AIIABaseShared
     assert isinstance(loaded_model, AIIABaseShared)
 
     # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
+    os.remove(os.path.join(save_pretrained_path, "model.safetensors"))
-    os.remove(os.path.join(save_path, "config.json"))
+    os.remove(os.path.join(save_pretrained_path, "config.json"))
-    os.rmdir(save_path)
+    os.rmdir(save_pretrained_path)
 
 def test_aiiaexpert_creation():
     config = AIIAConfig()
     model = AIIAExpert(config)
     assert isinstance(model, AIIAExpert)
 
-def test_aiiaexpert_save_load():
+def test_aiiaexpert_save_pretrained_from_pretrained():
     config = AIIAConfig()
     model = AIIAExpert(config)
-    save_path = "test_aiiaexpert_save_load"
+    save_pretrained_path = "test_aiiaexpert_save_pretrained_load"
 
     # Save the model
-    model.save(save_path)
+    model.save_pretrained(save_pretrained_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "model.safetensors"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "config.json"))
 
     # Load the model
-    loaded_model = AIIAExpert.load(save_path)
+    loaded_model = AIIAExpert.from_pretrained(save_pretrained_path)
 
     # Check if the loaded model is an instance of AIIAExpert
     assert isinstance(loaded_model, AIIAExpert)
 
     # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
+    os.remove(os.path.join(save_pretrained_path, "model.safetensors"))
-    os.remove(os.path.join(save_path, "config.json"))
+    os.remove(os.path.join(save_pretrained_path, "config.json"))
-    os.rmdir(save_path)
+    os.rmdir(save_pretrained_path)
 
 def test_aiiamoe_creation():
-    config = AIIAConfig()
+    config = AIIAConfig(num_experts=3)
-    model = AIIAmoe(config, num_experts=5)
+    model = AIIAmoe(config)
     assert isinstance(model, AIIAmoe)
 
-def test_aiiamoe_save_load():
+def test_aiiamoe_save_pretrained_from_pretrained():
-    config = AIIAConfig()
+    config = AIIAConfig(num_experts=3)
-    model = AIIAmoe(config, num_experts=5)
+    model = AIIAmoe(config)
-    save_path = "test_aiiamoe_save_load"
+    save_pretrained_path = "test_aiiamoe_save_pretrained_load"
 
     # Save the model
-    model.save(save_path)
+    model.save_pretrained(save_pretrained_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "model.safetensors"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "config.json"))
 
     # Load the model
-    loaded_model = AIIAmoe.load(save_path)
+    loaded_model = AIIAmoe.from_pretrained(save_pretrained_path)
 
     # Check if the loaded model is an instance of AIIAmoe
     assert isinstance(loaded_model, AIIAmoe)
 
     # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
+    os.remove(os.path.join(save_pretrained_path, "model.safetensors"))
-    os.remove(os.path.join(save_path, "config.json"))
+    os.remove(os.path.join(save_pretrained_path, "config.json"))
-    os.rmdir(save_path)
+    os.rmdir(save_pretrained_path)
 
 def test_aiiasparsemoe_creation():
-    config = AIIAConfig()
+    config = AIIAConfig(num_experts=5, top_k=2)
-    model = AIIASparseMoe(config, num_experts=5, top_k=2)
+    model = AIIASparseMoe(config, base_class=AIIABaseShared)
     assert isinstance(model, AIIASparseMoe)
 
-def test_aiiasparsemoe_save_load():
+def test_aiiasparsemoe_save_pretrained_from_pretrained():
-    config = AIIAConfig()
+    config = AIIAConfig(num_experts=3, top_k=1)
-    model = AIIASparseMoe(config, num_experts=3, top_k=1)
+    model = AIIASparseMoe(config)
-    save_path = "test_aiiasparsemoe_save_load"
+    save_pretrained_path = "test_aiiasparsemoe_save_pretrained_load"
 
     # Save the model
-    model.save(save_path)
+    model.save_pretrained(save_pretrained_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "model.safetensors"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
+    assert os.path.exists(os.path.join(save_pretrained_path, "config.json"))
 
     # Load the model
-    loaded_model = AIIASparseMoe.load(save_path)
+    loaded_model = AIIASparseMoe.from_pretrained(save_pretrained_path)
 
     # Check if the loaded model is an instance of AIIASparseMoe
     assert isinstance(loaded_model, AIIASparseMoe)
 
     # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
+    os.remove(os.path.join(save_pretrained_path, "model.safetensors"))
-    os.remove(os.path.join(save_path, "config.json"))
+    os.remove(os.path.join(save_pretrained_path, "config.json"))
-    os.rmdir(save_path)
+    os.rmdir(save_pretrained_path)
-
-def test_aiiachunked_creation():
-    config = AIIAConfig()
-    model = AIIAchunked(config)
-    assert isinstance(model, AIIAchunked)
-
-def test_aiiachunked_save_load():
-    config = AIIAConfig()
-    model = AIIAchunked(config)
-    save_path = "test_aiiachunked_save_load"
-
-    # Save the model
-    model.save(save_path)
-    assert os.path.exists(os.path.join(save_path, "model.pth"))
-    assert os.path.exists(os.path.join(save_path, "config.json"))
-
-    # Load the model
-    loaded_model = AIIAchunked.load(save_path)
-
-    # Check if the loaded model is an instance of AIIAchunked
-    assert isinstance(loaded_model, AIIAchunked)
-
-    # Clean up
-    os.remove(os.path.join(save_path, "model.pth"))
-    os.remove(os.path.join(save_path, "config.json"))
-    os.rmdir(save_path)

tests/model/test_config.py

@@ -1,75 +1,77 @@
 import os
 import tempfile
 import pytest
-import torch.nn as nn
 from aiia import AIIAConfig
 
+
 def test_aiia_config_initialization():
     config = AIIAConfig()
-    assert config.model_name == "AIIA"
+    assert config.model_type == "AIIA"
     assert config.kernel_size == 3
     assert config.activation_function == "GELU"
     assert config.hidden_size == 512
     assert config.num_hidden_layers == 12
     assert config.num_channels == 3
-    assert config.learning_rate == 5e-5
+
 
 def test_aiia_config_custom_initialization():
     config = AIIAConfig(
-        model_name="CustomModel",
+        model_type="CustomModel",
         kernel_size=5,
         activation_function="ReLU",
         hidden_size=1024,
         num_hidden_layers=8,
-        num_channels=1,
+        num_channels=1
-        learning_rate=1e-4
     )
-    assert config.model_name == "CustomModel"
+    assert config.model_type == "CustomModel"
     assert config.kernel_size == 5
     assert config.activation_function == "ReLU"
     assert config.hidden_size == 1024
     assert config.num_hidden_layers == 8
     assert config.num_channels == 1
-    assert config.learning_rate == 1e-4
+
 
 def test_aiia_config_invalid_activation_function():
     with pytest.raises(ValueError):
         AIIAConfig(activation_function="InvalidFunction")
 
+
 def test_aiia_config_to_dict():
     config = AIIAConfig()
     config_dict = config.to_dict()
     assert isinstance(config_dict, dict)
-    assert config_dict["model_name"] == "AIIA"
     assert config_dict["kernel_size"] == 3
 
-def test_aiia_config_save_and_load():
-    with tempfile.TemporaryDirectory() as tmpdir:
-        config = AIIAConfig(model_name="TempModel")
-        save_path = os.path.join(tmpdir, "config")
-        config.save(save_path)
 
-        loaded_config = AIIAConfig.load(save_path)
+def test_aiia_config_save_pretrained_and_from_pretrained():
-        assert loaded_config.model_name == "TempModel"
+    with tempfile.TemporaryDirectory() as tmpdir:
+        config = AIIAConfig(model_type="TempModel")
+        save_pretrained_path = os.path.join(tmpdir, "config")
+        config.save_pretrained(save_pretrained_path)
+
+        loaded_config = AIIAConfig.from_pretrained(save_pretrained_path)
+        assert loaded_config.model_type == "TempModel"
         assert loaded_config.kernel_size == 3
         assert loaded_config.activation_function == "GELU"
 
-def test_aiia_config_save_and_load_with_custom_attributes():
-    with tempfile.TemporaryDirectory() as tmpdir:
-        config = AIIAConfig(model_name="TempModel", custom_attr="value")
-        save_path = os.path.join(tmpdir, "config")
-        config.save(save_path)
 
-        loaded_config = AIIAConfig.load(save_path)
+def test_aiia_config_save_pretrained_and_load_with_custom_attributes():
-        assert loaded_config.model_name == "TempModel"
+    with tempfile.TemporaryDirectory() as tmpdir:
+        config = AIIAConfig(model_type="TempModel", custom_attr="value")
+        save_pretrained_path = os.path.join(tmpdir, "config")
+        config.save_pretrained(save_pretrained_path)
+
+        loaded_config = AIIAConfig.from_pretrained(save_pretrained_path)
+        assert loaded_config.model_type == "TempModel"
         assert loaded_config.custom_attr == "value"
 
-def test_aiia_config_save_and_load_with_nested_attributes():
-    with tempfile.TemporaryDirectory() as tmpdir:
-        config = AIIAConfig(model_name="TempModel", nested={"key": "value"})
-        save_path = os.path.join(tmpdir, "config")
-        config.save(save_path)
 
-        loaded_config = AIIAConfig.load(save_path)
+def test_aiia_config_save_pretrained_and_load_with_nested_attributes():
-        assert loaded_config.model_name == "TempModel"
+    with tempfile.TemporaryDirectory() as tmpdir:
+        config = AIIAConfig(model_type="TempModel", nested={"key": "value"})
+        save_pretrained_path = os.path.join(tmpdir, "config")
+        config.save_pretrained(save_pretrained_path)
+
+        loaded_config = AIIAConfig.from_pretrained(save_pretrained_path)
+        assert loaded_config.model_type == "TempModel"
         assert loaded_config.nested == {"key": "value"}

tests/pretrain/test_pretrainer.py

@@ -3,6 +3,8 @@ import torch
 from unittest.mock import MagicMock, patch, MagicMock, mock_open, call
 from aiia import Pretrainer, ProjectionHead, AIIABase, AIIAConfig
 import pandas as pd
+import os
+import datetime
 
 # Test the ProjectionHead class
 def test_projection_head():
@@ -53,11 +55,94 @@ def test_process_batch(mock_process_batch):
     loss = pretrainer._process_batch(batch_data, criterion_denoise, criterion_rotate)
     assert loss == 0.5
 
+# Error cases
+# New tests for checkpoint handling
+@patch('torch.save')
+@patch('os.path.join')
+def test_save_checkpoint(mock_join, mock_save):
+    """Test checkpoint saving functionality."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    pretrainer.projection_head = MagicMock()
+    pretrainer.optimizer = MagicMock()
+    
+    checkpoint_dir = "checkpoints"
+    epoch = 1
+    batch_count = 100
+    checkpoint_name = "test_checkpoint.pt"
+    
+    mock_join.return_value = os.path.join(checkpoint_dir, checkpoint_name)
+    
+    path = pretrainer._save_checkpoint(checkpoint_dir, epoch, batch_count, checkpoint_name)
+    
+    assert path == os.path.join(checkpoint_dir, checkpoint_name)
+    mock_save.assert_called_once()
+
+@patch('os.makedirs')
+@patch('os.path.exists')
+@patch('torch.load')
+def test_load_checkpoint_specific(mock_load, mock_exists, mock_makedirs):
+    """Test loading a specific checkpoint."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    pretrainer.projection_head = MagicMock()
+    pretrainer.optimizer = MagicMock()
+    
+    checkpoint_dir = "checkpoints"
+    specific_checkpoint = "specific_checkpoint.pt"
+    mock_exists.return_value = True
+    
+    mock_load.return_value = {
+        'epoch': 2,
+        'batch': 150,
+        'model_state_dict': {},
+        'projection_head_state_dict': {},
+        'optimizer_state_dict': {},
+        'train_losses': [],
+        'val_losses': []
+    }
+    
+    result = pretrainer.load_checkpoint(checkpoint_dir, specific_checkpoint)
+    assert result == (2, 150)
+
+@patch('os.listdir')
+@patch('os.path.getmtime')
+def test_load_checkpoint_most_recent(mock_getmtime, mock_listdir):
+    """Test loading the most recent checkpoint."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    
+    checkpoint_dir = "checkpoints"
+    mock_listdir.return_value = ["checkpoint_1.pt", "checkpoint_2.pt"]
+    mock_getmtime.side_effect = [100, 200]  # checkpoint_2.pt is more recent
+    
+    with patch.object(pretrainer, '_load_checkpoint_file', return_value=(2, 150)):
+        result = pretrainer.load_checkpoint(checkpoint_dir)
+        assert result == (2, 150)
+
+def test_initialize_checkpoint_variables():
+    """Test initialization of checkpoint variables."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    pretrainer._initialize_checkpoint_variables()
+    
+    assert hasattr(pretrainer, 'last_checkpoint_time')
+    assert pretrainer.checkpoint_interval == 2 * 60 * 60
+    assert pretrainer.last_22_date is None
+    assert pretrainer.recent_checkpoints == []
+
+@patch('torch.nn.MSELoss')
+@patch('torch.nn.CrossEntropyLoss')
+def test_initialize_loss_functions(mock_ce_loss, mock_mse_loss):
+    """Test loss function initialization."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    criterion_denoise, criterion_rotate, best_val_loss = pretrainer._initialize_loss_functions()
+    
+    assert mock_mse_loss.called
+    assert mock_ce_loss.called
+    assert best_val_loss == float('inf')
+
 @patch('pandas.concat')
 @patch('pandas.read_parquet')
 @patch('aiia.pretrain.pretrainer.AIIADataLoader')
 @patch('os.path.join', return_value='mocked/path/model.pt')
-@patch('builtins.print')  # Add this to mock the print function
+@patch('builtins.print')
 def test_train_happy_path(mock_print, mock_path_join, mock_data_loader, mock_read_parquet, mock_concat):
     """Test the train method under normal conditions with comprehensive verification."""
     # Setup test data and mocks
@@ -73,6 +158,7 @@ def test_train_happy_path(mock_print, mock_path_join, mock_data_loader, mock_rea
     pretrainer = Pretrainer(model=mock_model, config=AIIAConfig())
     pretrainer.projection_head = mock_projection_head
     pretrainer.optimizer = MagicMock()
+    pretrainer.checkpoint_dir = None  # Initialize checkpoint_dir
 
     # Setup dataset paths and mock batch data
     dataset_paths = ['path/to/dataset1.parquet', 'path/to/dataset2.parquet']
@@ -104,185 +190,118 @@ def test_train_happy_path(mock_print, mock_path_join, mock_data_loader, mock_rea
     assert mock_process_batch.call_count == 2
     assert mock_validate.call_count == 2
     
-    # Check for "Best model saved!" instead of model.save()
     mock_print.assert_any_call("Best model saved!")
-    
     mock_save_losses.assert_called_once()
 
-    # Verify state changes
     assert len(pretrainer.train_losses) == 2
     assert pretrainer.train_losses == [0.5, 0.5]
 
-
+@patch('datetime.datetime')
-# Error cases
+@patch('time.time')
-def test_train_no_dataset_paths():
+def test_handle_checkpoints(mock_time, mock_datetime):
-    """Test ValueError when no dataset paths are provided."""
+    """Test checkpoint handling logic."""
     pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    pretrainer.checkpoint_dir = "checkpoints"
+    pretrainer.current_epoch = 1
+    pretrainer._initialize_checkpoint_variables()
     
-    with pytest.raises(ValueError, match="No dataset paths provided"):
+    # Set a base time value
-        pretrainer.train([])
+    base_time = 1000
+    # Set the last checkpoint time to base_time
+    pretrainer.last_checkpoint_time = base_time
     
-@patch('pandas.read_parquet')
+    # Mock time to return base_time + interval + 1 to trigger checkpoint save
-def test_train_all_datasets_fail(mock_read_parquet):
+    mock_time.return_value = base_time + pretrainer.checkpoint_interval + 1
-    """Test handling when all datasets fail to load."""
-    mock_read_parquet.side_effect = Exception("Failed to load dataset")
     
+    # Mock datetime for 22:00 checkpoint
+    mock_dt = MagicMock()
+    mock_dt.hour = 22
+    mock_dt.minute = 0
+    mock_dt.date.return_value = datetime.date(2023, 1, 1)
+    mock_datetime.now.return_value = mock_dt
+    
+    with patch.object(pretrainer, '_save_checkpoint') as mock_save:
+        pretrainer._handle_checkpoints(100)
+        # Should be called twice - once for regular interval and once for 22:00
+        assert mock_save.call_count == 2
+
+def test_training_phase():
+    """Test the training phase logic."""
     pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
-    dataset_paths = ['path/to/dataset1.parquet', 'path/to/dataset2.parquet']
-
-    with pytest.raises(ValueError, match="No valid datasets could be loaded"):
-        pretrainer.train(dataset_paths)
-
-# Edge cases
-@patch('pandas.concat')
-@patch('pandas.read_parquet')
-@patch('aiia.pretrain.pretrainer.AIIADataLoader')
-def test_train_empty_loaders(mock_data_loader, mock_read_parquet, mock_concat):
-    """Test behavior with empty data loaders."""
-    real_df = pd.DataFrame({'image_bytes': [torch.randn(1, 3, 224, 224).tolist()]})
-    mock_read_parquet.return_value.head.return_value = real_df
-    mock_concat.return_value = real_df
-
-    loader_instance = MagicMock()
-    loader_instance.train_loader = []  # Empty train loader
-    loader_instance.val_loader = []    # Empty val loader
-    mock_data_loader.return_value = loader_instance
-
-    mock_model = MagicMock()
-    pretrainer = Pretrainer(model=mock_model, config=AIIAConfig())
-    pretrainer.projection_head = MagicMock()
     pretrainer.optimizer = MagicMock()
+    pretrainer.checkpoint_dir = None  # Initialize checkpoint_dir
+    pretrainer._initialize_checkpoint_variables()
+    pretrainer.current_epoch = 0
     
-    with patch.object(Pretrainer, 'save_losses') as mock_save_losses:
+    # Create mock batch data with requires_grad=True
-        pretrainer.train(['path/to/dataset.parquet'], num_epochs=1)
-
-    # Verify empty loader behavior
-    assert len(pretrainer.train_losses) == 1
-    assert pretrainer.train_losses[0] == 0.0
-    mock_save_losses.assert_called_once()
-
-@patch('pandas.concat')
-@patch('pandas.read_parquet')
-@patch('aiia.pretrain.pretrainer.AIIADataLoader')
-def test_train_none_batch_data(mock_data_loader, mock_read_parquet, mock_concat):
-    """Test behavior when batch_data is None."""
-    real_df = pd.DataFrame({'image_bytes': [torch.randn(1, 3, 224, 224).tolist()]})
-    mock_read_parquet.return_value.head.return_value = real_df
-    mock_concat.return_value = real_df
-
-    loader_instance = MagicMock()
-    loader_instance.train_loader = [None]  # Loader returns None
-    loader_instance.val_loader = []
-    mock_data_loader.return_value = loader_instance
-
-    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
-    pretrainer.projection_head = MagicMock()
-    pretrainer.optimizer = MagicMock()
-
-    with patch.object(Pretrainer, '_process_batch') as mock_process_batch, \
-         patch.object(Pretrainer, 'save_losses'):
-        pretrainer.train(['path/to/dataset.parquet'], num_epochs=1)
-
-    # Verify None batch handling
-    mock_process_batch.assert_not_called()
-    assert pretrainer.train_losses[0] == 0.0
-
-# Parameter variations
-@patch('pandas.concat')
-@patch('pandas.read_parquet')
-@patch('aiia.pretrain.pretrainer.AIIADataLoader')
-def test_train_with_custom_parameters(mock_data_loader, mock_read_parquet, mock_concat):
-    """Test that custom parameters are properly passed through."""
-    real_df = pd.DataFrame({'image_bytes': [torch.randn(1, 3, 224, 224).tolist()]})
-    mock_read_parquet.return_value.head.return_value = real_df
-    mock_concat.return_value = real_df
-
-    loader_instance = MagicMock()
-    loader_instance.train_loader = []
-    loader_instance.val_loader = []
-    mock_data_loader.return_value = loader_instance
-
-    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
-    pretrainer.projection_head = MagicMock()
-    pretrainer.optimizer = MagicMock()
-
-    # Custom parameters
-    custom_output_path = "custom/output/path"
-    custom_column = "custom_column"
-    custom_batch_size = 16
-    custom_sample_size = 5000
-
-    with patch.object(Pretrainer, 'save_losses'):
-        pretrainer.train(
-            ['path/to/dataset.parquet'],
-            output_path=custom_output_path,
-            column=custom_column,
-            batch_size=custom_batch_size,
-            sample_size=custom_sample_size
-        )
-
-    # Verify custom parameters were used
-    mock_read_parquet.return_value.head.assert_called_once_with(custom_sample_size)
-    assert mock_data_loader.call_args[1]['column'] == custom_column
-    assert mock_data_loader.call_args[1]['batch_size'] == custom_batch_size
-
-
-
-@patch('pandas.concat')
-@patch('pandas.read_parquet')
-@patch('aiia.pretrain.pretrainer.AIIADataLoader')
-@patch('builtins.print')  # Add this to mock the print function
-def test_train_validation_loss_improvement(mock_print, mock_data_loader, mock_read_parquet, mock_concat):
-    """Test that model is saved only when validation loss improves."""
-    real_df = pd.DataFrame({'image_bytes': [torch.randn(1, 3, 224, 224).tolist()]})
-    mock_read_parquet.return_value.head.return_value = real_df
-    mock_concat.return_value = real_df
-
-    # Create mock batch data with proper structure
     mock_batch_data = {
-        'denoise': (torch.randn(2, 3, 32, 32), torch.randn(2, 3, 32, 32)),
+        'denoise': (
-        'rotate': (torch.randn(2, 3, 32, 32), torch.tensor([0, 1]))
+            torch.randn(2, 3, 32, 32, requires_grad=True), 
+            torch.randn(2, 3, 32, 32, requires_grad=True)
+        ),
+        'rotate': (
+            torch.randn(2, 3, 32, 32, requires_grad=True), 
+            torch.tensor([0, 1], dtype=torch.long)  # Labels typically don't need gradients
+        )
+    }
+    mock_train_loader = [(0, mock_batch_data)]  # Include batch index
+    
+    # Mock the loss functions to return tensors that require gradients
+    criterion_denoise = MagicMock(return_value=torch.tensor(0.5, requires_grad=True))
+    criterion_rotate = MagicMock(return_value=torch.tensor(0.5, requires_grad=True))
+    
+    with patch.object(pretrainer, '_process_batch', return_value=torch.tensor(0.5, requires_grad=True)), \
+         patch.object(pretrainer, '_handle_checkpoints') as mock_handle_checkpoints:
+        
+        total_loss, batch_count = pretrainer._training_phase(
+            mock_train_loader, 0, criterion_denoise, criterion_rotate)
+        
+        assert total_loss == 0.5
+        assert batch_count == 1
+        mock_handle_checkpoints.assert_called_once_with(1)  # Check if checkpoint handling was called
+
+def test_validation_phase():
+    """Test the validation phase logic."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    pretrainer.projection_head = MagicMock()
+    
+    mock_val_loader = [MagicMock()]
+    criterion_denoise = MagicMock()
+    criterion_rotate = MagicMock()
+    
+    with patch.object(pretrainer, '_validate', return_value=0.4):
+        val_loss = pretrainer._validation_phase(
+            mock_val_loader, criterion_denoise, criterion_rotate)
+        
+        assert val_loss == 0.4
+
+@patch('pandas.read_parquet')
+def test_load_and_merge_datasets(mock_read_parquet):
+    """Test dataset loading and merging."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    
+    mock_df = pd.DataFrame({'col': [1, 2, 3]})
+    mock_read_parquet.return_value.head.return_value = mock_df
+    
+    result = pretrainer._load_and_merge_datasets(['path1.parquet', 'path2.parquet'], 1000)
+    assert len(result) == 6  # 2 datasets * 3 rows each
+
+def test_process_batch_none_tasks():
+    """Test processing batch with no tasks."""
+    pretrainer = Pretrainer(model=MagicMock(), config=AIIAConfig())
+    
+    batch_data = {
+        'denoise': None,
+        'rotate': None
     }
     
-    loader_instance = MagicMock()
+    loss = pretrainer._process_batch(
-    loader_instance.train_loader = [mock_batch_data]
+        batch_data, 
-    loader_instance.val_loader = [mock_batch_data]
+        criterion_denoise=MagicMock(),
-    mock_data_loader.return_value = loader_instance
+        criterion_rotate=MagicMock()
+    )
     
-    mock_model = MagicMock()
+    assert loss == 0
-    pretrainer = Pretrainer(model=mock_model, config=AIIAConfig())
-    pretrainer.projection_head = MagicMock()
-    pretrainer.optimizer = MagicMock()
-
-    # Initialize the best validation loss
-    pretrainer.best_val_loss = float('inf')
-
-    mock_batch_loss = torch.tensor(0.5, requires_grad=True)
-
-    # Test improving validation loss
-    with patch.object(Pretrainer, '_process_batch', return_value=mock_batch_loss), \
-        patch.object(Pretrainer, '_validate', side_effect=[3.0, 2.0, 1.0]), \
-        patch.object(Pretrainer, 'save_losses'):
-        pretrainer.train(['path/to/dataset.parquet'], num_epochs=3)
-
-    # Check for "Best model saved!" 3 times
-    assert mock_print.call_args_list.count(call("Best model saved!")) == 3
-
-    # Reset for next test
-    mock_print.reset_mock()
-    pretrainer.train_losses = []
-
-    # Reset best validation loss for the second test
-    pretrainer.best_val_loss = float('inf')
-
-    # Test fluctuating validation loss
-    with patch.object(Pretrainer, '_process_batch', return_value=mock_batch_loss), \
-        patch.object(Pretrainer, '_validate', side_effect=[3.0, 4.0, 2.0]), \
-        patch.object(Pretrainer, 'save_losses'):
-        pretrainer.train(['path/to/dataset.parquet'], num_epochs=3)
-
-    # Should print "Best model saved!" only on first and third epochs
-    assert mock_print.call_args_list.count(call("Best model saved!")) == 2
 
 
 @patch('aiia.pretrain.pretrainer.Pretrainer._process_batch')