""" Data processing utilities for Ruby method datasets. This module provides functions to load, preprocess, and prepare Ruby method data for GNN training. Includes custom Dataset class for AST to graph conversion. """ import json import random import os import logging from pathlib import Path from typing import List, Dict, Any, Tuple, Optional, Union try: import torch from torch_geometric.data import Data TORCH_AVAILABLE = True except ImportError: TORCH_AVAILABLE = False def load_methods_json(filepath: str) -> List[Dict[str, Any]]: """ Load Ruby methods from JSON file. Args: filepath: Path to the JSON file containing method data Returns: List of method dictionaries """ with open(filepath, 'r') as f: return json.load(f) def methods_to_dataframe(methods: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """ Convert list of method dictionaries to a structured format. Args: methods: List of method dictionaries Returns: List of method dictionaries (pass-through for compatibility) """ return methods def filter_methods_by_length(methods: List[Dict[str, Any]], min_lines: int = 5, max_lines: int = 100) -> List[Dict[str, Any]]: """ Filter methods by source code length. Args: methods: List of method dictionaries min_lines: Minimum number of lines max_lines: Maximum number of lines Returns: Filtered list of methods """ filtered = [] for method in methods: if 'raw_source' in method: line_count = len(method['raw_source'].split('\n')) if min_lines <= line_count <= max_lines: method['line_count'] = line_count filtered.append(method) return filtered """ Filter methods by source code length. Args: df: DataFrame containing method data min_lines: Minimum number of lines max_lines: Maximum number of lines Returns: Filtered DataFrame """ df['line_count'] = df['raw_source'].apply(lambda x: len(x.split('\n'))) return df[(df['line_count'] >= min_lines) & (df['line_count'] <= max_lines)] class ASTNodeEncoder: """ Encoder for mapping AST node types to feature vectors. This class maintains a vocabulary of AST node types found in Ruby code and maps them to dense feature vectors for GNN processing. """ def __init__(self): """Initialize the node encoder with common Ruby AST node types.""" # Common Ruby AST node types based on the parser gem self.node_types = [ 'def', 'defs', 'args', 'arg', 'begin', 'end', 'lvasgn', 'ivasgn', 'gvasgn', 'cvasgn', 'send', 'block', 'if', 'unless', 'while', 'until', 'for', 'case', 'when', 'rescue', 'ensure', 'retry', 'break', 'next', 'redo', 'return', 'yield', 'super', 'zsuper', 'lambda', 'proc', 'and', 'or', 'not', 'true', 'false', 'nil', 'self', 'int', 'float', 'str', 'sym', 'regexp', 'array', 'hash', 'pair', 'splat', 'kwsplat', 'block_pass', 'const', 'cbase', 'lvar', 'ivar', 'gvar', 'cvar', 'casgn', 'masgn', 'mlhs', 'op_asgn', 'and_asgn', 'or_asgn', 'back_ref', 'nth_ref', 'class', 'sclass', 'module', 'defined?', 'alias', 'undef', 'range', 'irange', 'erange', 'regopt' ] # Create mapping from node type to index self.type_to_idx = {node_type: idx for idx, node_type in enumerate(self.node_types)} self.unknown_idx = len(self.node_types) # Index for unknown node types self.vocab_size = len(self.node_types) + 1 # +1 for unknown def encode_node_type(self, node_type: str) -> int: """ Encode a node type to its integer index. Args: node_type: The AST node type string Returns: Integer index for the node type """ return self.type_to_idx.get(node_type, self.unknown_idx) def create_node_features(self, node_type: str) -> List[float]: """ Create feature vector for a node type. Args: node_type: The AST node type string Returns: Feature vector as list of floats """ # Simple one-hot encoding for now features = [0.0] * self.vocab_size idx = self.encode_node_type(node_type) features[idx] = 1.0 return features class ASTGraphConverter: """ Converter for transforming AST JSON to graph representation. This class parses the AST JSON structure and converts it into a graph format suitable for GNN processing. """ def __init__(self): """Initialize the AST to graph converter.""" self.node_encoder = ASTNodeEncoder() self.reset() def reset(self): """Reset the converter state for processing a new AST.""" self.nodes = [] # List of node features self.edges = [] # List of edge tuples (parent_idx, child_idx) self.edge_attrs = [] # List of edge attributes [child_index, depth, num_siblings] self.node_depths = [] # Depth of each node in the tree self.node_child_indices = [] # Position of each node among its siblings self.node_count = 0 def parse_ast_json(self, ast_json: str) -> Dict[str, Any]: """ Parse AST JSON string and convert to graph representation. Args: ast_json: JSON string representing the AST Returns: Dictionary containing node features, edge indices, and edge attributes. edge_attr contains [child_index, depth, num_siblings] per edge. node_pos contains [child_index, depth] per node for positional encoding. """ self.reset() try: ast_data = json.loads(ast_json) self._process_node(ast_data, parent_idx=None, depth=0, child_index=0, num_siblings=1) # Convert to appropriate format if not self.nodes: # Handle empty AST case node_features = [[0.0] * self.node_encoder.vocab_size] edge_index = [[], []] # Empty edge list edge_attr = [] node_pos = [[0, 0]] else: node_features = self.nodes if self.edges: # Transpose edge list to [2, num_edges] format edge_index = [[], []] for parent, child in self.edges: edge_index[0].append(parent) edge_index[1].append(child) else: edge_index = [[], []] edge_attr = self.edge_attrs node_pos = list(zip(self.node_child_indices, self.node_depths)) return { 'x': node_features, 'edge_index': edge_index, 'edge_attr': edge_attr, 'node_pos': node_pos, 'num_nodes': len(self.nodes) if self.nodes else 1 } except (json.JSONDecodeError, Exception): # Handle malformed JSON or other errors gracefully return { 'x': [[0.0] * self.node_encoder.vocab_size], 'edge_index': [[], []], 'edge_attr': [], 'node_pos': [[0, 0]], 'num_nodes': 1 } def _process_node(self, node: Union[Dict, List, str, int, float, None], parent_idx: Optional[int] = None, depth: int = 0, child_index: int = 0, num_siblings: int = 1) -> int: """ Recursively process an AST node and its children. Args: node: The AST node (dict, list, or primitive) parent_idx: Index of the parent node depth: Depth of the current node in the AST child_index: Position of this node among its siblings (0-based) num_siblings: Total number of siblings (including this node) Returns: Index of the current node """ if isinstance(node, dict) and 'type' in node: # This is an AST node with a type node_type = node['type'] current_idx = self.node_count self.node_count += 1 # Create node features features = self.node_encoder.create_node_features(node_type) self.nodes.append(features) self.node_depths.append(depth) self.node_child_indices.append(child_index) # Add edge from parent to current node if parent_idx is not None: self.edges.append((parent_idx, current_idx)) self.edge_attrs.append([child_index, depth, num_siblings]) # Process children with positional information if 'children' in node: children = node['children'] n_children = len(children) for i, child in enumerate(children): self._process_node(child, current_idx, depth=depth + 1, child_index=i, num_siblings=n_children) return current_idx elif isinstance(node, list): # Process list of nodes n_items = len(node) for i, child in enumerate(node): self._process_node(child, parent_idx, depth=depth, child_index=i, num_siblings=n_items) return parent_idx if parent_idx is not None else -1 else: # Leaf node (string, int, float, None) if parent_idx is not None: current_idx = self.node_count self.node_count += 1 # Create a generic leaf node leaf_type = 'leaf_' + type(node).__name__ features = self.node_encoder.create_node_features(leaf_type) self.nodes.append(features) self.node_depths.append(depth) self.node_child_indices.append(child_index) # Add edge from parent to leaf self.edges.append((parent_idx, current_idx)) self.edge_attrs.append([child_index, depth, num_siblings]) return current_idx return -1 def load_jsonl_file(filepath: str, limit: Optional[int] = None) -> List[Dict[str, Any]]: """ Load data from a JSONL file. Args: filepath: Path to the JSONL file limit: Optional maximum number of lines to load. Returns: List of dictionaries from the JSONL file """ data = [] with open(filepath, 'r', encoding='utf-8') as f: for i, line in enumerate(f): if limit is not None and i >= limit: break line = line.strip() if line: try: data.append(json.loads(line)) except json.JSONDecodeError: continue # Skip malformed lines return data class RubyASTDataset: """ Dataset class for loading Ruby AST data and converting to graph format. This class loads JSONL files containing Ruby method data and converts the AST representations to graph objects suitable for GNN training. """ def __init__(self, jsonl_path: str, transform=None, limit: Optional[int] = None): """ Initialize the dataset. Args: jsonl_path: Path to the JSONL file containing method data transform: Optional transform to apply to each sample limit: Optional maximum number of samples to load. """ self.jsonl_path = jsonl_path self.transform = transform self.converter = ASTGraphConverter() # Load the data self.data = load_jsonl_file(jsonl_path, limit=limit) print(f"Loaded {len(self.data)} samples from {jsonl_path}") def __len__(self) -> int: """Return the number of samples in the dataset.""" return len(self.data) def __getitem__(self, idx: int) -> Dict[str, Any]: """ Get a sample from the dataset. Args: idx: Index of the sample Returns: Dictionary containing graph data and target """ if idx < 0 or idx >= len(self.data): raise IndexError(f"Index {idx} out of range for dataset of size {len(self.data)}") sample = self.data[idx] # Convert AST to graph graph_data = self.converter.parse_ast_json(sample['ast_json']) # Create the data object result = { 'x': graph_data['x'], 'edge_index': graph_data['edge_index'], 'y': [sample.get('complexity_score', 5.0)], # Default complexity score if missing 'num_nodes': graph_data['num_nodes'], 'id': sample.get('id', f'sample_{idx}'), 'repo_name': sample.get('repo_name', ''), 'file_path': sample.get('file_path', '') } # Apply transform if provided if self.transform: result = self.transform(result) return result def get_feature_dim(self) -> int: """Return the dimension of node features.""" return self.converter.node_encoder.vocab_size def collate_graphs(batch: List[Dict[str, Any]]) -> Dict[str, Any]: """ Collate function for batching graph data. Args: batch: List of graph data dictionaries Returns: Batched graph data """ if not batch: raise ValueError("Cannot collate empty batch") # Collect all node features and edge indices all_x = [] all_edge_index = [[], []] # [source_nodes, target_nodes] all_y = [] batch_idx = [] node_offset = 0 metadata = { 'ids': [], 'repo_names': [], 'file_paths': [] } for i, sample in enumerate(batch): # Node features all_x.extend(sample['x']) # Edge indices (offset by current node count) edges = sample['edge_index'] if len(edges[0]) > 0: # Only offset if there are edges for j in range(len(edges[0])): all_edge_index[0].append(edges[0][j] + node_offset) all_edge_index[1].append(edges[1][j] + node_offset) # Target values all_y.extend(sample['y']) # Batch indices for each node num_nodes = sample['num_nodes'] batch_idx.extend([i] * num_nodes) node_offset += num_nodes # Metadata metadata['ids'].append(sample['id']) metadata['repo_names'].append(sample['repo_name']) metadata['file_paths'].append(sample['file_path']) return { 'x': all_x, 'edge_index': all_edge_index, 'y': all_y, 'batch': batch_idx, 'num_graphs': len(batch), 'metadata': metadata } class SimpleDataLoader: """ Simple DataLoader implementation for batching data. This provides a basic implementation that can be used when PyTorch DataLoader is not available, and can easily be replaced with the real PyTorch DataLoader when dependencies are installed. """ def __init__(self, dataset, batch_size: int = 1, shuffle: bool = False, collate_fn=None): """ Initialize the DataLoader. Args: dataset: Dataset to load from batch_size: Number of samples per batch shuffle: Whether to shuffle the data collate_fn: Function to collate samples into batches """ self.dataset = dataset self.batch_size = batch_size self.shuffle = shuffle self.collate_fn = collate_fn or collate_graphs # Create indices self.indices = list(range(len(dataset))) if shuffle: import random random.shuffle(self.indices) def __len__(self) -> int: """Return number of batches.""" return (len(self.dataset) + self.batch_size - 1) // self.batch_size def __iter__(self): """Iterate over batches.""" for i in range(0, len(self.dataset), self.batch_size): batch_indices = self.indices[i:i + self.batch_size] batch = [self.dataset[idx] for idx in batch_indices] yield self.collate_fn(batch) class PairedDataset: """ Dataset class for loading paired Ruby AST and text description data. This class loads the paired_data.jsonl file containing Ruby method data and converts AST representations to graph objects paired with text descriptions. For each method, it randomly samples one description from the available descriptions. """ def __init__(self, jsonl_path: str, transform=None, seed: Optional[int] = None, limit: Optional[int] = None): """ Initialize the paired dataset. Args: jsonl_path: Path to the paired_data.jsonl file transform: Optional transform to apply to each sample seed: Random seed for consistent description sampling limit: Optional maximum number of samples to load. """ self.jsonl_path = jsonl_path self.transform = transform self.converter = ASTGraphConverter() if seed is not None: random.seed(seed) # Load the data self.data = load_jsonl_file(jsonl_path, limit=limit) print(f"Loaded {len(self.data)} samples from {jsonl_path}") def __len__(self) -> int: """Return the number of samples in the dataset.""" return len(self.data) def __getitem__(self, idx: int) -> Tuple[Dict[str, Any], str]: """ Get a sample from the dataset. Args: idx: Index of the sample Returns: Tuple of (graph_data, text_description) """ if idx < 0 or idx >= len(self.data): raise IndexError(f"Index {idx} out of range for dataset of size {len(self.data)}") sample = self.data[idx] # Convert AST to graph graph_data = self.converter.parse_ast_json(sample['ast_json']) # Randomly sample one description descriptions = sample.get('descriptions', []) if descriptions: description = random.choice(descriptions) text_description = description['text'] else: # Fallback to method name if no descriptions available text_description = sample.get('method_name', 'unknown_method') # Create the graph data object graph_result = { 'x': graph_data['x'], 'edge_index': graph_data['edge_index'], 'num_nodes': graph_data['num_nodes'], 'id': sample.get('id', f'sample_{idx}'), 'repo_name': sample.get('repo_name', ''), 'file_path': sample.get('file_path', '') } # Apply transform if provided if self.transform: graph_result = self.transform(graph_result) return graph_result, text_description def get_feature_dim(self) -> int: """Return the dimension of node features.""" return self.converter.node_encoder.vocab_size def collate_paired_data(batch: List[Tuple[Dict[str, Any], str]]) -> Tuple[Dict[str, Any], List[str]]: """ Collate function for batching paired graph and text data. Args: batch: List of (graph_data, text_description) tuples Returns: Tuple of (batched_graph_data, list_of_text_descriptions) """ if not batch: raise ValueError("Cannot collate empty batch") # Separate graph data and text descriptions graph_batch = [item[0] for item in batch] text_batch = [item[1] for item in batch] # Collate graph data manually (similar to collate_graphs but without 'y' field) all_x = [] all_edge_index = [[], []] # [source_nodes, target_nodes] batch_idx = [] node_offset = 0 metadata = { 'ids': [], 'repo_names': [], 'file_paths': [] } for i, sample in enumerate(graph_batch): # Node features all_x.extend(sample['x']) # Edge indices (offset by current node count) edges = sample['edge_index'] if len(edges[0]) > 0: # Only offset if there are edges for j in range(len(edges[0])): all_edge_index[0].append(edges[0][j] + node_offset) all_edge_index[1].append(edges[1][j] + node_offset) # Batch indices for each node num_nodes = sample['num_nodes'] batch_idx.extend([i] * num_nodes) node_offset += num_nodes # Metadata metadata['ids'].append(sample['id']) metadata['repo_names'].append(sample['repo_name']) metadata['file_paths'].append(sample['file_path']) batched_graphs = { 'x': all_x, 'edge_index': all_edge_index, 'batch': batch_idx, 'num_graphs': len(batch), 'metadata': metadata } return batched_graphs, text_batch class PairedDataLoader: """ DataLoader for paired graph and text data. Extends SimpleDataLoader to handle paired (graph, text) data. """ def __init__(self, dataset, batch_size: int = 1, shuffle: bool = False): """ Initialize the PairedDataLoader. Args: dataset: PairedDataset to load from batch_size: Number of samples per batch shuffle: Whether to shuffle the data """ self.dataset = dataset self.batch_size = batch_size self.shuffle = shuffle # Create indices self.indices = list(range(len(dataset))) if shuffle: random.shuffle(self.indices) def __len__(self) -> int: """Return number of batches.""" return (len(self.dataset) + self.batch_size - 1) // self.batch_size def __iter__(self): """Iterate over batches.""" for i in range(0, len(self.dataset), self.batch_size): batch_indices = self.indices[i:i + self.batch_size] batch = [self.dataset[idx] for idx in batch_indices] yield collate_paired_data(batch) class PrecomputedRubyASTDataset: """ Dataset class for loading precomputed Ruby AST graph data. This class can load .pt files containing pre-converted PyTorch Geometric Data objects for speed, but also supports processing .jsonl files as a fallback. """ def __init__(self, path: str, transform=None): """ Initialize the dataset. Args: path: Path to the .pt or .jsonl file containing graph data. transform: Optional transform to apply to each sample. """ self.path = path self.transform = transform if not TORCH_AVAILABLE: raise ImportError("PyTorch and PyG are required for this dataset.") if path.endswith('.pt'): # Load the precomputed data into RAM self.data = torch.load(path, weights_only=False) print(f"Loaded {len(self.data)} precomputed graphs from {path}") elif path.endswith('.jsonl'): print(f"Processing JSONL file into graphs: {path}") jsonl_data = load_jsonl_file(path) converter = ASTGraphConverter() self.data = [] for sample in jsonl_data: graph_data = converter.parse_ast_json(sample['ast_json']) x = torch.tensor(graph_data['x'], dtype=torch.float) edge_index = torch.tensor(graph_data['edge_index'], dtype=torch.long) y = torch.tensor([sample.get('complexity_score', 5.0)], dtype=torch.float) data_obj = Data(x=x, edge_index=edge_index, y=y) # Add positional attributes — always set so PyG collation is consistent ea = graph_data.get('edge_attr', []) data_obj.edge_attr = torch.tensor( ea if ea else [], dtype=torch.float, ).reshape(-1, 3) if ea else torch.zeros((0, 3), dtype=torch.float) np_ = graph_data.get('node_pos', []) data_obj.node_pos = torch.tensor( np_ if np_ else [[0, 0]], dtype=torch.float, ) self.data.append(data_obj) print(f"Converted {len(self.data)} graphs from {path}") else: raise ValueError(f"Unsupported file type: {path}. Please provide a .pt or .jsonl file.") def __len__(self) -> int: """Return the number of samples in the dataset.""" return len(self.data) def __getitem__(self, idx: int): """ Get a sample from the dataset. Args: idx: Index of the sample Returns: PyTorch Geometric Data object """ if idx < 0 or idx >= len(self.data): raise IndexError(f"Index {idx} out of range for dataset of size {len(self.data)}") sample = self.data[idx] if self.transform: sample = self.transform(sample) return sample class PreCollatedDataset: """ Dataset class for loading pre-collated batches of graph data. This class loads a .pt file where each item is an already-collated `torch_geometric.data.Batch` object. This is the most efficient way to load data as it eliminates all real-time collation overhead. """ def __init__(self, pt_path: str): """ Initialize the dataset. Args: pt_path: Path to the .pt file containing pre-collated batches. """ # Load the list of pre-collated batches into RAM self.batches = torch.load(pt_path, weights_only=False) print(f"Loaded {len(self.batches)} pre-collated batches from {pt_path}") def __len__(self): return len(self.batches) def __getitem__(self, idx): return self.batches[idx] def create_data_loaders(train_path: str, val_path: str, batch_size: int = 32, shuffle: bool = True, num_workers: Optional[int] = None, pre_collated: bool = False): """ Create train and validation data loaders. Supports two modes: 1. Standard loading from a dataset of individual graphs (`pre_collated=False`). This uses a PyG DataLoader to perform real-time batching. 2. Pre-collated loading from a dataset of pre-batched graphs (`pre_collated=True`). This is the most performant option, as it has near-zero CPU overhead. Args: train_path: Path to training .pt file. val_path: Path to validation .pt file. batch_size: Batch size (used only if `pre_collated=False`). shuffle: Whether to shuffle training data. num_workers: Number of workers for data loading (used only if `pre_collated=False`). pre_collated: Whether the dataset files contain pre-collated batches. Returns: Tuple of (train_loader, val_loader) """ if not TORCH_AVAILABLE: raise ImportError("PyTorch is required to create data loaders.") if pre_collated: # --- Pre-collated path (most efficient) --- train_dataset = PreCollatedDataset(train_path) val_dataset = PreCollatedDataset(val_path) # The collate_fn simply returns the already-collated batch. # The input `batch` is a list of size 1 containing our pre-made Batch object. collate_fn = lambda x: x[0] # DataLoader is just a simple iterator here, no real collation work. # num_workers > 0 can actually be slower due to overhead of sending # already-large batches between processes. from torch.utils.data import DataLoader train_loader = DataLoader(train_dataset, batch_size=1, shuffle=shuffle, num_workers=0, collate_fn=collate_fn) val_loader = DataLoader(val_dataset, batch_size=1, shuffle=False, num_workers=0, collate_fn=collate_fn) print("✅ Using pre-collated data loader (maximum performance).") else: # --- Standard real-time collation path --- from torch_geometric.loader import DataLoader train_dataset = PrecomputedRubyASTDataset(train_path) val_dataset = PrecomputedRubyASTDataset(val_path) if num_workers is None: num_workers = os.cpu_count() train_loader = DataLoader( train_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=torch.cuda.is_available(), persistent_workers=num_workers > 0 ) val_loader = DataLoader( val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=torch.cuda.is_available(), persistent_workers=num_workers > 0 ) print(f"✅ Using standard PyG DataLoader with {num_workers} workers.") return train_loader, val_loader def create_paired_data_loaders(paired_data_path: str, batch_size: int = 32, shuffle: bool = True, seed: Optional[int] = None): """ Create data loader for paired graph and text data. Args: paired_data_path: Path to paired_data.jsonl file batch_size: Batch size for the loader shuffle: Whether to shuffle the data seed: Random seed for consistent description sampling Returns: PairedDataLoader instance """ dataset = PairedDataset(paired_data_path, seed=seed) loader = PairedDataLoader(dataset, batch_size=batch_size, shuffle=shuffle) return loader class AutoregressiveASTDataset: """ Dataset class for autoregressive AST generation training. This class loads paired Ruby AST and text description data and converts each AST into a sequence of (partial_graph, target_node) pairs for autoregressive training. Each method generates multiple training examples. """ def __init__(self, paired_data_path: str, max_sequence_length: int = 50, seed: Optional[int] = None, precomputed_embeddings_path: Optional[str] = None): """ Initialize the autoregressive dataset. Args: paired_data_path: Path to the paired_data.jsonl file max_sequence_length: Maximum number of nodes per sequence seed: Random seed for consistent description sampling precomputed_embeddings_path: Path to pre-computed text embeddings file (optional) """ self.paired_data_path = paired_data_path self.max_sequence_length = max_sequence_length self.converter = ASTGraphConverter() if seed is not None: random.seed(seed) # Load pre-computed embeddings if available self.precomputed_embeddings = {} if precomputed_embeddings_path and os.path.exists(precomputed_embeddings_path): try: if TORCH_AVAILABLE: self.precomputed_embeddings = torch.load(precomputed_embeddings_path, map_location='cpu', weights_only=True) print(f"✅ Loaded {len(self.precomputed_embeddings)} pre-computed text embeddings") else: print("⚠️ PyTorch not available, skipping pre-computed embeddings") except Exception as e: print(f"⚠️ Warning: Could not load pre-computed embeddings: {e}") elif precomputed_embeddings_path: print(f"⚠️ Warning: Pre-computed embeddings file not found: {precomputed_embeddings_path}") # Load the paired data self.paired_data = load_jsonl_file(paired_data_path) # Generate sequential training pairs from all methods self.sequential_pairs = [] self._generate_all_sequential_pairs() print(f"Loaded {len(self.paired_data)} methods from {paired_data_path}") print(f"Generated {len(self.sequential_pairs)} sequential training pairs") def _generate_all_sequential_pairs(self): """Generate sequential training pairs from all ASTs in the dataset.""" for sample in self.paired_data: try: # Get text description descriptions = sample.get('descriptions', []) if descriptions: description = random.choice(descriptions) text_description = description['text'] else: # Fallback to method name if no descriptions available text_description = sample.get('method_name', 'unknown_method') # Create sequential pairs for this AST sequential_pairs = self._create_sequential_pairs( sample['ast_json'], text_description ) # Add to global list self.sequential_pairs.extend(sequential_pairs) except Exception as e: # Skip malformed samples gracefully print(f"Warning: Skipping sample {sample.get('id', 'unknown')} due to error: {e}") continue def _create_sequential_pairs(self, ast_json: str, text_description: str) -> List[Dict[str, Any]]: """ Convert single AST into sequence of (partial_graph, target_node) pairs. Args: ast_json: JSON string representing the AST text_description: Text description for this method Returns: List of sequential training pairs """ pairs = [] try: # Extract nodes in proper order along with their connections nodes, connections = self._extract_nodes_and_connections_in_order(ast_json) # Limit sequence length if needed if len(nodes) > self.max_sequence_length: nodes = nodes[:self.max_sequence_length] # Also limit connections to only include those within the sequence filtered_connections = [] for src, tgt in connections: if src < self.max_sequence_length and tgt < self.max_sequence_length: filtered_connections.append((src, tgt)) connections = filtered_connections # Get pre-computed text embedding if available, otherwise store text text_embedding = None if text_description in self.precomputed_embeddings: text_embedding = self.precomputed_embeddings[text_description] # Create sequential pairs for i in range(len(nodes)): # Build partial graph with nodes 0 to i-1 partial_graph = self._build_partial_graph(nodes[:i]) # Target is the i-th node target_node = nodes[i] # Create target connections for this step # This represents which existing nodes (0 to i-1) the new node i should connect to target_connections = self._create_target_connections(i, connections) pair = { 'text_description': text_description, 'text_embedding': text_embedding, # Pre-computed embedding if available 'partial_graph': partial_graph, 'target_node': target_node, 'target_connections': target_connections, 'step': i, 'total_steps': len(nodes) } pairs.append(pair) except Exception as e: # Return empty list for malformed ASTs print(f"Warning: Failed to create sequential pairs: {e}") return pairs def _extract_nodes_and_connections_in_order(self, ast_json: str) -> Tuple[List[Dict[str, Any]], List[Tuple[int, int]]]: """ Extract nodes and their connections from AST in proper depth-first order. Args: ast_json: JSON string representing the AST Returns: Tuple of (nodes_list, connections_list) where connections are (parent_idx, child_idx) pairs """ try: ast_data = json.loads(ast_json) nodes = [] connections = [] self._traverse_ast_nodes_with_connections(ast_data, nodes, connections, parent_idx=None) return nodes, connections except (json.JSONDecodeError, Exception): # Return empty lists for malformed JSON return [], [] def _traverse_ast_nodes_with_connections(self, node: Union[Dict, List, str, int, float, None], nodes: List[Dict[str, Any]], connections: List[Tuple[int, int]], parent_idx: Optional[int] = None): """ Recursively traverse AST and collect nodes and connections in depth-first order. Args: node: Current AST node nodes: List to collect nodes connections: List to collect connections as (parent_idx, child_idx) pairs parent_idx: Index of parent node """ if isinstance(node, dict) and 'type' in node: # This is an AST node with a type current_idx = len(nodes) node_info = { 'node_type': node['type'], 'features': self.converter.node_encoder.create_node_features(node['type']), 'raw_node': node # Keep reference for debugging } nodes.append(node_info) # Add connection from parent to current node if parent_idx is not None: connections.append((parent_idx, current_idx)) # Traverse children if 'children' in node: for child in node['children']: self._traverse_ast_nodes_with_connections(child, nodes, connections, current_idx) elif isinstance(node, list): # Process list of nodes for child in node: self._traverse_ast_nodes_with_connections(child, nodes, connections, parent_idx) def _create_target_connections(self, node_idx: int, all_connections: List[Tuple[int, int]]) -> List[float]: """ Create target connection vector for a specific node being added. Args: node_idx: Index of the node being added to the graph all_connections: List of all connections in the full AST as (parent_idx, child_idx) pairs Returns: Binary vector of length max_nodes indicating which existing nodes to connect to """ # Initialize with zeros for all possible connections target_vector = [0.0] * 100 # max_nodes = 100 from model # Find all connections where this node is the target (child) # We want to know which existing nodes (with index < node_idx) should connect to this node for parent_idx, child_idx in all_connections: if child_idx == node_idx and parent_idx < node_idx and parent_idx < 100: target_vector[parent_idx] = 1.0 return target_vector def _traverse_ast_nodes(self, node: Union[Dict, List, str, int, float, None], nodes: List[Dict[str, Any]]): """ Recursively traverse AST and collect nodes in depth-first order. Args: node: Current AST node nodes: List to collect nodes """ if isinstance(node, dict) and 'type' in node: # This is an AST node with a type node_info = { 'node_type': node['type'], 'features': self.converter.node_encoder.create_node_features(node['type']), 'raw_node': node # Keep reference for debugging } nodes.append(node_info) # Traverse children if 'children' in node: for child in node['children']: self._traverse_ast_nodes(child, nodes) elif isinstance(node, list): # Process list of nodes for child in node: self._traverse_ast_nodes(child, nodes) def _build_partial_graph(self, nodes: List[Dict[str, Any]]) -> Dict[str, Any]: """ Build partial graph from first i nodes. Args: nodes: List of nodes to include in partial graph Returns: Partial graph representation """ if not nodes: # Empty graph case return { 'x': [], 'edge_index': [[], []], 'num_nodes': 0 } # Extract node features node_features = [node['features'] for node in nodes] # Create simple sequential connections (each node connects to next) # This is a simplified approach - in practice you'd want to preserve # the actual AST structure relationships edge_list = [] for i in range(len(nodes) - 1): edge_list.append([i, i + 1]) # Forward edge edge_list.append([i + 1, i]) # Backward edge for undirected if edge_list: edge_index = [[], []] for source, target in edge_list: edge_index[0].append(source) edge_index[1].append(target) else: edge_index = [[], []] return { 'x': node_features, 'edge_index': edge_index, 'num_nodes': len(nodes) } def __len__(self) -> int: """Return the number of sequential training pairs.""" return len(self.sequential_pairs) def __getitem__(self, idx: int) -> Dict[str, Any]: """ Get a sequential training pair. Args: idx: Index of the training pair Returns: Dictionary containing partial graph and target node data """ if idx < 0 or idx >= len(self.sequential_pairs): raise IndexError(f"Index {idx} out of range for dataset of size {len(self.sequential_pairs)}") return self.sequential_pairs[idx] def get_feature_dim(self) -> int: """Return the dimension of node features.""" return self.converter.node_encoder.vocab_size def collate_autoregressive_data(batch: List[Dict[str, Any]]) -> Dict[str, Any]: """ Collate function for batching autoregressive training data. Args: batch: List of sequential training pairs Returns: Batched autoregressive training data """ if not batch: raise ValueError("Cannot collate empty batch") # Separate different components text_descriptions = [item['text_description'] for item in batch] text_embeddings = [item.get('text_embedding') for item in batch] steps = [item['step'] for item in batch] total_steps = [item['total_steps'] for item in batch] # Collate partial graphs partial_graphs = [item['partial_graph'] for item in batch] # Collate node features from partial graphs all_x = [] all_edge_index = [[], []] batch_idx = [] node_offset = 0 for i, graph in enumerate(partial_graphs): # Node features if graph['x']: all_x.extend(graph['x']) # Edge indices (offset by current node count) edges = graph['edge_index'] if len(edges[0]) > 0: for j in range(len(edges[0])): all_edge_index[0].append(edges[0][j] + node_offset) all_edge_index[1].append(edges[1][j] + node_offset) # Batch indices for each node num_nodes = graph['num_nodes'] batch_idx.extend([i] * num_nodes) node_offset += num_nodes # Target nodes and connections target_nodes = [item['target_node'] for item in batch] target_node_types = [node['node_type'] for node in target_nodes] target_node_features = [node['features'] for node in target_nodes] target_connections = [item['target_connections'] for item in batch] return { 'text_descriptions': text_descriptions, 'text_embeddings': text_embeddings, # Can contain None values if not pre-computed 'partial_graphs': { 'x': all_x, 'edge_index': all_edge_index, 'batch': batch_idx, 'num_graphs': len(batch) }, 'target_node_types': target_node_types, 'target_node_features': target_node_features, 'target_connections': target_connections, 'steps': steps, 'total_steps': total_steps } class AutoregressiveDataLoader: """ DataLoader for autoregressive AST training data. """ def __init__(self, dataset: AutoregressiveASTDataset, batch_size: int = 8, shuffle: bool = True): """ Initialize the AutoregressiveDataLoader. Args: dataset: AutoregressiveASTDataset to load from batch_size: Number of sequential pairs per batch shuffle: Whether to shuffle the data """ self.dataset = dataset self.batch_size = batch_size self.shuffle = shuffle # Create indices self.indices = list(range(len(dataset))) if shuffle: random.shuffle(self.indices) def __len__(self) -> int: """Return number of batches.""" return (len(self.dataset) + self.batch_size - 1) // self.batch_size def __iter__(self): """Iterate over batches.""" for i in range(0, len(self.dataset), self.batch_size): batch_indices = self.indices[i:i + self.batch_size] batch = [self.dataset[idx] for idx in batch_indices] yield collate_autoregressive_data(batch) def create_autoregressive_data_loader(paired_data_path: str, batch_size: int = 8, shuffle: bool = True, max_sequence_length: int = 50, seed: Optional[int] = None, precomputed_embeddings_path: Optional[str] = None, num_workers: Optional[int] = None, pin_memory: bool = True): """ Create data loader for autoregressive AST training. Args: paired_data_path: Path to paired_data.jsonl file batch_size: Number of sequential pairs per batch shuffle: Whether to shuffle the data max_sequence_length: Maximum sequence length per method seed: Random seed for consistent description sampling precomputed_embeddings_path: Path to pre-computed text embeddings file num_workers: Number of worker processes for data loading (defaults to CPU count) pin_memory: Whether to use pinned memory for faster GPU transfer Returns: DataLoader instance (PyTorch DataLoader if available, otherwise AutoregressiveDataLoader) """ dataset = AutoregressiveASTDataset( paired_data_path, max_sequence_length=max_sequence_length, seed=seed, precomputed_embeddings_path=precomputed_embeddings_path ) # Use PyTorch DataLoader if available for better performance if TORCH_AVAILABLE: import os if num_workers is None: num_workers = os.cpu_count() try: from torch.utils.data import DataLoader # Create PyTorch DataLoader with optimizations loader = DataLoader( dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=pin_memory and torch.cuda.is_available(), collate_fn=collate_autoregressive_data, persistent_workers=num_workers > 0, # Keep workers alive between epochs prefetch_factor=2 if num_workers > 0 else 2 # Prefetch batches ) print(f"✅ Using optimized PyTorch DataLoader with {num_workers} workers, pin_memory={pin_memory and torch.cuda.is_available()}") return loader except Exception as e: print(f"⚠️ Warning: Could not create PyTorch DataLoader ({e}), falling back to custom loader") # Fallback to custom loader loader = AutoregressiveDataLoader(dataset, batch_size=batch_size, shuffle=shuffle) print("ℹ️ Using custom AutoregressiveDataLoader") return loader class HierarchicalASTDataset(RubyASTDataset): """ Dataset for loading a single level of a hierarchical AST dataset. This class inherits from RubyASTDataset to reuse the same AST-to-graph conversion logic. It is used to load one of the `_level_N.jsonl` files. """ def __init__(self, jsonl_path: str, transform=None): """ Initialize the dataset for a specific AST level. Args: jsonl_path: Path to the JSONL file for a specific level. transform: Optional transform to apply to each sample. """ super().__init__(jsonl_path, transform) def create_hierarchical_data_loader(dataset_path: str, batch_size: int, shuffle: bool, num_workers: Optional[int] = None): """ Creates a data loader for a specific level of the hierarchical dataset. Args: dataset_path: The full path to the `_level_N.jsonl` file. batch_size: The batch size for the data loader. shuffle: Whether to shuffle the data. num_workers: The number of worker processes for data loading. Returns: A DataLoader instance for the specified dataset level. """ dataset = HierarchicalASTDataset(dataset_path) if TORCH_AVAILABLE: try: from torch_geometric.loader import DataLoader if num_workers is None: num_workers = os.cpu_count() loader = DataLoader( dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=torch.cuda.is_available(), persistent_workers=num_workers > 0, collate_fn=collate_graphs # Reusing the existing collate function ) logging.info(f"Created PyG DataLoader for {dataset_path} with {num_workers} workers.") return loader except ImportError: logging.warning("PyTorch Geometric not found. Falling back to SimpleDataLoader.") # Fallback to SimpleDataLoader return SimpleDataLoader(dataset, batch_size=batch_size, shuffle=shuffle, collate_fn=collate_graphs) class HierarchicalPairedDataset(PairedDataset): """ Dataset for loading a single level of a hierarchical dataset with paired text. This class inherits from PairedDataset to reuse the same logic for processing graph data and randomly sampling text descriptions. """ def __init__(self, jsonl_path: str, transform=None, seed: Optional[int] = None, limit: Optional[int] = None): """ Initialize the dataset for a specific AST level. Args: jsonl_path: Path to the JSONL file for a specific level (e.g., train_paired_data_level_0.jsonl). transform: Optional transform to apply to each sample. seed: Random seed for consistent description sampling. limit: Optional maximum number of samples to load. """ super().__init__(jsonl_path, transform, seed, limit) def create_hierarchical_paired_data_loader(dataset_path: str, batch_size: int, shuffle: bool, num_workers: Optional[int] = None, limit: Optional[int] = None): """ Creates a data loader for a specific level of the hierarchical paired dataset. Args: dataset_path: The full path to the `_level_N.jsonl` file. batch_size: The batch size for the data loader. shuffle: Whether to shuffle the data. num_workers: The number of worker processes for data loading. limit: Optional maximum number of samples to load. Returns: A DataLoader instance for the specified dataset level. """ dataset = HierarchicalPairedDataset(dataset_path, limit=limit) if TORCH_AVAILABLE: try: from torch.utils.data import DataLoader if num_workers is None: num_workers = 0 # Disabled for now to prevent file handle exhaustion loader = DataLoader( dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=torch.cuda.is_available(), persistent_workers=num_workers > 0, collate_fn=collate_paired_data ) logging.info(f"Created PyTorch DataLoader for {dataset_path} with {num_workers} workers.") return loader except (ImportError, Exception) as e: logging.warning(f"PyTorch DataLoader creation failed ({e}). Falling back to PairedDataLoader.") # Fallback to custom PairedDataLoader return PairedDataLoader(dataset, batch_size=batch_size, shuffle=shuffle)