bertose-affinose-training-code / code /training /multimodal_dataset.py
supanthadey1's picture
Add BERTose and AFFINose training code release
1d6f391 verified
Raw
History Blame Contribute Delete
24.5 kB
"""
Multimodal Glycan Dataset
Combines sequence (WURCS), MS, and 3D structure data for multimodal BERT training.
Handles optional modalities (MS and 3D structure).
"""
import torch
from torch.utils.data import Dataset
import pickle
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import numpy as np
class MultimodalGlycanDataset(Dataset):
"""
Dataset for multimodal glycan BERT training.
Combines:
- Sequence tokens (WURCS atomic tokenization)
- MS tokens (mass spectrometry peaks, RT, intensity)
- 3D structure tokens (VQ-VAE discrete tokens, 4 per residue)
Each modality can be enabled/disabled via flags.
"""
def __init__(
self,
sequences_path: str,
ms_tokens_path: str,
structure_data_path: Optional[str] = None,
max_seq_length: int = 512,
max_ms_length: int = 150,
max_mono_length: int = 50,
max_struct_tokens: int = 200,
max_atoms: int = 300,
include_ms: bool = True,
include_3d: bool = True,
):
"""
Initialize multimodal dataset.
Args:
sequences_path: Path to sequences.pkl (contains token_ids, residue_ids, has_ms, has_3d, monosaccharide_indices)
ms_tokens_path: Path to ms_tokens.pkl (contains MS token IDs per WURCS)
structure_data_path: Path to training_dataset.pkl (contains VQ-VAE tokens and attention masks)
max_seq_length: Maximum sequence length (truncate/pad)
max_ms_length: Maximum MS token length (truncate/pad)
max_mono_length: Maximum number of monosaccharides (truncate/pad)
max_struct_tokens: Maximum structural tokens (truncate/pad)
max_atoms: Maximum number of atoms (for cross-attention mask padding)
include_ms: Whether to include MS modality
include_3d: Whether to include 3D structure modality
"""
self.max_seq_length = max_seq_length
self.max_ms_length = max_ms_length
self.max_mono_length = max_mono_length
self.max_struct_tokens = max_struct_tokens
self.max_atoms = max_atoms
self.include_ms = include_ms
self.include_3d = include_3d
# Load sequences
print(f"Loading sequences from {sequences_path}...")
with open(sequences_path, 'rb') as f:
sequences_raw = pickle.load(f)
# Convert to list if it's a dict, but keep WURCS key
if isinstance(sequences_raw, dict):
self.sequences = []
for wurcs, seq_data in sequences_raw.items():
# Validate that seq_data is a dict with required fields
if not isinstance(seq_data, dict):
print(f"Warning: Skipping invalid entry for WURCS: {wurcs[:50]}...")
continue
if 'token_ids' not in seq_data:
print(f"Warning: Skipping entry without token_ids for WURCS: {wurcs[:50]}...")
continue
# Add WURCS key to the data if not present
if 'wurcs' not in seq_data:
seq_data['wurcs'] = wurcs
self.sequences.append(seq_data)
else:
self.sequences = sequences_raw
print(f" Loaded {len(self.sequences)} sequences")
# Load MS tokens
self.ms_tokens = {}
if self.include_ms:
print(f"Loading MS tokens from {ms_tokens_path}...")
with open(ms_tokens_path, 'rb') as f:
self.ms_tokens = pickle.load(f)
print(f" Loaded {len(self.ms_tokens)} MS token sets")
# Load 3D structure data
self.structure_data = {}
if self.include_3d and structure_data_path:
struct_path = Path(structure_data_path)
if struct_path.exists():
print(f"Loading 3D structure data from {structure_data_path}...")
with open(structure_data_path, 'rb') as f:
struct_pkl = pickle.load(f)
# Index by WURCS
if isinstance(struct_pkl, dict) and 'full_multimodal' in struct_pkl:
samples = struct_pkl['full_multimodal']
self.structure_data = {s['wurcs']: s for s in samples}
else:
self.structure_data = {s['wurcs']: s for s in struct_pkl}
print(f" Loaded {len(self.structure_data)} structure samples")
else:
print(f" Warning: Structure data file not found at {structure_data_path}")
print(f" Continuing without 3D structure modality...")
# Statistics
self._compute_stats()
def _compute_stats(self):
"""Compute dataset statistics."""
# Count actual data availability based on loaded dictionaries, not stored flags
ms_available = 0
struct_available = 0
for s in self.sequences:
wurcs = s.get('wurcs', '')
if wurcs in self.ms_tokens:
ms_available += 1
if wurcs in self.structure_data:
struct_available += 1
self.stats = {
'total': len(self.sequences),
'with_ms_available': ms_available,
'with_3d_available': struct_available,
'with_ms_tokens': len(self.ms_tokens),
'with_structure_tokens': len(self.structure_data),
}
print(f"\nDataset Statistics:")
print(f" Total sequences: {self.stats['total']:,}")
print(f" With MS data: {self.stats['with_ms_available']:,} ({100*self.stats['with_ms_available']/self.stats['total']:.2f}%)")
print(f" With 3D data: {self.stats['with_3d_available']:,} ({100*self.stats['with_3d_available']/self.stats['total']:.2f}%)")
print(f" Include MS: {self.include_ms}")
print(f" Include 3D: {self.include_3d}")
print()
def __len__(self) -> int:
return len(self.sequences)
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
"""
Get a single multimodal sample.
Returns:
Dictionary containing:
- seq_token_ids: Sequence token IDs (padded/truncated)
- seq_attention_mask: Sequence attention mask
- seq_residue_ids: Residue position IDs for sequence tokens
- ms_token_ids: MS token IDs (padded/truncated, or empty if no MS)
- ms_attention_mask: MS attention mask
- ms_residue_ids: Residue IDs for MS tokens (all -2 for global)
- mono_indices: Monosaccharide indices (padded/truncated)
- mono_residue_ids: Residue IDs for each monosaccharide
- has_ms: Whether this sample has MS data
- has_3d: Whether this sample has 3D data (future)
- has_residue_error: Whether this sample has [RESIDUE_ERROR] tokens
"""
seq_data = self.sequences[idx]
wurcs = seq_data['wurcs']
# ===== Sequence Modality =====
seq_token_ids = seq_data['token_ids']
seq_residue_ids = seq_data.get('residue_ids', [-1] * len(seq_token_ids))
# NEW: Extract branch depths and linkage types for tree-aware encoding
seq_branch_depths = seq_data.get('branch_depths', [0] * len(seq_token_ids))
seq_linkage_types = seq_data.get('linkage_types', [0] * len(seq_token_ids))
# Truncate/pad sequence
if len(seq_token_ids) > self.max_seq_length:
seq_token_ids = seq_token_ids[:self.max_seq_length]
seq_residue_ids = seq_residue_ids[:self.max_seq_length]
seq_branch_depths = seq_branch_depths[:self.max_seq_length]
seq_linkage_types = seq_linkage_types[:self.max_seq_length]
seq_len = len(seq_token_ids)
seq_attention_mask = [1] * seq_len
# Pad to max length
padding_len = self.max_seq_length - seq_len
seq_token_ids = seq_token_ids + [0] * padding_len # 0 = [PAD]
seq_residue_ids = seq_residue_ids + [-1] * padding_len
seq_branch_depths = seq_branch_depths + [0] * padding_len
seq_linkage_types = seq_linkage_types + [0] * padding_len
# Pad attention mask
seq_attention_mask = seq_attention_mask + [0] * padding_len
# ===== Topology (Distance Matrix) =====
dist_labels = seq_data.get('distance_matrix', None)
if dist_labels is not None:
# Convert to tensor and pad
# dist_labels is list of lists
# 1. Pad rows (already done in tokenizer? assume yes, but re-checking)
# Tokenizer guarantees square matrix of size `length`. We need to pad to `max_seq_length`.
# Create full -1 matrix
padded_dist = [[-1] * self.max_seq_length for _ in range(self.max_seq_length)]
# Fill in valid part
current_len = len(dist_labels) # This is the valid length
# Truncate if too long (unlikely due to tokenizer limit)
trunc_len = min(current_len, self.max_seq_length)
for i in range(trunc_len):
row = dist_labels[i]
valid_row_len = min(len(row), self.max_seq_length)
for j in range(valid_row_len):
padded_dist[i][j] = row[j]
dist_labels = torch.tensor(padded_dist, dtype=torch.float)
else:
# Should not happen if data is regenerated, but fail safe
dist_labels = torch.full((self.max_seq_length, self.max_seq_length), -1.0)
# ===== MS Modality =====
has_ms = False
ms_token_ids = []
ms_residue_ids = []
ms_attention_mask = []
if self.include_ms and wurcs in self.ms_tokens:
has_ms = True
ms_data = self.ms_tokens[wurcs]
# Handle different data formats
if isinstance(ms_data, dict) and 'ms_token_ids' in ms_data:
ms_token_ids = ms_data['ms_token_ids']
elif isinstance(ms_data, str):
# If ms_data is a string (token sequence), skip it
has_ms = False
ms_token_ids = []
elif isinstance(ms_data, list):
# If ms_data is directly a list of token IDs
ms_token_ids = ms_data
else:
# Unknown format, skip
has_ms = False
ms_token_ids = []
# Ensure ms_token_ids is a list of integers
if not isinstance(ms_token_ids, list):
has_ms = False
ms_token_ids = []
elif len(ms_token_ids) > 0 and isinstance(ms_token_ids[0], str):
# If list contains strings, skip this entry
has_ms = False
ms_token_ids = []
# Truncate/pad MS tokens
if has_ms and len(ms_token_ids) > 0:
if len(ms_token_ids) > self.max_ms_length:
ms_token_ids = ms_token_ids[:self.max_ms_length]
ms_len = len(ms_token_ids)
ms_attention_mask = [1] * ms_len
# MS tokens are global (apply to whole glycan), so residue_id = -2
ms_residue_ids = [-2] * ms_len
# Pad to max length
padding_len = self.max_ms_length - ms_len
ms_token_ids = ms_token_ids + [0] * padding_len
ms_residue_ids = ms_residue_ids + [-1] * padding_len
ms_attention_mask = ms_attention_mask + [0] * padding_len
# Ensure MS tensors are always properly sized (handles invalid/missing MS data)
if len(ms_token_ids) != self.max_ms_length:
has_ms = False
ms_token_ids = [0] * self.max_ms_length
ms_residue_ids = [-1] * self.max_ms_length
ms_attention_mask = [0] * self.max_ms_length
# ===== Monosaccharide Indices =====
mono_indices = seq_data.get('monosaccharide_indices', [])
mono_residue_ids = seq_data.get('monosaccharide_residue_ids', [])
# Validate mono_indices format
if not isinstance(mono_indices, list):
mono_indices = []
mono_residue_ids = []
elif len(mono_indices) > 0:
# Check if elements are valid integers or can be converted
validated_indices = []
validated_residue_ids = []
for i, idx in enumerate(mono_indices):
if isinstance(idx, (int, np.integer)):
validated_indices.append(int(idx))
if i < len(mono_residue_ids) and isinstance(mono_residue_ids[i], (int, np.integer)):
validated_residue_ids.append(int(mono_residue_ids[i]))
else:
validated_residue_ids.append(-1)
elif isinstance(idx, str):
# Try to convert string to int
try:
validated_indices.append(int(idx))
if i < len(mono_residue_ids):
try:
validated_residue_ids.append(int(mono_residue_ids[i]))
except (ValueError, TypeError):
validated_residue_ids.append(-1)
else:
validated_residue_ids.append(-1)
except (ValueError, TypeError):
# Skip invalid entries
continue
# Skip non-convertible types
mono_indices = validated_indices
mono_residue_ids = validated_residue_ids
# Truncate/pad monosaccharide indices
if len(mono_indices) > self.max_mono_length:
mono_indices = mono_indices[:self.max_mono_length]
mono_residue_ids = mono_residue_ids[:self.max_mono_length]
mono_len = len(mono_indices)
padding_len = self.max_mono_length - mono_len
mono_indices = mono_indices + [0] * padding_len # 0 = <PAD>
mono_residue_ids = mono_residue_ids + [-1] * padding_len
# ===== 3D Structure Modality =====
has_3d = False
struct_token_ids = []
struct_attention_mask = []
struct_residue_ids = []
if self.include_3d and wurcs in self.structure_data:
has_3d = True
struct_sample = self.structure_data[wurcs]
# Get WURCS-to-GraphML residue mapping for cross-attention alignment
# This maps WURCS residue IDs to GraphML residue indices
wurcs_to_graphml = struct_sample.get('wurcs_to_graphml_mapping', {})
# Create reverse mapping: graphml_idx -> wurcs_residue_id
graphml_to_wurcs = {v: k for k, v in wurcs_to_graphml.items()}
# Flatten VQ-VAE tokens (4 tokens per residue)
struct_tokens_per_residue = struct_sample['structural_tokens_per_residue']
for graphml_idx, residue_tokens in enumerate(struct_tokens_per_residue):
struct_token_ids.extend(residue_tokens)
# Map GraphML index to WURCS residue ID for cross-attention
# Use -1 for unmapped residues (e.g., ROH reducing end)
wurcs_res_id = graphml_to_wurcs.get(graphml_idx, -1)
struct_residue_ids.extend([wurcs_res_id] * len(residue_tokens))
# Truncate/pad structural tokens
if len(struct_token_ids) > self.max_struct_tokens:
struct_token_ids = struct_token_ids[:self.max_struct_tokens]
struct_residue_ids = struct_residue_ids[:self.max_struct_tokens]
struct_len = len(struct_token_ids)
struct_attention_mask = [1] * struct_len
# Pad to max length
padding_len = self.max_struct_tokens - struct_len
struct_token_ids = struct_token_ids + [0] * padding_len
struct_residue_ids = struct_residue_ids + [-1] * padding_len
struct_attention_mask = struct_attention_mask + [0] * padding_len
# Ensure structure tensors are always properly sized
if len(struct_token_ids) != self.max_struct_tokens:
has_3d = False
struct_token_ids = [0] * self.max_struct_tokens
struct_residue_ids = [-1] * self.max_struct_tokens
struct_attention_mask = [0] * self.max_struct_tokens
has_residue_error = seq_data.get('has_residue_error', False)
# Convert to tensors
result = {
'seq_token_ids': torch.tensor(seq_token_ids, dtype=torch.long),
'seq_attention_mask': torch.tensor(seq_attention_mask, dtype=torch.long),
'seq_residue_ids': torch.tensor(seq_residue_ids, dtype=torch.long),
'seq_branch_depths': torch.tensor(seq_branch_depths, dtype=torch.long), # NEW
'seq_linkage_types': torch.tensor(seq_linkage_types, dtype=torch.long), # NEW
'dist_labels': dist_labels, # NEW: Topology Target
# MS Modality
'ms_token_ids': torch.tensor(ms_token_ids, dtype=torch.long),
'ms_attention_mask': torch.tensor(ms_attention_mask, dtype=torch.long),
'ms_residue_ids': torch.tensor(ms_residue_ids, dtype=torch.long),
'struct_token_ids': torch.tensor(struct_token_ids, dtype=torch.long),
'struct_attention_mask': torch.tensor(struct_attention_mask, dtype=torch.long),
'struct_residue_ids': torch.tensor(struct_residue_ids, dtype=torch.long),
'mono_indices': torch.tensor(mono_indices, dtype=torch.long),
'mono_residue_ids': torch.tensor(mono_residue_ids, dtype=torch.long),
'has_ms': torch.tensor(has_ms, dtype=torch.bool),
'has_3d': torch.tensor(has_3d, dtype=torch.bool),
'has_residue_error': torch.tensor(has_residue_error, dtype=torch.bool),
}
# Note: atom_coords, atom_types, and atom-level attention_mask are available
# in self.structure_data but not used (residue-level VQ-VAE tokens are used instead)
return result
def collate_fn(batch: List[Dict[str, torch.Tensor]]) -> Dict[str, torch.Tensor]:
"""
Collate function for batching multimodal samples.
Args:
batch: List of samples from __getitem__
Returns:
Batched tensors
"""
result = {
'seq_token_ids': torch.stack([item['seq_token_ids'] for item in batch]),
'seq_attention_mask': torch.stack([item['seq_attention_mask'] for item in batch]),
'seq_residue_ids': torch.stack([item['seq_residue_ids'] for item in batch]),
'seq_branch_depths': torch.stack([item['seq_branch_depths'] for item in batch]), # NEW
'seq_linkage_types': torch.stack([item['seq_linkage_types'] for item in batch]), # NEW
'ms_token_ids': torch.stack([item['ms_token_ids'] for item in batch]),
'ms_attention_mask': torch.stack([item['ms_attention_mask'] for item in batch]),
'ms_residue_ids': torch.stack([item['ms_residue_ids'] for item in batch]),
'struct_token_ids': torch.stack([item['struct_token_ids'] for item in batch]),
'struct_attention_mask': torch.stack([item['struct_attention_mask'] for item in batch]),
'struct_residue_ids': torch.stack([item['struct_residue_ids'] for item in batch]),
'mono_indices': torch.stack([item['mono_indices'] for item in batch]),
'mono_residue_ids': torch.stack([item['mono_residue_ids'] for item in batch]),
'has_ms': torch.stack([item['has_ms'] for item in batch]),
'has_3d': torch.stack([item['has_3d'] for item in batch]),
'has_residue_error': torch.stack([item['has_residue_error'] for item in batch]),
'dist_labels': torch.stack([item['dist_labels'] for item in batch]), # NEW: Topology
}
return result
def create_multimodal_dataloaders(
sequences_path: str,
ms_tokens_path: str,
structure_data_path: str,
batch_size: int = 64,
num_workers: int = 4,
max_seq_length: int = 512,
max_ms_length: int = 150,
max_struct_length: int = 200,
train_split: float = 0.8,
):
"""
Create train and validation dataloaders for multimodal training.
Args:
sequences_path: Path to sequences.pkl
ms_tokens_path: Path to ms_tokens.pkl
structure_data_path: Path to training_dataset.pkl (VQ-VAE tokens)
batch_size: Batch size
num_workers: Number of data loading workers
max_seq_length: Maximum sequence length
max_ms_length: Maximum MS token length
max_struct_length: Maximum structural token length
train_split: Fraction of data for training (default 0.8 = 80/20 split)
Returns:
train_loader, val_loader
"""
from torch.utils.data import DataLoader, random_split
# Create full dataset
full_dataset = MultimodalGlycanDataset(
sequences_path=sequences_path,
ms_tokens_path=ms_tokens_path,
structure_data_path=structure_data_path,
max_seq_length=max_seq_length,
max_ms_length=max_ms_length,
max_struct_tokens=max_struct_length,
include_ms=True,
include_3d=True,
)
# Split into train and val
total_size = len(full_dataset)
train_size = int(train_split * total_size)
val_size = total_size - train_size
train_dataset, val_dataset = random_split(
full_dataset,
[train_size, val_size],
generator=torch.Generator().manual_seed(42)
)
# Create dataloaders
train_loader = DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
collate_fn=collate_fn,
pin_memory=True,
)
val_loader = DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
collate_fn=collate_fn,
pin_memory=True,
)
print(f"Created dataloaders: {train_size} train, {val_size} val")
return train_loader, val_loader
if __name__ == "__main__":
# Test the dataset
import sys
from pathlib import Path
base_path = Path(__file__).parent.parent / "data"
dataset = MultimodalGlycanDataset(
sequences_path=str(base_path / "sequences.pkl"),
ms_tokens_path=str(base_path / "ms_tokens.pkl"),
structure_data_path=str(Path(__file__).parent.parent.parent / "structure/cluster_upload/files/multimodal_training_package/training_dataset.pkl"),
max_seq_length=512,
max_ms_length=150,
max_struct_tokens=200,
max_atoms=300,
include_ms=True,
include_3d=True,
)
print("="*80)
print("Testing Dataset")
print("="*80)
# Test single sample
sample = dataset[0]
print(f"\nSample 0:")
for key, value in sample.items():
if isinstance(value, torch.Tensor):
print(f" {key}: shape={value.shape}, dtype={value.dtype}")
if key in ['seq_token_ids', 'ms_token_ids', 'struct_token_ids']:
non_zero = (value != 0).sum().item()
print(f" Non-padding tokens: {non_zero}")
else:
print(f" {key}: {value}")
# Test batch
print(f"\n{'='*80}")
print("Testing Batch")
print("="*80)
from torch.utils.data import DataLoader
dataloader = DataLoader(
dataset,
batch_size=4,
shuffle=False,
collate_fn=collate_fn,
)
batch = next(iter(dataloader))
print(f"\nBatch shapes:")
for key, value in batch.items():
print(f" {key}: {value.shape}")
print(f"\nBatch MS availability:")
print(f" Samples with MS: {batch['has_ms'].sum().item()}/{len(batch['has_ms'])}")