DNA-CASINO / dna_diffusion_model.py
lucapinello
Add application file
b074e28
raw
history blame
11.8 kB
"""
DNA-Diffusion Model Wrapper
Singleton class to handle model loading and sequence generation
"""
import os
import sys
import torch
import numpy as np
import logging
from typing import Optional, Dict, List
import time
logger = logging.getLogger(__name__)
class DNADiffusionModel:
"""Singleton wrapper for DNA-Diffusion model"""
_instance = None
_initialized = False
# Cell type mapping from simple names to dataset identifiers
CELL_TYPE_MAPPING = {
'K562': 'K562_ENCLB843GMH',
'GM12878': 'GM12878_ENCLB441ZZZ',
'HepG2': 'HepG2_ENCLB029COU',
'hESCT0': 'hESCT0_ENCLB449ZZZ'
}
def __new__(cls):
if cls._instance is None:
cls._instance = super().__new__(cls)
return cls._instance
def __init__(self):
"""Initialize the model (only runs once due to singleton pattern)"""
if not self._initialized:
self._initialize()
self._initialized = True
def _initialize(self):
"""Load model and setup components"""
try:
logger.info("Initializing DNA-Diffusion model...")
# Add DNA-Diffusion to path
dna_diffusion_path = os.path.join(os.path.dirname(__file__), 'DNA-Diffusion')
if os.path.exists(dna_diffusion_path):
sys.path.insert(0, os.path.join(dna_diffusion_path, 'src'))
# Import DNA-Diffusion components
from dnadiffusion.models.pretrained_unet import PretrainedUNet
from dnadiffusion.models.diffusion import Diffusion
from dnadiffusion.data.dataloader import get_dataset_for_sampling
# Load pretrained model from HuggingFace
logger.info("Loading pretrained model from HuggingFace...")
self.model = PretrainedUNet.from_pretrained("ssenan/DNA-Diffusion")
# Move to GPU if available
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {self.device}")
self.model = self.model.to(self.device)
self.model.eval()
# Initialize diffusion sampler with the model
self.diffusion = Diffusion(
model=self.model,
timesteps=50,
beta_start=0.0001,
beta_end=0.2
)
# Ensure output_attention is set to False initially
if hasattr(self.model, 'output_attention'):
self.model.output_attention = False
if hasattr(self.model.model, 'output_attention'):
self.model.model.output_attention = False
# Setup dataset for sampling
data_path = os.path.join(dna_diffusion_path, "data/K562_hESCT0_HepG2_GM12878_12k_sequences_per_group.txt")
saved_data_path = os.path.join(dna_diffusion_path, "data/encode_data.pkl")
# Get dataset info
train_data, val_data, cell_num_list, numeric_to_tag_dict = get_dataset_for_sampling(
data_path=data_path,
saved_data_path=saved_data_path,
load_saved_data=True,
debug=False,
cell_types=None # Load all cell types
)
# Store dataset info
self.train_data = train_data
self.val_data = val_data
self.cell_num_list = cell_num_list
self.numeric_to_tag_dict = numeric_to_tag_dict
# Get available cell types
self.available_cell_types = [numeric_to_tag_dict[num] for num in cell_num_list]
logger.info(f"Available cell types: {self.available_cell_types}")
# Warm up the model with a test generation
logger.info("Warming up model...")
self._warmup()
logger.info("Model initialization complete!")
except Exception as e:
logger.error(f"Failed to initialize model: {str(e)}")
self.model = None
self.diffusion = None
self.dataset = None
raise
def _warmup(self):
"""Warm up the model with a test generation"""
try:
# Generate one sequence for the first available cell type
if self.available_cell_types:
cell_type = list(self.CELL_TYPE_MAPPING.keys())[0]
self.generate(cell_type, guidance_scale=1.0)
except Exception as e:
logger.warning(f"Warmup generation failed: {str(e)}")
def is_ready(self) -> bool:
"""Check if model is loaded and ready"""
return self.model is not None and self.diffusion is not None and self.train_data is not None
def generate(self, cell_type: str, guidance_scale: float = 1.0) -> Dict[str, any]:
"""
Generate a DNA sequence for the specified cell type
Args:
cell_type: Simple cell type name (K562, GM12878, HepG2, hESCT0)
guidance_scale: Guidance scale for generation (1.0-10.0)
Returns:
Dict with 'sequence' (200bp string) and 'metadata'
"""
if not self.is_ready():
raise RuntimeError("Model is not initialized")
# Validate inputs
if cell_type not in self.CELL_TYPE_MAPPING:
raise ValueError(f"Invalid cell type: {cell_type}. Must be one of {list(self.CELL_TYPE_MAPPING.keys())}")
if not 1.0 <= guidance_scale <= 10.0:
raise ValueError(f"Guidance scale must be between 1.0 and 10.0, got {guidance_scale}")
# Map to full cell type identifier
full_cell_type = self.CELL_TYPE_MAPPING[cell_type]
# Find the numeric index for this cell type
tag_to_numeric = {tag: num for num, tag in self.numeric_to_tag_dict.items()}
# Find matching cell type (case-insensitive partial match)
cell_type_numeric = None
for tag, num in tag_to_numeric.items():
if full_cell_type.lower() in tag.lower() or tag.lower() in full_cell_type.lower():
cell_type_numeric = num
logger.info(f"Matched '{full_cell_type}' to '{tag}'")
break
if cell_type_numeric is None:
raise ValueError(f"Cell type {full_cell_type} not found in dataset. Available: {list(self.numeric_to_tag_dict.values())}")
try:
logger.info(f"Generating sequence for {cell_type} (guidance={guidance_scale})...")
start_time = time.time()
# For now, use simple generation without classifier-free guidance
# TODO: Fix classifier-free guidance implementation
sequence = self._generate_simple(cell_type_numeric, guidance_scale)
generation_time = time.time() - start_time
logger.info(f"Generated sequence in {generation_time:.2f}s")
return {
'sequence': sequence,
'metadata': {
'cell_type': cell_type,
'full_cell_type': full_cell_type,
'guidance_scale': guidance_scale,
'generation_time': generation_time,
'sequence_length': len(sequence)
}
}
except Exception as e:
logger.error(f"Generation failed: {str(e)}")
raise
def _generate_simple(self, cell_type_idx: int, guidance_scale: float) -> str:
"""Simple generation using the diffusion model's sample method"""
with torch.no_grad():
# For guidance_scale = 1.0, use simple generation without classifier-free guidance
if guidance_scale == 1.0:
# Create initial noise
img = torch.randn((1, 1, 4, 200), device=self.device)
# Simple denoising loop without guidance
for i in reversed(range(self.diffusion.timesteps)):
t = torch.full((1,), i, device=self.device, dtype=torch.long)
# Get model prediction with classes
classes = torch.tensor([cell_type_idx], device=self.device, dtype=torch.long)
noise_pred = self.model(img, time=t, classes=classes)
# Denoising step
betas_t = self.diffusion.betas[i]
sqrt_one_minus_alphas_cumprod_t = self.diffusion.sqrt_one_minus_alphas_cumprod[i]
sqrt_recip_alphas_t = self.diffusion.sqrt_recip_alphas[i]
# Predict x0
model_mean = sqrt_recip_alphas_t * (img - betas_t * noise_pred / sqrt_one_minus_alphas_cumprod_t)
if i == 0:
img = model_mean
else:
posterior_variance_t = self.diffusion.posterior_variance[i]
noise = torch.randn_like(img)
img = model_mean + torch.sqrt(posterior_variance_t) * noise
final_image = img[0] # Remove batch dimension
else:
# Use the diffusion model's built-in sample method with guidance
# This requires proper context mask handling which is complex
# For now, fall back to simple generation
logger.warning(f"Guidance scale {guidance_scale} not fully implemented, using simple generation")
return self._generate_simple(cell_type_idx, 1.0)
# Convert to sequence
final_array = final_image.cpu().numpy()
sequence = self._array_to_sequence(final_array)
return sequence
def _array_to_sequence(self, array: np.ndarray) -> str:
"""Convert model output array to DNA sequence string"""
# Get nucleotide mapping
nucleotides = ['A', 'C', 'G', 'T']
# array shape is (1, 4, 200) - channels, nucleotides, sequence_length
# Reshape to (4, 200) and get argmax along nucleotide dimension
array = array.squeeze(0) # Remove channel dimension -> (4, 200)
indices = np.argmax(array, axis=0) # Get max nucleotide for each position
# Convert indices to nucleotides
sequence = ''.join(nucleotides[int(idx)] for idx in indices)
return sequence
def get_model_info(self) -> Dict[str, any]:
"""Get information about the loaded model"""
if not self.is_ready():
return {'status': 'not_initialized'}
return {
'status': 'ready',
'device': str(self.device),
'cell_types': list(self.CELL_TYPE_MAPPING.keys()),
'full_cell_types': self.available_cell_types,
'model_name': 'ssenan/DNA-Diffusion',
'sequence_length': 200,
'guidance_scale_range': [1.0, 10.0]
}
# Convenience functions for direct usage
_model_instance = None
def get_model() -> DNADiffusionModel:
"""Get or create the singleton model instance"""
global _model_instance
if _model_instance is None:
_model_instance = DNADiffusionModel()
return _model_instance
def generate_sequence(cell_type: str, guidance_scale: float = 1.0) -> str:
"""Generate a DNA sequence (convenience function)"""
model = get_model()
result = model.generate(cell_type, guidance_scale)
return result['sequence']