cda-hmoc-classifier / scripts /train_hmoc_ensemble.py

Upload folder using huggingface_hub

1c51b77 verified 15 days ago

7.54 kB

import pandas as pd
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import precision_recall_curve, auc, f1_score, roc_auc_score
import os

# --- Configurations ---
DATA_PATH = "./data/cda_qm_spectra_pre2008277_lvl3_train.parquet"
MODELS_DIR = "./hmoc_binary_classifier/models"
os.makedirs(MODELS_DIR, exist_ok=True)

BATCH_SIZE = 64
EPOCHS = 100
PATIENCE = 15
LEARNING_RATE = 1e-3
WEIGHT_DECAY = 1e-4

# --- Focal Loss ---
class FocalLoss(nn.Module):
    def __init__(self, alpha=0.85, gamma=2.0, reduction='mean'):
        super(FocalLoss, self).__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.reduction = reduction

    def forward(self, inputs, targets):
        BCE_loss = nn.functional.binary_cross_entropy_with_logits(inputs, targets, reduction='none')
        targets = targets.type(torch.float32)
        pt = torch.exp(-BCE_loss)
        F_loss = self.alpha * (1-pt)**self.gamma * BCE_loss

        if self.reduction == 'mean':
            return torch.mean(F_loss)
        else:
            return F_loss

# --- Model Architecture ---
class Spectra1DCNN(nn.Module):
    def __init__(self, seq_len):
        super(Spectra1DCNN, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv1d(1, 32, kernel_size=7, padding=3),
            nn.BatchNorm1d(32),
            nn.ReLU(),
            nn.Conv1d(32, 64, kernel_size=5, padding=2),
            nn.BatchNorm1d(64),
            nn.ReLU(),
            nn.Conv1d(64, 128, kernel_size=3, padding=1),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.AdaptiveAvgPool1d(1)
        )
        self.fc = nn.Sequential(
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Dropout(0.4),
            nn.Linear(64, 1)
        )

    def forward(self, x):
        # x shape: (batch, 1, seq_len)
        x = self.conv(x)
        x = x.squeeze(-1) # (batch, 128)
        x = self.fc(x)
        return x

class SpectraDataset(Dataset):
    def __init__(self, X, y):
        self.X = torch.tensor(X, dtype=torch.float32).unsqueeze(1) # Add channel dim
        self.y = torch.tensor(y, dtype=torch.float32).unsqueeze(1)
        
    def __len__(self):
        return len(self.X)
    
    def __getitem__(self, idx):
        return self.X[idx], self.y[idx]

def pad_or_truncate(x, target_len=1000):
    if len(x) > target_len:
        return x[:target_len]
    elif len(x) < target_len:
        return np.pad(x, (0, target_len - len(x)))
    return x

def main():
    print("Loading data...")
    df = pd.read_parquet(DATA_PATH)
    
    # Filter negatives
    exclude_classes = ['?', 'Noise']
    df_filtered = df[~df['class'].isin(exclude_classes)].copy()
    
    # Map classes
    df_filtered['label'] = (df_filtered['class'] == 'HMOC').astype(int)
    print(f"Dataset shape after filtering: {df_filtered.shape}")
    print(f"Class distribution:\n{df_filtered['label'].value_counts()}")
    
    # Process sequences
    print("Processing spectra arrays...")
    TARGET_LEN = 1000
    X = np.stack(df_filtered['spectrum'].apply(lambda s: pad_or_truncate(np.array(s), TARGET_LEN)).values)
    
    # Simple Min-Max normalization per spectrum
    X_min = X.min(axis=1, keepdims=True)
    X_max = X.max(axis=1, keepdims=True)
    X = (X - X_min) / (X_max - X_min + 1e-8)
    
    y = df_filtered['label'].values
    
    print(f"X shape: {X.shape}, y shape: {y.shape}")
    
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Using device: {device}")
    
    skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
    ensemble_seeds = [42, 100, 2026, 777, 888]
    
    for seed_idx, seed in enumerate(ensemble_seeds):
        print(f"\\n--- Training Ensemble Model {seed_idx+1}/{len(ensemble_seeds)} (Seed {seed}) ---")
        torch.manual_seed(seed)
        np.random.seed(seed)
        
        # We will just train on one fold per seed for the ensemble to maximize data efficiency,
        # or we could do 5 models total, each being a full train on 4/5ths of the data.
        # Let's use the fold corresponding to the seed index
        fold = seed_idx % 5
        train_idx, val_idx = list(skf.split(X, y))[fold]
        
        X_train, y_train = X[train_idx], y[train_idx]
        X_val, y_val = X[val_idx], y[val_idx]
        
        train_dataset = SpectraDataset(X_train, y_train)
        val_dataset = SpectraDataset(X_val, y_val)
        
        # Use simple augmentation (Label smoothing handled in loss conceptually, but let's just train standard PyTorch)
        # To handle imbalance in batches, we use a balanced sampler
        class_sample_count = np.array([len(np.where(y_train == t)[0]) for t in np.unique(y_train)])
        weight = 1. / class_sample_count
        samples_weight = np.array([weight[t] for t in y_train])
        samples_weight = torch.from_numpy(samples_weight)
        sampler = torch.utils.data.WeightedRandomSampler(samples_weight.type('torch.DoubleTensor'), len(samples_weight))
        
        train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, sampler=sampler, num_workers=2)
        val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False)
        
        model = Spectra1DCNN(TARGET_LEN).to(device)
        criterion = FocalLoss(alpha=0.85, gamma=2.0)
        optimizer = optim.AdamW(model.parameters(), lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)
        
        best_auprc = 0.0
        patience_counter = 0
        
        for epoch in range(EPOCHS):
            model.train()
            train_loss = 0.0
            for batch_x, batch_y in train_loader:
                batch_x, batch_y = batch_x.to(device), batch_y.to(device)
                
                optimizer.zero_grad()
                outputs = model(batch_x)
                loss = criterion(outputs, batch_y)
                loss.backward()
                optimizer.step()
                
                train_loss += loss.item()
                
            model.eval()
            val_preds = []
            val_targs = []
            with torch.no_grad():
                for batch_x, batch_y in val_loader:
                    batch_x = batch_x.to(device)
                    outputs = torch.sigmoid(model(batch_x))
                    val_preds.extend(outputs.cpu().numpy().flatten())
                    val_targs.extend(batch_y.numpy().flatten())
            
            val_preds = np.array(val_preds)
            val_targs = np.array(val_targs)
            
            # Metrics
            precision, recall, _ = precision_recall_curve(val_targs, val_preds)
            auprc = auc(recall, precision)
            
            if auprc > best_auprc:
                best_auprc = auprc
                torch.save(model.state_dict(), os.path.join(MODELS_DIR, f"hmoc_model_seed_{seed}.pth"))
                patience_counter = 0
            else:
                patience_counter += 1
                
            if np.isnan(auprc) or (epoch + 1) % 5 == 0:
                print(f"Epoch {epoch+1:03d} | Train Loss: {train_loss/len(train_loader):.4f} | Val AUPRC: {auprc:.4f}")
                
            if patience_counter >= PATIENCE:
                print(f"Early stopping at epoch {epoch+1}")
                break
                
        print(f"Best Val AUPRC for seed {seed}: {best_auprc:.4f}")

if __name__ == "__main__":
    main()

	import pandas as pd
	import numpy as np
	import torch
	import torch.nn as nn
	import torch.optim as optim
	from torch.utils.data import Dataset, DataLoader
	from sklearn.model_selection import StratifiedKFold
	from sklearn.metrics import precision_recall_curve, auc, f1_score, roc_auc_score
	import os

	# --- Configurations ---
	DATA_PATH = "./data/cda_qm_spectra_pre2008277_lvl3_train.parquet"
	MODELS_DIR = "./hmoc_binary_classifier/models"
	os.makedirs(MODELS_DIR, exist_ok=True)

	BATCH_SIZE = 64
	EPOCHS = 100
	PATIENCE = 15
	LEARNING_RATE = 1e-3
	WEIGHT_DECAY = 1e-4

	# --- Focal Loss ---
	class FocalLoss(nn.Module):
	def __init__(self, alpha=0.85, gamma=2.0, reduction='mean'):
	super(FocalLoss, self).__init__()
	self.alpha = alpha
	self.gamma = gamma
	self.reduction = reduction

	def forward(self, inputs, targets):
	BCE_loss = nn.functional.binary_cross_entropy_with_logits(inputs, targets, reduction='none')
	targets = targets.type(torch.float32)
	pt = torch.exp(-BCE_loss)
	F_loss = self.alpha * (1-pt)*self.gamma BCE_loss

	if self.reduction == 'mean':
	return torch.mean(F_loss)
	else:
	return F_loss

	# --- Model Architecture ---
	class Spectra1DCNN(nn.Module):
	def __init__(self, seq_len):
	super(Spectra1DCNN, self).__init__()
	self.conv = nn.Sequential(
	nn.Conv1d(1, 32, kernel_size=7, padding=3),
	nn.BatchNorm1d(32),
	nn.ReLU(),
	nn.Conv1d(32, 64, kernel_size=5, padding=2),
	nn.BatchNorm1d(64),
	nn.ReLU(),
	nn.Conv1d(64, 128, kernel_size=3, padding=1),
	nn.BatchNorm1d(128),
	nn.ReLU(),
	nn.AdaptiveAvgPool1d(1)
	)
	self.fc = nn.Sequential(
	nn.Linear(128, 64),
	nn.ReLU(),
	nn.Dropout(0.4),
	nn.Linear(64, 1)
	)

	def forward(self, x):
	# x shape: (batch, 1, seq_len)
	x = self.conv(x)
	x = x.squeeze(-1) # (batch, 128)
	x = self.fc(x)
	return x

	class SpectraDataset(Dataset):
	def __init__(self, X, y):
	self.X = torch.tensor(X, dtype=torch.float32).unsqueeze(1) # Add channel dim
	self.y = torch.tensor(y, dtype=torch.float32).unsqueeze(1)

	def __len__(self):
	return len(self.X)

	def __getitem__(self, idx):
	return self.X[idx], self.y[idx]

	def pad_or_truncate(x, target_len=1000):
	if len(x) > target_len:
	return x[:target_len]
	elif len(x) < target_len:
	return np.pad(x, (0, target_len - len(x)))
	return x

	def main():
	print("Loading data...")
	df = pd.read_parquet(DATA_PATH)

	# Filter negatives
	exclude_classes = ['?', 'Noise']
	df_filtered = df[~df['class'].isin(exclude_classes)].copy()

	# Map classes
	df_filtered['label'] = (df_filtered['class'] == 'HMOC').astype(int)
	print(f"Dataset shape after filtering: {df_filtered.shape}")
	print(f"Class distribution:\n{df_filtered['label'].value_counts()}")

	# Process sequences
	print("Processing spectra arrays...")
	TARGET_LEN = 1000
	X = np.stack(df_filtered['spectrum'].apply(lambda s: pad_or_truncate(np.array(s), TARGET_LEN)).values)

	# Simple Min-Max normalization per spectrum
	X_min = X.min(axis=1, keepdims=True)
	X_max = X.max(axis=1, keepdims=True)
	X = (X - X_min) / (X_max - X_min + 1e-8)

	y = df_filtered['label'].values

	print(f"X shape: {X.shape}, y shape: {y.shape}")

	device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
	print(f"Using device: {device}")

	skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
	ensemble_seeds = [42, 100, 2026, 777, 888]

	for seed_idx, seed in enumerate(ensemble_seeds):
	print(f"\\n--- Training Ensemble Model {seed_idx+1}/{len(ensemble_seeds)} (Seed {seed}) ---")
	torch.manual_seed(seed)
	np.random.seed(seed)

	# We will just train on one fold per seed for the ensemble to maximize data efficiency,
	# or we could do 5 models total, each being a full train on 4/5ths of the data.
	# Let's use the fold corresponding to the seed index
	fold = seed_idx % 5
	train_idx, val_idx = list(skf.split(X, y))[fold]

	X_train, y_train = X[train_idx], y[train_idx]
	X_val, y_val = X[val_idx], y[val_idx]

	train_dataset = SpectraDataset(X_train, y_train)
	val_dataset = SpectraDataset(X_val, y_val)

	# Use simple augmentation (Label smoothing handled in loss conceptually, but let's just train standard PyTorch)
	# To handle imbalance in batches, we use a balanced sampler
	class_sample_count = np.array([len(np.where(y_train == t)[0]) for t in np.unique(y_train)])
	weight = 1. / class_sample_count
	samples_weight = np.array([weight[t] for t in y_train])
	samples_weight = torch.from_numpy(samples_weight)
	sampler = torch.utils.data.WeightedRandomSampler(samples_weight.type('torch.DoubleTensor'), len(samples_weight))

	train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, sampler=sampler, num_workers=2)
	val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False)

	model = Spectra1DCNN(TARGET_LEN).to(device)
	criterion = FocalLoss(alpha=0.85, gamma=2.0)
	optimizer = optim.AdamW(model.parameters(), lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)

	best_auprc = 0.0
	patience_counter = 0

	for epoch in range(EPOCHS):
	model.train()
	train_loss = 0.0
	for batch_x, batch_y in train_loader:
	batch_x, batch_y = batch_x.to(device), batch_y.to(device)

	optimizer.zero_grad()
	outputs = model(batch_x)
	loss = criterion(outputs, batch_y)
	loss.backward()
	optimizer.step()

	train_loss += loss.item()

	model.eval()
	val_preds = []
	val_targs = []
	with torch.no_grad():
	for batch_x, batch_y in val_loader:
	batch_x = batch_x.to(device)
	outputs = torch.sigmoid(model(batch_x))
	val_preds.extend(outputs.cpu().numpy().flatten())
	val_targs.extend(batch_y.numpy().flatten())

	val_preds = np.array(val_preds)
	val_targs = np.array(val_targs)

	# Metrics
	precision, recall, _ = precision_recall_curve(val_targs, val_preds)
	auprc = auc(recall, precision)

	if auprc > best_auprc:
	best_auprc = auprc
	torch.save(model.state_dict(), os.path.join(MODELS_DIR, f"hmoc_model_seed_{seed}.pth"))
	patience_counter = 0
	else:
	patience_counter += 1

	if np.isnan(auprc) or (epoch + 1) % 5 == 0:
	print(f"Epoch {epoch+1:03d} \| Train Loss: {train_loss/len(train_loader):.4f} \| Val AUPRC: {auprc:.4f}")

	if patience_counter >= PATIENCE:
	print(f"Early stopping at epoch {epoch+1}")
	break

	print(f"Best Val AUPRC for seed {seed}: {best_auprc:.4f}")

	if __name__ == "__main__":
	main()