SparseDataset Class Reference

Inheritance diagram for SparseDataset:

Collaboration diagram for SparseDataset:

Public Member Functions
	__init__ (self, features_sparse, labels, mbc_values=None, extra_features=None, weights=None)
	__len__ (self)
	__getitem__ (self, idx)

Public Attributes
	features = features_sparse.tocsr()
	Feature matrix in CSR format.
	labels = labels
	Target class labels.
	mbc_values = mbc_values
	Mbc values per event for DisCo loss, or None.
	extra_features = extra_features
	Additional dense features concatenated at retrieval time, or None.
	weights = weights
	Per-event loss weights, or None.

Detailed Description

Dataset for sparse feature matrices.

Converts each row to dense on-the-fly during batching.
Slower than dense loading but significantly more memory-efficient.

Definition at line 1096 of file train.py.

Constructor & Destructor Documentation

__init__()

__init__	(		self,
			features_sparse,
			labels,
			mbc_values = None,
			extra_features = None,
			weights = None )

Parameters:
    features_sparse (scipy.sparse matrix): Sparse feature matrix of shape
        (n_samples, n_features). CSR format recommended.
    labels (np.ndarray): Target labels, shape (n_samples,).
    mbc_values (np.ndarray, optional): Mbc values per event for DisCo loss,
        shape (n_samples,).
    extra_features (np.ndarray, optional): Additional dense features concatenated
        per sample, shape (n_samples, n_extra).
    weights (np.ndarray, optional): Per-event loss weights, shape (n_samples,)
        (e.g. FEI calibration weights).

Definition at line 1103 of file train.py.

                 weights=None):
        """
        Parameters:
            features_sparse (scipy.sparse matrix): Sparse feature matrix of shape
                (n_samples, n_features). CSR format recommended.
            labels (np.ndarray): Target labels, shape (n_samples,).
            mbc_values (np.ndarray, optional): Mbc values per event for DisCo loss,
                shape (n_samples,).
            extra_features (np.ndarray, optional): Additional dense features concatenated
                per sample, shape (n_samples, n_extra).
            weights (np.ndarray, optional): Per-event loss weights, shape (n_samples,)
                (e.g. FEI calibration weights).
        """
        
        self.features = features_sparse.tocsr()
        
        self.labels = labels
        
        self.mbc_values = mbc_values
        
        self.extra_features = extra_features
        
        self.weights = weights
 

Member Function Documentation

__getitem__()

__getitem__	(		self,
			idx )

Return one sample (features, label, [mbc, [weight]]) for the given index.

Definition at line 1132 of file train.py.

    def __getitem__(self, idx):
        """Return one sample (features, label, [mbc, [weight]]) for the given index."""
        # Convert sparse row to dense 1D tensor
        feature_row = self.features[idx].toarray().astype(np.float32).squeeze()
        if self.extra_features is not None:
            feature_row = np.concatenate([feature_row, self.extra_features[idx]]).astype(np.float32)
        feature_row = torch.from_numpy(feature_row)
        label = torch.tensor(self.labels[idx], dtype=torch.long)
 
        if self.mbc_values is not None:
            mbc = torch.tensor(self.mbc_values[idx], dtype=torch.float32)
            if self.weights is not None:
                w = torch.tensor(self.weights[idx], dtype=torch.float32)
                return feature_row, label, mbc, w
            return feature_row, label, mbc
 
        if self.weights is not None:
            w = torch.tensor(self.weights[idx], dtype=torch.float32)
            return feature_row, label, w
 
        return feature_row, label
 
 

__len__()

__len__

(

self

)

Return the number of samples in the dataset.

Definition at line 1128 of file train.py.

    def __len__(self):
        """Return the number of samples in the dataset."""
        return self.features.shape[0]
 

Member Data Documentation

extra_features

extra_features = extra_features

Additional dense features concatenated at retrieval time, or None.

Definition at line 1124 of file train.py.

features

features = features_sparse.tocsr()

Feature matrix in CSR format.

Definition at line 1118 of file train.py.

labels

labels = labels

Target class labels.

Definition at line 1120 of file train.py.

mbc_values

mbc_values = mbc_values

Mbc values per event for DisCo loss, or None.

Definition at line 1122 of file train.py.

weights

weights = weights

Per-event loss weights, or None.

Definition at line 1126 of file train.py.

---

The documentation for this class was generated from the following file:

• analysis/scripts/modeSelector/training/train.py