PerfectEvent Class Reference

Inheritance diagram for PerfectEvent:

Public Member Functions
def	__init__ (self, ignore_index, output_transform, device='cpu')
def	reset (self)
def	update (self, output)
def	compute (self)

Public Attributes
	ignore_index
	Ignore index.
	device
	CPU or GPU.

Protected Attributes
	_per_corrects
	Good samples.
	_num_examples
	Total samples.

Detailed Description

Computes the rate of events with perfectly predicted mass hypotheses and LCAS matrices over a batch.

``output_transform`` should return the following items:
``(x_pred, x_y, edge_pred, edge_y, edge_index, u_y, batch, num_graphs)``.

* ``x_pred`` must contain node prediction logits and have shape (num_nodes_in_batch, node_classes);
* ``x_y`` must contain node ground-truth class indices and have shape (num_nodes_in_batch, 1);
* ``edge_pred`` must contain edge prediction logits and have shape (num_edges_in_batch, edge_classes);
* ``edge_y`` must contain edge ground-truth class indices and have shape (num_edges_in_batch, 1);
* ``edge index`` maps edges to its nodes;
* ``u_y`` is the signal/background class (always 1 in the current setting);
* ``batch`` maps nodes to their graph;
* ``num_graphs`` is the number of graph in a batch (could be derived from ``batch`` also).

.. seealso::
    `Ignite metrics <https://pytorch.org/ignite/metrics.html>`_

:param ignore_index: Class or list of classes to ignore during the computation (e.g. padding).
:type ignore_index: list[int]
:param output_transform: Function to transform engine's output to desired output.
:type output_transform: `function <https://docs.python.org/3/glossary.html#term-function>`_
:param device: ``cpu`` or ``gpu``.
:type device: str

Definition at line 205 of file metrics.py.

Constructor & Destructor Documentation

__init__()

def __init__	(		self,
			ignore_index,
			output_transform,
			device = 'cpu'
	)

Initialization.

Definition at line 232 of file metrics.py.

    def __init__(self, ignore_index, output_transform, device='cpu'):
        """
        Initialization.
        """
        
        self.ignore_index = ignore_index if isinstance(ignore_index, list) else [ignore_index]
        
        self.device = device
        
        self._per_corrects = None
        
        self._num_examples = None
 
        super(PerfectEvent, self).__init__(output_transform=output_transform, device=device)
 

Member Function Documentation

compute()

def compute

(

self

)

Final computation.

Definition at line 304 of file metrics.py.

    def compute(self):
        """
        Final computation.
        """
        if self._num_examples == 0:
            raise NotComputableError(
                "CustomAccuracy must have at least one example before it can be computed."
            )
        return self._per_corrects / self._num_examples

reset()

def reset

(

self

)

Resets counts.

Definition at line 248 of file metrics.py.

    def reset(self):
        """
        Resets counts.
        """
        self._per_corrects = 0
        self._num_examples = 0
 
        super(PerfectEvent, self).reset()
 

update()

def update	(		self,
			output
	)

Updates counts.

Definition at line 258 of file metrics.py.

    def update(self, output):
        """
        Updates counts.
        """
        x_pred, x_y, edge_pred, edge_y, edge_index, u_y, batch, num_graphs = output
 
        num_graphs = num_graphs.item()
 
        x_probs = torch.softmax(x_pred, dim=1)
        x_winners = x_probs.argmax(dim=1)
        edge_probs = torch.softmax(edge_pred, dim=1)
        edge_winners = edge_probs.argmax(dim=1)
 
        assert x_winners.shape == x_y.shape, 'Mass predictions shape does not match target shape'
        assert edge_winners.shape == edge_y.shape, 'Edge predictions shape does not match target shape'
 
        # Create a mask for the zeroth elements (padded entries)
        x_mask = torch.ones(x_y.size(), dtype=torch.long, device=self.device)
        edge_mask = torch.ones(edge_y.size(), dtype=torch.long, device=self.device)
        for ig_class in self.ignore_index:
            x_mask &= (x_y != ig_class)
            edge_mask &= (edge_y != ig_class)
 
        # Zero the respective entries in the predictions
        x_pred_mask = x_winners * x_mask
        x_mask = x_y * x_mask
        edge_pred_mask = edge_winners * edge_mask
        edge_mask = edge_y * edge_mask
 
        # (N) compare the masked predictions with the target. The padded will be equal due to masking
        # Masses
        x_truth = x_pred_mask.eq(x_mask) + 0  # +0 so it's not bool but 0 and 1
        x_truth = scatter(x_truth, batch, reduce="min")
        # Edges
        edge_truth = edge_pred_mask.eq(edge_mask) + 0  # +0 so it's not bool but 0 and 1
        edge_truth = scatter(edge_truth, edge_index[0], reduce="min")
        edge_truth = scatter(edge_truth, batch, reduce="min")
 
        # Count the number of zero wrong predictions across the batch
        truth = x_truth.bool() & edge_truth.bool()
        batch_perfect = (truth + 0).sum().item()
 
        self._per_corrects += batch_perfect
        self._num_examples += num_graphs
 

Member Data Documentation

_num_examples

_num_examples

protected

Total samples.

Definition at line 243 of file metrics.py.

_per_corrects

_per_corrects

protected

Good samples.

Definition at line 241 of file metrics.py.

device

device

CPU or GPU.

Definition at line 239 of file metrics.py.

ignore_index

ignore_index

Ignore index.

Definition at line 237 of file metrics.py.

---

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

• analysis/scripts/grafei/model/metrics.py