ISubGVQA.sampling.methods.simple_scheme
=======================================

.. py:module:: ISubGVQA.sampling.methods.simple_scheme


Attributes
----------

.. autoapisummary::

   ISubGVQA.sampling.methods.simple_scheme.LARGE_NUMBER


Classes
-------

.. autoapisummary::

   ISubGVQA.sampling.methods.simple_scheme.EdgeSIMPLEBatched


Functions
---------

.. autoapisummary::

   ISubGVQA.sampling.methods.simple_scheme.logsigmoid


Module Contents
---------------

.. py:data:: LARGE_NUMBER
   :value: 10000000000.0


.. py:function:: logsigmoid(x)

.. py:class:: EdgeSIMPLEBatched(k, device, policy, val_ensemble=1, train_ensemble=1, logits_activation=None)

   Bases: :py:obj:`torch.nn.Module`


   .. py:attribute:: k


   .. py:attribute:: device


   .. py:attribute:: policy


   .. py:attribute:: layer_configs


   .. py:attribute:: adj
      :value: None



   .. py:attribute:: val_ensemble
      :value: 1



   .. py:attribute:: train_ensemble
      :value: 1



   .. py:attribute:: logits_activation
      :value: None



   .. py:method:: forward(scores, train=True)


   .. py:method:: validation(scores)

      during the inference we need to margin-out the stochasticity
      thus we do top-k once or sample multiple times

      Args:
          scores: shape B x N x N x E

      Returns:
          mask: shape B x N x N x (E x VE)