Key Code Blocks of Pytorch RNN Dropout Implementation

weight_drop.py

# https://github.com/salesforce/awd-lstm-lm/blob/dfd3cb0235d2caf2847a4d53e1cbd495b781b5d2/weight_drop.py#L5
class WeightDrop(torch.nn.Module):
    def __init__(self, module, weights, dropout=0, variational=False):
        # ...
        self._setup()
    # ...
    def _setup(self):
        # Terrible temporary solution to an issue regarding compacting weights re: CUDNN RNN
        if issubclass(type(self.module), torch.nn.RNNBase):
            self.module.flatten_parameters = self.widget_demagnetizer_y2k_edition
        for name_w in self.weights:
            print('Applying weight drop of {} to {}'.format(self.dropout, name_w))
            w = getattr(self.module, name_w)
            del self.module._parameters[name_w]
            self.module.register_parameter(name_w + '_raw', Parameter(w.data))

    def _setweights(self):
        for name_w in self.weights:
            raw_w = getattr(self.module, name_w + '_raw')
            w = None
            if self.variational:
                mask = torch.autograd.Variable(torch.ones(raw_w.size(0), 1))
                if raw_w.is_cuda: mask = mask.cuda()
                mask = torch.nn.functional.dropout(mask, p=self.dropout, training=True)
                w = mask.expand_as(raw_w) * raw_w
            else:
                w = torch.nn.functional.dropout(raw_w, p=self.dropout, training=self.training)
            setattr(self.module, name_w, w)

    def forward(self, *args):
        self._setweights()