Source code for deeppy.feedforward.dropout_layers

import cudarray as ca
from .layers import Layer, FullyConnected
from ..base import PhaseMixin


class Dropout(Layer, PhaseMixin):
    def __init__(self, dropout=0.5):
        self.name = 'dropout'
        self.dropout = dropout
        self._tmp_mask = None

    def fprop(self, x):
        if self.dropout > 0.0:
            if self.phase == 'train':
                self._tmp_mask = self.dropout < ca.random.uniform(size=x.shape)
                y = x * self._tmp_mask
            elif self.phase == 'test':
                y = x * (1.0 - self.dropout)
            else:
                raise ValueError('Invalid phase: %s' % self.phase)
        return y

    def bprop(self, y_grad):
        if self.dropout > 0.0:
            return y_grad * self._tmp_mask
        else:
            return y_grad

    def y_shape(self, x_shape):
        return x_shape


class DropoutFullyConnected(FullyConnected, PhaseMixin):
    def __init__(self, n_out, weights, bias=0.0, dropout=0.5):
        super(DropoutFullyConnected, self).__init__(
            n_out=n_out, weights=weights, bias=bias
        )
        self.name = 'fc_drop'
        self.dropout = dropout
        self._tmp_mask = None

    def fprop(self, x):
        y = super(DropoutFullyConnected, self).fprop(x)
        if self.dropout > 0.0:
            if self.phase == 'train':
                self._tmp_mask = self.dropout < ca.random.uniform(size=y.shape)
                y *= self._tmp_mask
            elif self.phase == 'test':
                y *= (1.0 - self.dropout)
            else:
                raise ValueError('Invalid phase: %s' % self.phase)
        return y

    def bprop(self, y_grad):
        if self.dropout > 0.0:
            y_grad *= self._tmp_mask
        return super(DropoutFullyConnected, self).bprop(y_grad)