def test_get_output_for_shared_axes(self, shared_axes):
from lasagne.layers.noise import DropoutLayer
layer = DropoutLayer((2, 4, 7, 9), shared_axes=shared_axes)
input = theano.shared(numpy.ones((2, 4, 7, 9)))
result = layer.get_output_for(input)
result_eval = result.eval()
# check if the dropout mask is the same across the specified axes:
# compute the mean across these axes and compare against the full
# output, broadcasting across the shared axes, to see if it matches
assert np.allclose(result_eval.mean(axis=shared_axes, keepdims=True),
result_eval)
def test_get_output_for_shared_axes(self, shared_axes):
from lasagne.layers.noise import DropoutLayer
layer = DropoutLayer((2, 4, 7, 9), shared_axes=shared_axes)
input = theano.shared(numpy.ones((2, 4, 7, 9)))
result = layer.get_output_for(input)
result_eval = result.eval()
# check if the dropout mask is the same across the specified axes:
# compute the mean across these axes and compare against the full
# output, broadcasting across the shared axes, to see if it matches
assert np.allclose(result_eval.mean(axis=shared_axes, keepdims=True),
result_eval)
def test_get_output_for_p_float16(self, input_layer):
from lasagne.layers.noise import DropoutLayer
layer = DropoutLayer(input_layer, p=numpy.float16(0.5))
input = theano.shared(numpy.ones((100, 100), dtype=numpy.float16))
assert layer.get_output_for(input).dtype == input.dtype
def test_get_output_for_p_float32(self, input_layer):
from lasagne.layers.noise import DropoutLayer
layer = DropoutLayer(input_layer, p=numpy.float32(0.5))
input = theano.shared(numpy.ones((100, 100), dtype=numpy.float32))
assert layer.get_output_for(input).dtype == input.dtype
