def test_not_implemented(self):
try:
from lasagne.layers.cuda_convnet import MaxPool2DCCLayer
except ImportError:
pytest.skip("cuda_convnet not available")
input_layer = self.input_layer((128, 4, 12, 12))
with pytest.raises(RuntimeError) as exc:
layer = MaxPool2DCCLayer(input_layer, pool_size=2, pad=2)
assert "MaxPool2DCCLayer does not support padding" in exc.value.args[0]
with pytest.raises(RuntimeError) as exc:
layer = MaxPool2DCCLayer(input_layer, pool_size=(2, 3))
assert ("MaxPool2DCCLayer only supports square pooling regions"
in exc.value.args[0])
with pytest.raises(RuntimeError) as exc:
layer = MaxPool2DCCLayer(input_layer, pool_size=2, stride=(1, 2))
assert (("MaxPool2DCCLayer only supports using the same stride in "
"both directions") in exc.value.args[0])
with pytest.raises(RuntimeError) as exc:
layer = MaxPool2DCCLayer(input_layer, pool_size=2, stride=3)
assert ("MaxPool2DCCLayer only supports stride <= pool_size"
in exc.value.args[0])
with pytest.raises(RuntimeError) as exc:
layer = MaxPool2DCCLayer(input_layer,
pool_size=2,
ignore_border=True)
assert ("MaxPool2DCCLayer does not support ignore_border"
in exc.value.args[0])
def build_model(input_width, input_height, output_dim, batch_size=BATCH_SIZE):
l_in = lasagne.layers.InputLayer(shape=(batch_size, 1, input_width,
input_height), )
l_in_c01b = bc01_to_c01b(l_in)
l_conv1 = layers.NervanaConvLayer(
l_in_c01b,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
dimshuffle=False,
)
l_pool1 = MaxPool2DCCLayer(l_conv1, ds=(2, 2), dimshuffle=False)
l_conv2 = layers.NervanaConvLayer(
l_pool1,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
dimshuffle=False,
)
l_pool2 = MaxPool2DCCLayer(l_conv2, ds=(2, 2), dimshuffle=False)
l_pool2_bc01 = c01b_to_bc01(l_pool2)
l_hidden1 = lasagne.layers.DenseLayer(
l_pool2_bc01,
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
)
l_hidden1_dropout = lasagne.layers.DropoutLayer(l_hidden1, p=0.5)
# l_hidden2 = lasagne.layers.DenseLayer(
# l_hidden1_dropout,
# num_units=256,
# nonlinearity=lasagne.nonlinearities.rectify,
# )
# l_hidden2_dropout = lasagne.layers.DropoutLayer(l_hidden2, p=0.5)
l_out = lasagne.layers.DenseLayer(
l_hidden1_dropout,
num_units=output_dim,
nonlinearity=lasagne.nonlinearities.softmax,
W=lasagne.init.GlorotUniform(),
)
return l_out
def layer(self, input_layer, pool_size, stride):
try:
from lasagne.layers.cuda_convnet import MaxPool2DCCLayer
except ImportError:
pytest.skip("cuda_convnet not available")
return MaxPool2DCCLayer(
input_layer,
pool_size=pool_size,
stride=stride,
)
def test_dimshuffle_false(self):
try:
from lasagne.layers.cuda_convnet import MaxPool2DCCLayer
except ImportError:
pytest.skip("cuda_convnet not available")
from lasagne.layers.input import InputLayer
input_layer = InputLayer((4, 12, 12, 16)) # c01b order
layer = MaxPool2DCCLayer(input_layer, pool_size=2, dimshuffle=False)
assert layer.output_shape == (4, 6, 6, 16)
input = floatX(np.random.randn(4, 12, 12, 16))
output = max_pool_2d(input.transpose(3, 0, 1, 2), (2, 2), (2, 2))
output = output.transpose(1, 2, 3, 0)
actual = layer.get_output_for(input).eval()
assert np.allclose(output, actual)
