from treeano.sandbox.nodes import activation_transformation
concat_node = activation_transformation.ConcatenateNegationNode
num_filters_1 = 96
num_filters_2 = 192
# based off of architecture from "Scalable Bayesian Optimization Using
# Deep Neural Networks" http://arxiv.org/abs/1502.05700
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(None, 3, 32, 32)),
tn.DnnConv2DWithBiasNode("conv1", num_filters=num_filters_1),
concat_node("concat1"),
tn.ReLUNode("relu1"),
tn.DropoutNode("do1"),
tn.DnnConv2DWithBiasNode("conv2", num_filters=num_filters_1),
concat_node("concat2"),
tn.ReLUNode("relu2"),
tn.DropoutNode("do2"),
tn.MaxPool2DNode("mp1"),
tn.DnnConv2DWithBiasNode("conv3", num_filters=num_filters_2),
concat_node("concat3"),
tn.ReLUNode("relu3"),
tn.DropoutNode("do3"),
tn.DnnConv2DWithBiasNode("conv4", num_filters=num_filters_2),
concat_node("concat4"),
tn.ReLUNode("relu4"),
tn.DropoutNode("do4"),
tn.DnnConv2DWithBiasNode("conv5", num_filters=num_filters_2),
concat_node("concat5"),
# - the batch size can be provided as `None` to make the network
# work for multiple different batch sizes
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(None, 1, 28, 28)),
tn.DnnConv2DWithBiasNode("conv1"),
tn.ReLUNode("relu1"),
tn.DnnMaxPoolNode("mp1"),
tn.DnnConv2DWithBiasNode("conv2"),
tn.ReLUNode("relu2"),
tn.DnnMaxPoolNode("mp2"),
tn.DenseNode("fc1"),
tn.ReLUNode("relu3"),
tn.DropoutNode("do1"),
tn.DenseNode("fc2", num_units=10),
tn.SoftmaxNode("pred"),
]),
num_filters=32,
filter_size=(5, 5),
pool_size=(2, 2),
num_units=256,
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
"adam", {
def make_network(p):
return tn.SequentialNode(
"s",
[tn.InputNode("i", shape=(3, 4, 5)),
tn.DropoutNode("do", p=p)]).network()
# ############################### prepare data ###############################
train, valid, test = canopy.sandbox.datasets.mnist()
# ############################## prepare model ##############################
BATCH_SIZE = 500
NUM_EPOCHS = 25
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(None, 1, 28, 28)),
tn.DenseNode("fc1"),
tn.ReLUNode("relu1"),
eb.GradualDropoutNode("gdo1", tn.DropoutNode("do1")),
tn.DenseNode("fc2"),
tn.ReLUNode("relu2"),
eb.GradualDropoutNode("gdo2", tn.DropoutNode("do2")),
tn.DenseNode("fc3", num_units=10),
tn.SoftmaxNode("pred"),
]),
num_units=512,
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
expected_batches=NUM_EPOCHS * len(train["x"]) / BATCH_SIZE,
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
def test_dropout_node_serialization():
tn.check_serialization(tn.DropoutNode("a"))
tn.check_serialization(tn.DropoutNode("a", p=0.5))