def test_dense_node_and_dense_combine_node1():
# testing that dense node and dense combine node with identity child
# return the same thing
network1 = tn.HyperparameterNode("hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.DenseNode("fc1", num_units=6),
tn.DenseNode("fc2", num_units=7),
tn.DenseNode("fc3", num_units=8)
]),
inits=[treeano.inits.ConstantInit(1)
]).network()
network2 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.DenseCombineNode("fc1", [tn.IdentityNode("i1")], num_units=6),
tn.DenseCombineNode("fc2", [tn.IdentityNode("i2")], num_units=7),
tn.DenseCombineNode("fc3", [tn.IdentityNode("i3")], num_units=8)
]),
inits=[treeano.inits.ConstantInit(1)]).network()
x = np.random.randn(3, 4, 5).astype(fX)
fn1 = network1.function(["in"], ["fc3"])
fn2 = network2.function(["in"], ["fc3"])
np.testing.assert_allclose(fn1(x), fn2(x))
def test_dense_combine_node_uses_children():
network1 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.MultiplyConstantNode("mul", value=2),
tn.DenseCombineNode("fc",
[tn.IdentityNode("i1"),
tn.IdentityNode("i2")],
num_units=6)
]),
inits=[treeano.inits.ConstantInit(1)]).network()
network2 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.DenseCombineNode("fc", [
tn.MultiplyConstantNode("mul1", value=2),
tn.MultiplyConstantNode("mul2", value=2)
],
num_units=6)
]),
inits=[treeano.inits.ConstantInit(1)]).network()
x = np.random.randn(3, 4, 5).astype(fX)
fn1 = network1.function(["in"], ["hp"])
fn2 = network2.function(["in"], ["hp"])
np.testing.assert_allclose(fn1(x), fn2(x))
def test_postwalk_node():
names = []
def f(node):
names.append(node.name)
return node
node = tn.HyperparameterNode(
"1", tn.HyperparameterNode("2", tn.IdentityNode("3")))
canopy.node_utils.postwalk_node(node, f)
nt.assert_equal(names, ["3", "2", "1"])
def test_remove_parents():
network1 = tn.SequentialNode("seq", [
tn.InputNode("i", shape=()),
tn.HyperparameterNode("hp1",
tn.HyperparameterNode(
"hp2", tn.AddConstantNode("ac"), value=1),
value=2)
]).network()
network2 = canopy.transforms.remove_parents(network1, "ac")
nt.assert_equal(tn.AddConstantNode("ac"), network2.root_node)
def test_affine_spatial_transformer_node_build():
localization_network = tn.HyperparameterNode(
"loc",
tn.SequentialNode(
"loc_seq",
[tn.DenseNode("loc_fc1", num_units=50),
tn.ReLUNode("loc_relu3"),
tn.DenseNode("loc_fc2",
num_units=6,
inits=[treeano.inits.ZeroInit()])]),
num_filters=32,
filter_size=(5, 5),
pool_size=(2, 2),
)
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[tn.InputNode("x", shape=(None, 1, 60, 60)),
spatial_transformer.AffineSpatialTransformerNode(
"st",
localization_network,
output_shape=(20, 20)),
tn.DenseNode("fc1"),
tn.ReLUNode("relu1"),
tn.DropoutNode("do1"),
tn.DenseNode("fc2", num_units=10),
tn.SoftmaxNode("pred"),
]),
num_filters=32,
filter_size=(3, 3),
pool_size=(2, 2),
num_units=256,
dropout_probability=0.5,
inits=[treeano.inits.HeNormalInit()],
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
"adam",
{"subtree": model,
"cost": tn.TotalCostNode("cost", {
"pred": tn.ReferenceNode("pred_ref", reference="model"),
"target": tn.InputNode("y", shape=(None,), dtype="int32")},
)}),
cost_function=treeano.utils.categorical_crossentropy_i32,
)
network = with_updates.network()
network.build() # build eagerly to share weights
def test_auxiliary_cost_node():
network = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(3, 4, 5)),
tn.AuxiliaryCostNode(
"cost1", {"target": tn.InputNode("y1", shape=(3, 4, 5))}),
tn.AddConstantNode("a1", value=2),
tn.AuxiliaryCostNode(
"cost2", {"target": tn.InputNode("y2", shape=(3, 4, 5))}),
tn.MultiplyConstantNode("m1", value=2),
tn.AuxiliaryCostNode(
"cost3", {"target": tn.InputNode("y3", shape=(3, 4, 5))}),
tn.ConstantNode("const", value=0),
tn.InputElementwiseSumNode("cost")
]),
cost_reference="cost",
cost_function=treeano.utils.squared_error,
).network()
fn = network.function(["x", "y1", "y2", "y3"], ["cost"])
x = np.random.rand(3, 4, 5).astype(fX)
ys = [np.random.rand(3, 4, 5).astype(fX) for _ in range(3)]
def mse(x, y):
return ((x - y)**2).mean()
expected_output = (mse(x, ys[0]) + mse(x + 2, ys[1]) +
mse(2 * (x + 2), ys[2]))
np.testing.assert_allclose(fn(x, *ys)[0], expected_output, rtol=1e-5)
def axes(ndim, pos, neg):
network = tn.HyperparameterNode(
"a",
tn.InputNode("b", shape=()),
pos=pos,
neg=neg,
).network()["a"]
return treeano.utils.find_axes(network, ndim, ["pos"], ["neg"])
def test_remove_nodes():
network1 = tn.SequentialNode("seq", [
tn.InputNode("i", shape=()),
tn.HyperparameterNode("hp1",
tn.HyperparameterNode(
"hp2", tn.AddConstantNode("ac"), value=1),
value=2)
]).network()
fn1 = network1.function(["i"], ["seq"])
nt.assert_equal(1, fn1(0)[0])
network2 = canopy.transforms.remove_nodes(network1, {"hp2"},
keep_child=True)
fn2 = network2.function(["i"], ["seq"])
nt.assert_equal(2, fn2(0)[0])
network3 = canopy.transforms.remove_nodes(network1, {"ac"})
fn3 = network3.function(["i"], ["seq"])
nt.assert_equal(0, fn3(0)[0])
def GradualBatchNormalization(name, **kwargs):
from treeano.sandbox.nodes import batch_normalization as bn
return tn.HyperparameterNode(
name,
LinearInterpolationNode(
name + "_interpolate", {
"early": bn.BatchNormalizationNode(name + "_bn"),
"late": tn.IdentityNode(name + "_identity")
}), **kwargs)
def AverageSamplesDropoutDnnMaxPoolNode(name, *args, **kwargs):
return tn.HyperparameterNode(
name,
AverageSamplesNode(
name + "_samples",
tn.SequentialNode(
name + "_seq",
[tn.DropoutNode(name + "_dropout"),
tn.DnnMaxPoolNode(name + "_maxpool")])),
*args,
**kwargs)
def HighwayDenseNode(name, nonlinearity_node, **hyperparameters):
return tn.HyperparameterNode(
name,
HighwayNode(
name + "_highway", {
"transform":
tn.SequentialNode(name + "_transform", [
tn.DenseNode(name + "_transformdense"), nonlinearity_node
]),
"gate":
tn.DenseNode(name + "_gatedense")
}), **hyperparameters)
def test_dense_node_and_dense_combine_node2():
# testing that summing the output of 2 dense nodes is the same as
# applying a dense combine node with 2 identities (+ bias)
# and the same as multiplying the output of 1 dense node by 2
network0 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.DenseNode("dense1", num_units=6),
tn.MultiplyConstantNode("mul", value=2)
]),
inits=[treeano.inits.ConstantInit(1)]).network()
network1 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.ElementwiseSumNode("sum", [
tn.DenseNode("dense1", num_units=6),
tn.DenseNode("dense2", num_units=6)
])
]),
inits=[treeano.inits.ConstantInit(1)]).network()
network2 = tn.HyperparameterNode(
"hp",
tn.SequentialNode("seq", [
tn.InputNode("in", shape=(3, 4, 5)),
tn.DenseCombineNode("fc",
[tn.IdentityNode("i1"),
tn.IdentityNode("i2")],
num_units=6),
tn.AddBiasNode("bias")
]),
inits=[treeano.inits.ConstantInit(1)]).network()
x = np.random.randn(3, 4, 5).astype(fX)
fn0 = network0.function(["in"], ["hp"])
fn1 = network1.function(["in"], ["hp"])
fn2 = network2.function(["in"], ["hp"])
np.testing.assert_allclose(fn0(x), fn1(x))
np.testing.assert_allclose(fn0(x), fn2(x))
def test_remove_parent():
network1 = tn.SequentialNode("seq", [
tn.InputNode("i", shape=()),
tn.HyperparameterNode("hp1",
tn.HyperparameterNode(
"hp2", tn.AddConstantNode("ac"), value=1),
value=2)
]).network()
fn1 = network1.function(["i"], ["seq"])
nt.assert_equal(1, fn1(0)[0])
network2 = canopy.transforms.remove_parent(network1, {"ac"})
fn2 = network2.function(["i"], ["seq"])
nt.assert_equal(2, fn2(0)[0])
network3 = canopy.transforms.remove_parent(network1, {"i"})
@nt.raises(Exception)
def fails(name):
network3.function(["i"], [name])
# testing that these nodes are removed
fails("ac")
fails("seq")
network3.function(["i"], ["i"])
def MultiPool2DNode(name, **kwargs):
# TODO tests
# TODO make a node that verifies hyperparameters
return tn.HyperparameterNode(
name,
tn.ConcatenateNode(name + "_concat", [
tn.SequentialNode(name + "_seq0", [
PartitionAxisNode(name + "_part0", split_idx=0, num_splits=2),
tn.MaxPool2DNode(name + "_max", ignore_border=True)
]),
tn.SequentialNode(name + "_seq1", [
PartitionAxisNode(name + "_part1", split_idx=1, num_splits=2),
tn.MeanPool2DNode(name + "_mean")
])
]), **kwargs)
def test_scale_hyperparameter():
network = tn.HyperparameterNode(
"hp",
eb.ScaleHyperparameterNode("scale", tn.ConstantNode("c")),
value=42.0,
hyperparameter="value",
start_percent=0.,
end_percent=1.0,
start_scale=1.0,
end_scale=0.1,
expected_batches=2,
).network()
fn = network.function([], ["c"], include_updates=True)
np.testing.assert_allclose(42.0, fn()[0], rtol=1e-5)
np.testing.assert_allclose(42.0 * 0.55, fn()[0], rtol=1e-5)
np.testing.assert_allclose(42.0 * 0.1, fn()[0], rtol=1e-5)
np.testing.assert_allclose(42.0 * 0.1, fn()[0], rtol=1e-5)
def test_auxiliary_kl_sparsity_penalty_node():
# testing that both sparsity penalty versions return the same thing
network = tn.HyperparameterNode(
"hp",
tn.SequentialNode(
"s",
[
tn.InputNode("i", shape=(10, 3)),
tn.DenseNode("d", num_units=9),
sp.AuxiliaryKLSparsityPenaltyNode("scp", cost_reference="sum"),
sp.ElementwiseKLSparsityPenaltyNode("sp"),
tn.AggregatorNode("a"),
# zero out rest of network, so that value of sum is just the value
# from auxiliary sparsity pentalty node
tn.ConstantNode("foo", value=0),
tn.InputElementwiseSumNode("sum")
]),
sparsity=0.1,
).network()
fn = network.function(["i"], ["sum", "a"])
x = np.random.rand(10, 3).astype(fX)
res = fn(x)
np.testing.assert_equal(res[0], res[1])
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[
tn.InputNode("x", shape=(None, 1, 28, 28)),
tn.Conv2DWithBiasNode("conv1"),
# bn.BatchNormalizationNode("bn1"),
timesout.IndexedTimesoutNode("to1"),
tn.TanhNode("tanh1"),
tn.MaxPool2DNode("mp1"),
tn.Conv2DWithBiasNode("conv2"),
# bn.BatchNormalizationNode("bn2"),
timesout.IndexedTimesoutNode("to2"),
tn.TanhNode("tanh2"),
tn.MaxPool2DNode("mp2"),
tn.DenseNode("fc1"),
# bn.BatchNormalizationNode("bn3"),
timesout.IndexedTimesoutNode("to3"),
tn.TanhNode("tanh3"),
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,
num_pieces=2,
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
)
def vgg_16_nodes(conv_only):
"""
conv_only:
whether or not to only return conv layers (before FC layers)
"""
assert conv_only
return tn.HyperparameterNode(
"vgg16",
tn.SequentialNode(
"vgg16_seq",
[
tn.HyperparameterNode(
"conv_group_1",
tn.SequentialNode("conv_group_1_seq", [
tn.DnnConv2DWithBiasNode("conv1_1"),
tn.ReLUNode("relu1_1"),
tn.DnnConv2DWithBiasNode("conv1_2"),
tn.ReLUNode("relu1_2")
]),
num_filters=64),
tn.MaxPool2DNode("pool1"),
tn.HyperparameterNode(
"conv_group_2",
tn.SequentialNode("conv_group_2_seq", [
tn.DnnConv2DWithBiasNode("conv2_1"),
tn.ReLUNode("relu2_1"),
tn.DnnConv2DWithBiasNode("conv2_2"),
tn.ReLUNode("relu2_2")
]),
num_filters=128),
tn.MaxPool2DNode("pool2"),
tn.HyperparameterNode(
"conv_group_3",
tn.SequentialNode("conv_group_3_seq", [
tn.DnnConv2DWithBiasNode("conv3_1"),
tn.ReLUNode("relu3_1"),
tn.DnnConv2DWithBiasNode("conv3_2"),
tn.ReLUNode("relu3_2"),
tn.DnnConv2DWithBiasNode("conv3_3"),
tn.ReLUNode("relu3_3")
]),
num_filters=256),
tn.MaxPool2DNode("pool3"),
tn.HyperparameterNode(
"conv_group_4",
tn.SequentialNode("conv_group_4_seq", [
tn.DnnConv2DWithBiasNode("conv4_1"),
tn.ReLUNode("relu4_1"),
tn.DnnConv2DWithBiasNode("conv4_2"),
tn.ReLUNode("relu4_2"),
tn.DnnConv2DWithBiasNode("conv4_3"),
tn.ReLUNode("relu4_3")
]),
num_filters=512),
tn.MaxPool2DNode("pool4"),
tn.HyperparameterNode(
"conv_group_5",
tn.SequentialNode("conv_group_5_seq", [
tn.DnnConv2DWithBiasNode("conv5_1"),
tn.ReLUNode("relu5_1"),
tn.DnnConv2DWithBiasNode("conv5_2"),
tn.ReLUNode("relu5_2"),
tn.DnnConv2DWithBiasNode("conv5_3"),
tn.ReLUNode("relu5_3")
]),
num_filters=512),
tn.MaxPool2DNode("pool5"),
# TODO add dense nodes
]),
pad="same",
filter_size=(3, 3),
pool_size=(2, 2),
# VGG net uses cross-correlation by default
conv_mode="cross",
)
# - ReLU
# - 50% dropout
# - fully connected 10 units
# - softmax
# - 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.DenseNode("fc1"),
tn.ReLUNode("relu1"),
tn.DropoutNode("do1"),
tn.DenseNode("fc2"),
tn.ReLUNode("relu2"),
tn.DropoutNode("do2"),
tn.DenseNode("fc3", num_units=10),
tn.SoftmaxNode("pred"),
]),
num_units=512,
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
"adam", {
"subtree":
model,
def test_suffix_node():
node1 = tn.HyperparameterNode(
"1", tn.HyperparameterNode("2", tn.IdentityNode("3")))
node2 = tn.HyperparameterNode(
"1_foo", tn.HyperparameterNode("2_foo", tn.IdentityNode("3_foo")))
nt.assert_equal(canopy.node_utils.suffix_node(node1, "_foo"), node2)
nodes.append(
resnet.residual_block_conv_2d("resblock_%d_%d" %
(group, block),
num_filters=num_filters,
num_layers=num_layers))
nodes += [
tn.GlobalMeanPool2DNode("global_pool"),
tn.DenseNode("logit", num_units=10),
tn.SoftmaxNode("pred"),
]
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", nodes),
filter_size=(3, 3),
inits=[treeano.inits.OrthogonalInit()],
pad="same",
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
"adam", {
"subtree":
model,
"cost":
tn.TotalCostNode(
"cost",
{
"pred": tn.ReferenceNode("pred_ref", reference="model"),
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[
tn.InputNode("x", shape=(None, 28 * 28)),
tn.DenseNode("fc1"),
# nbn.GradualBatchToNoBatchNormalizationNode("bn1"),
nbn.NoBatchNormalizationNode("bn1"),
# bn.BatchNormalizationNode("bn1"),
tn.ReLUNode("relu1"),
# tn.DropoutNode("do2", p=0.5),
tn.DenseNode("fc2"),
# nbn.GradualBatchToNoBatchNormalizationNode("bn2"),
nbn.NoBatchNormalizationNode("bn2"),
# bn.BatchNormalizationNode("bn2"),
tn.ReLUNode("relu2"),
# tn.DropoutNode("do3", p=0.5),
tn.DenseNode("fc3", num_units=10),
# nbn.GradualBatchToNoBatchNormalizationNode("bn3"),
# nbn.NoBatchNormalizationNode("bn3"),
# bn.BatchNormalizationNode("bn3"),
tn.SoftmaxNode("pred"),
]),
num_units=512,
inits=[treeano.inits.XavierNormalInit()],
current_mean_weight=1. / 8,
current_var_weight=1. / 8,
rolling_mean_rate=0.99,
rolling_var_rate=0.99,
expected_batches=25 * len(X_train) / BATCH_SIZE,
)
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[tn.InputNode("x", shape=(None, 28 * 28)),
cp.AuxiliaryContractionPenaltyNode(
"cp1",
tn.SequentialNode(
"cp_seq1",
[tn.DenseNode("fc1"),
# the cost has nan's when using ReLU's
# TODO look into why
tn.AbsNode("abs1")]),
cost_weight=1e1),
# the cost has nan's when this is enabled
# TODO look into why
# tn.DropoutNode("do1"),
cp.AuxiliaryContractionPenaltyNode(
"cp2",
tn.SequentialNode(
"cp_seq2",
[tn.DenseNode("fc2"),
# the cost has nan's when using ReLU's
# TODO look into why
tn.AbsNode("abs2")]),
cost_weight=1e1),
tn.DropoutNode("do2"),
tn.DenseNode("fc3", num_units=10),
tn.SoftmaxNode("pred"),
tn.TotalCostNode(
"cost",
{"pred": tn.IdentityNode("pred_id"),
"target": tn.InputNode("y", shape=(None,), dtype="int32")},
cost_function=treeano.utils.categorical_crossentropy_i32),
tn.InputElementwiseSumNode("total_cost")]),
num_units=32,
cost_reference="total_cost",
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
)
train, valid, test = canopy.sandbox.datasets.mnist()
# ############################## prepare model ##############################
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[tn.InputNode("x", shape=(None, 1, 28, 28)),
tn.Conv2DWithBiasNode("conv1"),
tn.ReLUNode("relu1"),
dropout_max_pool.AverageSamplesDropoutDnnMaxPoolNode("mp1"),
tn.Conv2DWithBiasNode("conv2"),
tn.ReLUNode("relu2"),
dropout_max_pool.AverageSamplesDropoutDnnMaxPoolNode("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(
in_valid = {"x": X_valid, "y": y_valid}
# ############################## prepare model ##############################
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(None, 1, 28, 28)),
inception.InceptionNode("i1"),
tn.DnnMaxPoolNode("mp1"),
bn.BatchNormalizationNode("bn1"),
inception.InceptionNode("i2"),
tn.DnnMaxPoolNode("mp2"),
bn.BatchNormalizationNode("bn2"),
tn.DenseNode("fc1"),
tn.ReLUNode("relu3"),
tn.DenseNode("fc2", num_units=10),
tn.SoftmaxNode("pred"),
]),
num_filters_1x1=32,
num_filters_3x3reduce=16,
num_filters_3x3=32,
num_filters_5x5reduce=16,
num_filters_5x5=32,
num_filters_poolproj=32,
pool_size=(2, 2),
num_units=32,
inits=[treeano.inits.XavierNormalInit()],
)
with_updates = tn.HyperparameterNode(
"with_updates",
# ############################## prepare model ##############################
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(None, 28 * 28)),
tn.DenseNode("fc1"),
tn.SigmoidNode("sigmoid1"),
sp.AuxiliaryKLSparsityPenaltyNode("sp1", cost_weight=1e1),
tn.DropoutNode("do1"),
tn.DenseNode("fc2"),
tn.SigmoidNode("sigmoid2"),
sp.AuxiliaryKLSparsityPenaltyNode("sp2", cost_weight=1e1),
tn.DropoutNode("do2"),
tn.DenseNode("fc3", num_units=10),
tn.SoftmaxNode("pred"),
tn.TotalCostNode(
"cost", {
"pred": tn.IdentityNode("pred_id"),
"target": tn.InputNode("y", shape=(None, ), dtype="int32")
},
cost_function=treeano.utils.categorical_crossentropy_i32),
tn.InputElementwiseSumNode("total_cost")
]),
num_units=512,
sparsity=0.1,
cost_reference="total_cost",
dropout_probability=0.5,
inits=[treeano.inits.XavierNormalInit()],
)
with_updates = tn.HyperparameterNode(
# ############################## 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"),
eb.GradualBatchNormalization("bn1"),
tn.ReLUNode("relu1"),
tn.DenseNode("fc2"),
eb.GradualBatchNormalization("bn2"),
tn.ReLUNode("relu2"),
tn.DenseNode("fc3", num_units=10),
eb.GradualBatchNormalization("bn3"),
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(
"adam", {
"subtree":
train, valid, _ = canopy.sandbox.datasets.cluttered_mnist()
# ############################## prepare model ##############################
localization_network = tn.HyperparameterNode(
"loc",
tn.SequentialNode("loc_seq", [
tn.DnnMaxPoolNode("loc_pool1"),
tn.DnnConv2DWithBiasNode("loc_conv1"),
tn.DnnMaxPoolNode("loc_pool2"),
bn.NoScaleBatchNormalizationNode("loc_bn1"),
tn.ReLUNode("loc_relu1"),
tn.DnnConv2DWithBiasNode("loc_conv2"),
bn.SimpleBatchNormalizationNode("loc_bn2"),
tn.SpatialSoftmaxNode("loc_spatial_softmax"),
spatial_attention.SpatialFeaturePointNode("loc_feature_point"),
tn.DenseNode("loc_fc1", num_units=50),
bn.NoScaleBatchNormalizationNode("loc_bn3"),
tn.ReLUNode("loc_relu3"),
tn.DenseNode("loc_fc2",
num_units=6,
inits=[treeano.inits.NormalWeightInit(std=0.001)])
]),
num_filters=20,
filter_size=(5, 5),
pool_size=(2, 2),
)
st_node = st.AffineSpatialTransformerNode("st",
localization_network,
output_shape=(20, 20))
model = tn.HyperparameterNode(
"model",
tn.SequentialNode("seq", [
tn.InputNode("x", shape=(BATCH_SIZE, 3, 32, 32)),
tn.DnnConv2DWithBiasNode("conv1", num_filters=96),
tn.ReLUNode("relu1"),
tn.DnnConv2DWithBiasNode("conv2", num_filters=96),
tn.ReLUNode("relu2"),
tn.MaxPool2DNode("mp1"),
tn.DropoutNode("do1", dropout_probability=0.1),
tn.DnnConv2DWithBiasNode("conv3", num_filters=192),
tn.ReLUNode("relu3"),
tn.DnnConv2DWithBiasNode("conv4", num_filters=192),
tn.ReLUNode("relu4"),
tn.DnnConv2DWithBiasNode("conv5", num_filters=192),
tn.ReLUNode("relu5"),
tn.MaxPool2DNode("mp2"),
tn.DropoutNode("do2", dropout_probability=0.5),
tn.DnnConv2DWithBiasNode("conv6", num_filters=192),
tn.ReLUNode("relu6"),
tn.DnnConv2DWithBiasNode("conv7", num_filters=192, filter_size=(1, 1)),
tn.ReLUNode("relu7"),
tn.DnnConv2DWithBiasNode("conv8", num_filters=10, filter_size=(1, 1)),
tn.GlobalMeanPool2DNode("mean_pool"),
tn.SoftmaxNode("pred"),
]),
filter_size=(3, 3),
conv_pad="same",
pool_size=(3, 3),
pool_stride=(2, 2),
pool_pad=(1, 1),
inits=[treeano.inits.OrthogonalInit()],
)
i, o = binary_toy_data(lag, length)
inputs.append(i)
outputs.append(o)
return np.array(inputs)[..., np.newaxis], np.array(outputs)[..., np.newaxis]
# ############################## prepare model ##############################
model = tn.HyperparameterNode(
"model",
tn.SequentialNode(
"seq",
[tn.InputNode("x", shape=(None, None, 1)),
recurrent_hc.GRUNode("gru1"),
tn.LinearMappingNode("y_linear", output_dim=1),
tn.AddBiasNode("y_bias", broadcastable_axes=(0, 1)),
tn.SigmoidNode("sigmoid"),
]),
inits=[treeano.inits.OrthogonalInit()],
num_units=HIDDEN_STATE_SIZE,
learn_init=True,
grad_clip=1,
)
with_updates = tn.HyperparameterNode(
"with_updates",
tn.AdamNode(
"adam",
{"subtree": model,
"cost": tn.TotalCostNode("cost", {
"pred": tn.ReferenceNode("pred_ref", reference="model"),
