def test_FlowNode_invertibility():
flow = mdp.Flow([mdp.nodes.PolynomialExpansionNode(degree=2)])
flownode = mh.FlowNode(flow)
assert flownode.is_invertible() is False
flow = mdp.Flow([
mdp.nodes.PCANode(output_dim=15),
mdp.nodes.SFANode(),
mdp.nodes.PCANode(output_dim=3)
])
flownode = mh.FlowNode(flow)
assert flownode.is_invertible() is True
def test_FlowNode_pretrained_flow():
flow = mdp.Flow([
mdp.nodes.PolynomialExpansionNode(degree=2),
mdp.nodes.PCANode(output_dim=15, reduce=True),
mdp.nodes.PolynomialExpansionNode(degree=2),
mdp.nodes.PCANode(output_dim=3, reduce=True)
])
flownode = mh.FlowNode(flow)
x = numx_rand.random([300, 20])
while flownode.get_remaining_train_phase() > 0:
flownode.train(x)
flownode.stop_training()
# build new flownode with the trained nodes
flownode = mh.FlowNode(flow)
assert not flownode.is_training()
flownode.execute(x)
def test_FlowNode_fix_nodes_dimensions3():
flow = mdp.Flow([mdp.nodes.IdentityNode()])
flownode = mh.FlowNode(flow)
# for a single node this should not raise an Exception
flownode.set_output_dim(10)
x = numx_rand.random([100, 10])
flownode.execute(x)
def test_FlowNode_fix_nodes_dimensions2():
flow = mdp.Flow([mdp.nodes.IdentityNode(), mdp.nodes.IdentityNode()])
flownode = mh.FlowNode(flow)
# this should fail, since the internal nodes don't have fixed dims
py.test.raises(mdp.InconsistentDimException,
lambda: flownode.set_output_dim(10))
x = numx_rand.random([100, 10])
flownode.execute(x)
assert flownode.output_dim == 10
def test_SFA_net(noisenode):
sfa_node = mdp.nodes.SFANode(input_dim=20 * 20, output_dim=10, dtype='f')
switchboard = mh.Rectangular2dSwitchboard(in_channels_xy=100,
field_channels_xy=20,
field_spacing_xy=10)
flownode = mh.FlowNode(mdp.Flow([noisenode, sfa_node]))
sfa_layer = mh.CloneLayer(flownode, switchboard.output_channels)
flow = mdp.Flow([switchboard, sfa_layer])
train_gen = numx.cast['f'](numx_rand.random((3, 10, 100 * 100)))
flow.train([None, train_gen])
def test_FlowNode_copy2():
# Test that the FlowNode copy method delegates to internal nodes.
class CopyFailException(Exception):
pass
class CopyFailNode(mdp.Node):
def copy(self, protocol=None):
raise CopyFailException()
flow = mdp.Flow([mdp.Node(), CopyFailNode()])
flownode = mh.FlowNode(flow)
py.test.raises(CopyFailException, flownode.copy)
def test_flownode_training():
flow = mdp.Flow([
mdp.nodes.PolynomialExpansionNode(degree=2),
mdp.nodes.PCANode(output_dim=15, reduce=True),
mdp.nodes.PolynomialExpansionNode(degree=2),
mdp.nodes.PCANode(output_dim=3, reduce=True)
])
flownode = mh.FlowNode(flow)
x = numx_rand.random([300, 20])
while flownode.get_remaining_train_phase() > 0:
flownode.train(x)
flownode.stop_training()
flownode.execute(x)
def test_FlowNode_fix_nodes_dimensions1():
x = numx_rand.random([100, 10])
last_node = mdp.nodes.IdentityNode()
flow = mdp.Flow(
[mdp.nodes.PCANode(output_dim=3),
mdp.nodes.IdentityNode(), last_node])
flownode = mh.FlowNode(flow)
flownode.train(x)
flownode.stop_training()
# check that the dimensions of NoiseNode and FlowNode where all set
# by calling _fix_nodes_dimensions
assert flownode.output_dim == 3
assert last_node.input_dim == 3
assert last_node.output_dim == 3
def testHiNetHTML(noisenode):
# create some flow for testing
sfa_node = mdp.nodes.SFANode(input_dim=20 * 20, output_dim=10)
switchboard = mh.Rectangular2dSwitchboard(in_channels_xy=100,
field_channels_xy=20,
field_spacing_xy=10)
flownode = mh.FlowNode(mdp.Flow([noisenode, sfa_node]))
sfa_layer = mh.CloneLayer(flownode, switchboard.output_channels)
flow = mdp.Flow([switchboard, sfa_layer])
# create dummy file to write the HTML representation
html_file = io.StringIO()
hinet_html = mdp.hinet.HiNetHTMLVisitor(html_file)
hinet_html.convert_flow(flow)
html_file.close()
def test_hinet_simple_net():
switchboard = mh.Rectangular2dSwitchboard(in_channels_xy=(12, 8),
field_channels_xy=4,
field_spacing_xy=2,
in_channel_dim=3)
node = mdp.nodes.PCANode(input_dim=4 * 4 * 3, output_dim=5)
flownode = mh.FlowNode(mdp.Flow([
node,
]))
layer = mh.CloneLayer(flownode, switchboard.output_channels)
flow = mdp.Flow([switchboard, layer])
x = numx_rand.random([5, switchboard.input_dim])
flow.train(x)
def test_FlowNode_pretrained_node():
x = numx_rand.random([100, 10])
pretrained_node = mdp.nodes.PCANode(output_dim=6)
pretrained_node.train(x)
pretrained_node.stop_training()
flow = mdp.Flow([
pretrained_node,
mdp.nodes.PolynomialExpansionNode(degree=2),
mdp.nodes.PCANode(output_dim=3)
])
flownode = mh.FlowNode(flow)
while flownode.get_remaining_train_phase() > 0:
flownode.train(x)
flownode.stop_training()
flownode.execute(x)
def test_FlowNode_copy1():
flow = mdp.Flow([mdp.nodes.PCANode(), mdp.nodes.SFANode()])
flownode = mh.FlowNode(flow)
flownode.copy()