def test_raises_on_malformed(self):
# simple smoke test from the example file
tree1 = ((0, 1), ((1, 2), ((0, 1), ), ((-1, 0), )),
((-3, 0), ((2, 3), ), ((1, 2), )))
tree2 = ((16, 3, 2), ((0, 1), ((5, 3), ), ((2, 6), )), ((2, 9), ))
trees = [tree1, tree2]
# function to extract the left child of a node
def left_child(x):
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[1]
# function to extract the right child of node
def right_child(x):
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[2]
# function to transform a node into a (feature) vector,
# should be a numpy array.
def transformer(x):
return np.array(x[0])
with self.assertRaises(TreeConvolutionError):
prepare_trees(trees, transformer, left_child, right_child)
def test_example(self):
# simple smoke test from the example file
tree1 = ((0, 1), ((1, 2), ((0, 1), ), ((-1, 0), )),
((-3, 0), ((2, 3), ), ((1, 2), )))
tree2 = ((16, 3), ((0, 1), ((5, 3), ), ((2, 6), )), ((2, 9), ))
trees = [tree1, tree2]
# function to extract the left child of a node
def left_child(x):
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[1]
# function to extract the right child of node
def right_child(x):
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[2]
# function to transform a node into a (feature) vector,
# should be a numpy array.
def transformer(x):
return np.array(x[0])
prepared_trees = prepare_trees(trees, transformer, left_child,
right_child)
net = nn.Sequential(tcnn.BinaryTreeConv(2, 16), tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.BinaryTreeConv(16, 8), tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.BinaryTreeConv(8, 4), tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.DynamicPooling())
# output: torch.Size([2, 4])
shape = tuple(net(prepared_trees).shape)
self.assertEqual(shape, (2, 4))
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[1]
def right_child(x):
assert isinstance(x, tuple)
if len(x) == 1:
# leaf.
return None
return x[2]
def transformer(x):
return np.array(x[0])
net = nn.Sequential(
tcnn.BinaryTreeConv(10, 16),
tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.BinaryTreeConv(16, 8),
tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.BinaryTreeConv(8, 4),
tcnn.TreeLayerNorm(),
tcnn.TreeActivation(nn.ReLU()),
tcnn.DynamicPooling()
)
prepared_trees = prepare_trees(trees, transformer, left_child, right_child)
print(net(prepared_trees))
# prepared_tredddes_train = prepare_trees(trees, transformer, left_child, right_child)
# training_set = WatDivDataset(prepared_tredddes_train, np.array([1., 2., 3.]))
# training_generator = torch.utils.data.DataLoader(training_set, **params)
#
# prepared_tredddes_val = prepare_trees(treesVal, transformer, left_child, right_child)
# validation_set = WatDivDataset(prepared_tredddes_val, np.array([1., 2., 3.]))
# validation_generator = torch.utils.data.DataLoader(validation_set, **params)
#
# for train_step, (trees, target) in enumerate(training_generator):
# print(trees, target)
#scale target in log scale and StandardScaller
scalery, y_train_log_std, y_val_log_std, y_test_log_std = prepare_log_std_target(
y_train, y_val, y_test)
prepared_trees_train = prepare_trees(x_train.values, transformer, left_child,
right_child)
prepared_trees_val = prepare_trees(x_val.values, transformer, left_child,
right_child)
prepared_trees_test = prepare_trees(x_test.values, transformer, left_child,
right_child)
# Parameters
# Generators
training_set = WatDivDataset(prepared_trees_train,
y_train_log_std.reshape(-1, 1))
training_generator = torch.utils.data.DataLoader(training_set, **params)
validation_set = WatDivDataset(prepared_trees_val,
y_val_log_std.reshape(-1, 1))
validation_generator = torch.utils.data.DataLoader(validation_set, **params)