def _build_feature_spec(self):
num_attrs = len(self._attr_types)
numeric_types = ("float", "int")
embedding_types = ("int", "string")
self._fspec = FeatureSpec(num_attrs, self._weighted, self._labeled)
if not self._attr_dims:
self._attr_dims = [None for _ in range(num_attrs)]
if num_attrs != len(self._attr_dims):
raise ValueError(
"The size of attr_dims must be equal with attr_types.")
def check(dim, attr_type, bucket):
if not dim:
assert type_name in numeric_types and bucket_size is None, \
"Must assign an attr_dim for {}, and bucket_size should None." \
.format(type_name)
else:
assert type_name in embedding_types, \
"Must assign an attr_dim with None for {}".format(type_name)
for attr_type, dim in zip(self._attr_types, self._attr_dims):
type_name, bucket_size, is_multival = self.parse(attr_type)
check(dim, type_name, bucket_size)
if is_multival:
self._fspec.append_multival(bucket_size, dim, ",")
elif dim:
self._fspec.append_sparse(bucket_size, dim, type_name == "int")
else:
self._fspec.append_dense(type_name == "float")
def get_sub_graph_data(self):
spec = FeatureSpec(10)
for i in range(3):
spec.append_dense()
total_dim = 3
for i in range(7):
dim = random.randint(8, 10)
spec.append_sparse(20 + 10 * i, dim, False)
total_dim += dim
N = 10
# [f_num, batch_size] = [3, N]
batch_floats = np.random.random([3, N])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, N]
batch_ints = np.array([[i * j for j in range(N)] for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
vertices = Vertex(floats=batch_floats, ints=batch_ints)
adj = np.zeros([N, N], dtype=np.float32)
for i in range(N):
for j in range(N):
adj[i][j] = i
tf_adj = tf.convert_to_tensor(adj, dtype=tf.float32)
g = SubGraph(vertices, tf_adj, schema=("nodes", spec))
return g, N, total_dim
def test_homogeneous_graph(self):
spec = FeatureSpec(10)
for i in range(3):
spec.append_dense()
total_dim = 3
for i in range(7):
dim = random.randint(8, 10)
spec.append_sparse(20 + 10 * i, dim, False)
total_dim += dim
hops = [4, 5]
# the centric vertices share the same spec with 2-hop neighbors
schema = [("nodes", spec), ("nodes", spec), ("nodes", spec)]
# [f_num, batch_size] = [3, 2]
batch_floats = np.array([[1.0 * i, 2.0 * i] for i in range(3)])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2]
batch_ints = np.array([[i, 2 * i] for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
vertices = Vertex(floats=batch_floats, ints=batch_ints)
# [f_num, batch_size] = [3, 2 * 4]
hop1_floats = np.array([[1.0 * i, 2.0 * i] * hops[0]
for i in range(3)])
hop1_floats = tf.convert_to_tensor(hop1_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4]
hop1_ints = np.array([[i, 2 * i] * hops[0] for i in range(7)])
hop1_ints = tf.convert_to_tensor(hop1_ints, dtype=tf.int64)
neighbor_hop_1 = Vertex(floats=hop1_floats, ints=hop1_ints)
# [f_num, batch_size] = [3, 2 * 4 * 5]
hop2_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] * hops[1]
for i in range(3)])
hop2_floats = tf.convert_to_tensor(hop2_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4 * 5]
hop2_ints = np.array([[i, 2 * i] * hops[0] * hops[1]
for i in range(7)])
hop2_ints = tf.convert_to_tensor(hop2_ints, dtype=tf.int64)
neighbor_hop_2 = Vertex(floats=hop2_floats, ints=hop2_ints)
g = EgoGraph(vertices, [neighbor_hop_1, neighbor_hop_2], schema, hops)
dims = np.array([total_dim, 16, 8])
model = HomoEgoGIN(dims,
num_head=5,
bn_fn=None,
active_fn=tf.nn.relu,
droput=0.1)
embeddings = model.forward(g)
with tf.Session() as sess:
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
ret = sess.run(embeddings)
self.assertListEqual([2, 8],
list(ret.shape)) # [batch_size, output_dim]
def get_model_and_graph(self):
spec = FeatureSpec(10)
for i in range(3):
spec.append_dense()
total_dim = 3
for i in range(7):
dim = random.randint(8, 10)
spec.append_sparse(20 + 10 * i, dim, False)
total_dim += dim
hops = [4, 5]
# the centric vertices share the same spec with 2-hop neighbors
schema = [("nodes", spec), ("nodes", spec), ("nodes", spec)]
# [f_num, batch_size] = [3, 2]
batch_floats = np.array([[1.0 * i, 2.0 * i] for i in range(3)])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2]
batch_ints = np.array([[i, 2 * i] for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
# [batch_size] = [2]
batch_labels = np.array([1, 0])
batch_labels = tf.convert_to_tensor(batch_labels, dtype=tf.int32)
vertices = Vertex(floats=batch_floats,
ints=batch_ints,
labels=batch_labels)
# [f_num, batch_size] = [3, 2 * 4]
hop1_floats = np.array([[1.0 * i, 2.0 * i] * hops[0]
for i in range(3)])
hop1_floats = tf.convert_to_tensor(hop1_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4]
hop1_ints = np.array([[i, 2 * i] * hops[0] for i in range(7)])
hop1_ints = tf.convert_to_tensor(hop1_ints, dtype=tf.int64)
neighbor_hop_1 = Vertex(floats=hop1_floats, ints=hop1_ints)
# [f_num, batch_size] = [3, 2 * 4 * 5]
hop2_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] * hops[1]
for i in range(3)])
hop2_floats = tf.convert_to_tensor(hop2_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4 * 5]
hop2_ints = np.array([[i, 2 * i] * hops[0] * hops[1]
for i in range(7)])
hop2_ints = tf.convert_to_tensor(hop2_ints, dtype=tf.int64)
neighbor_hop_2 = Vertex(floats=hop2_floats, ints=hop2_ints)
g = EgoGraph(vertices, [neighbor_hop_1, neighbor_hop_2], schema, hops)
dims = np.array([total_dim, 16, 8])
model = HomoEgoGraphSAGE(dims,
bn_fn=None,
active_fn=tf.nn.relu,
droput=0.1)
return model, g, 2 # batch_size
def get_graph(cls, hops, neg=None):
spec = FeatureSpec(10)
for i in range(3):
spec.append_dense()
total_dim = 3
for i in range(7):
dim = i + 1
spec.append_sparse(20 + 10 * i, dim, False)
total_dim += dim
neg = 1 if not neg else int(neg)
hops[0] = int(hops[0] * neg)
# the centric vertices share the same spec with 2-hop neighbors
schema = [("nodes", spec), ("nodes", spec), ("nodes", spec)]
# batch_size = 2
# [f_num, batch_size] = [3, 2 * neg]
batch_floats = np.array([[1.0 * i, 2.0 * i] * neg for i in range(3)])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * neg]
batch_ints = np.array([[i, 2 * i] * neg for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
# [batch_size] = [2]
batch_labels = np.array([1, 0])
batch_labels = tf.convert_to_tensor(batch_labels, dtype=tf.int32)
vertices = Vertex(floats=batch_floats,
ints=batch_ints,
labels=batch_labels)
# [f_num, batch_size] = [3, 2 * neg * hop0]
hop1_floats = np.array([[1.0 * i, 2.0 * i] * hops[0]
for i in range(3)])
hop1_floats = tf.convert_to_tensor(hop1_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * neg * hop0]
hop1_ints = np.array([[i, 2 * i] * hops[0] for i in range(7)])
hop1_ints = tf.convert_to_tensor(hop1_ints, dtype=tf.int64)
neighbor_hop_1 = Vertex(floats=hop1_floats, ints=hop1_ints)
# [f_num, batch_size] = [3, 2 * neg * hop0 * hop1]
hop2_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] * hops[1]
for i in range(3)])
hop2_floats = tf.convert_to_tensor(hop2_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * neg * hop0 * hop1]
hop2_ints = np.array([[i, 2 * i] * hops[0] * hops[1]
for i in range(7)])
hop2_ints = tf.convert_to_tensor(hop2_ints, dtype=tf.int64)
neighbor_hop_2 = Vertex(floats=hop2_floats, ints=hop2_ints)
hops[0] = int(hops[0] / neg)
g = EgoGraph(vertices, [neighbor_hop_1, neighbor_hop_2], schema, hops)
return g, total_dim
def test_homogeneous_graph(self):
spec = FeatureSpec(10)
for i in range(3):
spec.append_dense()
total_dim = 3
for i in range(7):
dim = random.randint(8, 10)
spec.append_sparse(20 + 10 * i, dim, False)
total_dim += dim
hops = [4, 5]
# the centric vertices share the same spec with 2-hop neighbors
schema = [("nodes", spec), ("nodes", spec), ("nodes", spec)]
# [f_num, batch_size] = [3, 2]
batch_floats = np.array([[1.0 * i, 2.0 * i] for i in range(3)])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2]
batch_ints = np.array([[i, 2 * i] for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
vertices = Vertex(floats=batch_floats, ints=batch_ints)
# [f_num, batch_size] = [3, 2 * 4]
hop1_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] for i in range(3)])
hop1_floats = tf.convert_to_tensor(hop1_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4]
hop1_ints = np.array([[i, 2 * i] * hops[0] for i in range(7)])
hop1_ints = tf.convert_to_tensor(hop1_ints, dtype=tf.int64)
neighbor_hop_1 = Vertex(floats=hop1_floats, ints=hop1_ints)
# [f_num, batch_size] = [3, 2 * 4 * 5]
hop2_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] * hops[1] for i in range(3)])
hop2_floats = tf.convert_to_tensor(hop2_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2 * 4 * 5]
hop2_ints = np.array([[i, 2 * i] * hops[0] * hops[1] for i in range(7)])
hop2_ints = tf.convert_to_tensor(hop2_ints, dtype=tf.int64)
neighbor_hop_2 = Vertex(floats=hop2_floats, ints=hop2_ints)
g1 = EgoGraph(vertices, [neighbor_hop_1, neighbor_hop_2], schema, hops)
g2 = g1.forward()
self.assertListEqual(list(g2.expands), hops)
x_list = [g2.nodes, g2.hop(0), g2.hop(1)]
with tf.Session() as sess:
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
ret = sess.run(x_list)
self.assertListEqual(list(ret[0].shape), [2, total_dim])
self.assertListEqual(list(ret[1].shape), [2 * hops[0], total_dim])
self.assertListEqual(list(ret[2].shape), [2 * hops[0] * hops[1], total_dim])
class Decoder(object):
""" Decoder is used for Graph.node() and Graph.edge() to describe the schema
of data source.
"""
def __init__(self,
weighted=False,
labeled=False,
attr_types=None,
attr_delimiter=":",
attr_dims=None):
""" Initialize a data source decoder.
Args:
weighted (boolean, Optional): Whether the data source has weights.
Default is False.
labeled (boolean, Optional): Whether the data source has labels.
Default is False.
attr_types (list, Optional): Attribute type list if attributes exist.
Default is None, which means no attribute exists. Each element must
be string, tuple or list.
Valid types like below:
attr_types = ['int', 'float', 'string']
attr_types = ['int', ('string', 10), 'string'] # 10 means bucket size
For ('string', 10), we will get the string and hash it into 10 buckets
directly. The raw attribute can by any string.
# True means multi-val splited by ',', only for string attribute
attr_types = ['int', ('string', 10, True)]
# 10 means bucket size
attr_types = ['int', ('int', 10), 'string']
For ('int', 10), we will cast the string to int first and then hash it
into 10 buckets. In this way, the raw attribute must be an integer.
When `attr_dims` is assigned, be sure that the string attribute must
be configured with a bucket size. Bucket size for int attribute is
optional, and it will be considered as a continuous attribute if bucket
size is not assigned.
attr_delimiter (string, Optional): The delimiter to seperate attributes.
Default is ':'. If attributes exist, all of them are concatenated
together with a delimiter in the raw storage. We need to know how to
parse them.
attr_dims (list, Optional): An integer list, the element of which
represents the dimension of the corresponding attribute that will be
encodeding to. Default is None, which means no attribute exists.
All valid configurations of attr_type and attr_dim are as shown below.
| attr_type |attr_dim| encoded into |
| --------- | -- | -------- |
| "string" | 8 | Dynamic bucket embedding variable, dim=8 |
| ("string",10) | 8 | Embedding variable, bucketsize=10, dim=8 |
|("string",10,True)| 8 |Sparse embedding variable, bucketsize=10,dim=8|
|("string",None,True)| 8| Sparse dynamic embedding variable, dim=8 |
| "int" |None| Continues numeric tensor |
| "int" | 8 | Dynamic bucket embedding variable, dim=8 |
| ("int",10) | 8 | Embedding variable, bucket size=10, dim=8 |
| "float" |None| Continues numeric tensor |
Note that dynamic bucket embedding variable is only supported in PAI-TF.
For continues numeric attribute, attr_dim should be either None or 0.
"""
self._weighted = weighted
self._labeled = labeled
self._attr_types = attr_types
self._attr_delimiter = attr_delimiter
self._attr_dims = attr_dims
self._int_attr_num = 0
self._float_attr_num = 0
self._string_attr_num = 0
self._fspec = None
self._attributed = self._parse_attributes()
def _parse_attributes(self):
if not self._attr_types:
return False
if not isinstance(self._attr_types, list):
raise ValueError(
"attr_types for Decoder must be a list, got {}.".format(
type(self._attr_types)))
for i in range(len(self._attr_types)):
type_name, bucket_size, is_multival = self.parse(
self._attr_types[i])
self._int_attr_num += int(type_name == "int")
self._float_attr_num += int(type_name == "float")
if is_multival:
self._string_attr_num += 1
else:
self._int_attr_num += int(type_name == "string"
and bucket_size is not None)
self._string_attr_num += int(type_name == "string"
and bucket_size is None)
return True
def _build_feature_spec(self):
num_attrs = len(self._attr_types)
numeric_types = ("float", "int")
embedding_types = ("int", "string")
self._fspec = FeatureSpec(num_attrs, self._weighted, self._labeled)
if not self._attr_dims:
self._attr_dims = [None for _ in range(num_attrs)]
if num_attrs != len(self._attr_dims):
raise ValueError(
"The size of attr_dims must be equal with attr_types.")
def check(dim, attr_type, bucket):
if not dim:
assert type_name in numeric_types and bucket_size is None, \
"Must assign an attr_dim for {}, and bucket_size should None." \
.format(type_name)
else:
assert type_name in embedding_types, \
"Must assign an attr_dim with None for {}".format(type_name)
for attr_type, dim in zip(self._attr_types, self._attr_dims):
type_name, bucket_size, is_multival = self.parse(attr_type)
check(dim, type_name, bucket_size)
if is_multival:
self._fspec.append_multival(bucket_size, dim, ",")
elif dim:
self._fspec.append_sparse(bucket_size, dim, type_name == "int")
else:
self._fspec.append_dense(type_name == "float")
def parse(self, attr_type):
if isinstance(attr_type, tuple) or isinstance(attr_type, list):
type_name = attr_type[0]
bucket_size = attr_type[1] if len(attr_type) >= 2 else None
is_multival = attr_type[2] if len(attr_type) >= 3 else False
else:
type_name = attr_type
bucket_size = None
is_multival = False
assert type_name in {"int", "float", "string"}
if is_multival and type_name != "string":
raise ValueError("multi-value attribute must be string type.")
return type_name, bucket_size, is_multival
@property
def weighted(self):
return self._weighted
@property
def labeled(self):
return self._labeled
@property
def attributed(self):
return self._attributed
@property
def attr_types(self):
return self._attr_types
@property
def attr_delimiter(self):
return self._attr_delimiter
@property
def data_format(self):
# attributed << 3 | labeled << 2 | weighted << 1
return int(self._weighted * 2 + \
self._labeled * 4 + self._attributed * 8)
@property
def int_attr_num(self):
return self._int_attr_num
@property
def float_attr_num(self):
return self._float_attr_num
@property
def string_attr_num(self):
return self._string_attr_num
@property
def feature_spec(self):
if not self._fspec:
self._build_feature_spec()
return self._fspec
def format_attrs(self, int_attrs, float_attrs, string_attrs):
""" Reshape and format attributes with int_attr_num, float_attr_num
and string_attr_num calculated by decoder.attr_types.
Return:
Reshaped int_attrs, float_attrs, string_attrs
"""
if int_attrs is not None:
int_attrs = int_attrs.reshape(-1, self._int_attr_num)
if float_attrs is not None:
float_attrs = float_attrs.reshape(-1, self._float_attr_num)
if string_attrs is not None:
string_attrs = string_attrs.reshape(-1, self._string_attr_num)
return int_attrs, float_attrs, string_attrs
def test_heterogeneous_graph(self):
u_spec = FeatureSpec(10)
for i in range(3):
u_spec.append_dense()
u_total_dim = 3
for i in range(7):
dim = random.randint(8, 10)
u_spec.append_sparse(20 + 10 * i, dim, False)
u_total_dim += dim
i_spec = FeatureSpec(19)
for i in range(6):
i_spec.append_dense()
i_total_dim = 6
for i in range(13):
dim = random.randint(8, 11)
i_spec.append_sparse(30 + 10 * i, dim, False)
i_total_dim += dim
u_out_dim = 16
i_out_dim = 12
out_dim = 9
hops = [4, 5]
# the centric vertices share the same spec with 2-hop neighbors
# metapath: u--i--i
schema = [("u_nodes", u_spec), ("nbr", i_spec), ("nbr", i_spec)]
# [f_num, batch_size] = [3, 2]
batch_floats = np.array([[1.0 * i, 2.0 * i] for i in range(3)])
batch_floats = tf.convert_to_tensor(batch_floats, dtype=tf.float32)
# [i_num, batch_size] = [7, 2]
batch_ints = np.array([[i, 2 * i] for i in range(7)])
batch_ints = tf.convert_to_tensor(batch_ints, dtype=tf.int64)
vertices = Vertex(floats=batch_floats, ints=batch_ints)
# [f_num, batch_size] = [6, 2 * 4]
hop1_floats = np.array([[1.0 * i, 2.0 * i] * hops[0]
for i in range(6)])
hop1_floats = tf.convert_to_tensor(hop1_floats, dtype=tf.float32)
# [i_num, batch_size] = [13, 2 * 4]
hop1_ints = np.array([[i, 2 * i] * hops[0] for i in range(13)])
hop1_ints = tf.convert_to_tensor(hop1_ints, dtype=tf.int64)
neighbor_hop_1 = Vertex(floats=hop1_floats, ints=hop1_ints)
# [f_num, batch_size] = [6, 2 * 4 * 5]
hop2_floats = np.array([[1.0 * i, 2.0 * i] * hops[0] * hops[1]
for i in range(6)])
hop2_floats = tf.convert_to_tensor(hop2_floats, dtype=tf.float32)
# [i_num, batch_size] = [13, 2 * 4 * 5]
hop2_ints = np.array([[i, 2 * i] * hops[0] * hops[1]
for i in range(13)])
hop2_ints = tf.convert_to_tensor(hop2_ints, dtype=tf.int64)
neighbor_hop_2 = Vertex(floats=hop2_floats, ints=hop2_ints)
g = EgoGraph(vertices, [neighbor_hop_1, neighbor_hop_2], schema, hops)
layer_ui = EgoSAGELayer("heter_uv",
input_dim=(u_total_dim, i_total_dim),
output_dim=u_out_dim,
agg_type="mean",
com_type="concat")
layer_ii = EgoSAGELayer("heter_ii",
input_dim=i_total_dim,
output_dim=i_out_dim,
agg_type="mean",
com_type="concat")
layer_uii = EgoSAGELayer("heter_uii",
input_dim=(u_out_dim, i_out_dim),
output_dim=out_dim,
agg_type="sum",
com_type="concat")
layer_group_1 = EgoSAGELayerGroup([layer_ui, layer_ii])
layer_group_2 = EgoSAGELayerGroup([layer_uii])
model = EgoGraphSAGE([layer_group_1, layer_group_2],
bn_fn=None,
active_fn=tf.nn.relu,
droput=0.1)
embeddings = model.forward(g)
with tf.Session() as sess:
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
ret = sess.run(embeddings)
self.assertListEqual([2, 9],
list(ret.shape)) # [batch_size, output_dim]