from deepsnap.dataset import GraphDataset
from torch_geometric.datasets import *
from graphgym.register import register_loader
def load_dataset_example(format, name, dataset_dir):
dataset_dir = '{}/{}'.format(dataset_dir, name)
if format == 'PyG':
if name == 'QM7b':
dataset_raw = QM7b(dataset_dir)
graphs = GraphDataset.pyg_to_graphs(dataset_raw)
return graphs
register_loader('example', load_dataset_example)
TODO: use dataset.Generator instead?
"""
if cfg.dataset.interpretation != 'bipartite' or cfg.dataset.format != 'SBM':
return None
from util.graph_generators import get_kHSBM_graph
cd = cfg.dataset
memberships, nw = get_kHSBM_graph(cd.sbm_n, cd.sbm_m, cd.sbm_k, cd.sbm_p,
cd.sbm_q, cd.sbm_c)
nxG = nw.to_networkx()
dsG = deepsnap.graph.Graph(nxG)
return [dsG]
register_loader('SBM_single_bipartite', SBM_single_bipartite)
def SBM_single_from_generated(format, name, dataset_dir):
if cfg.dataset.interpretation != 'bipartite' or cfg.dataset.format != 'SBM_gen':
return None
# todo generate if not found; coalesce this and `generate_nx_to_file`
import loader
dir = files('data').joinpath(dataset_dir)
dataset = loader.load_nx(name, dir)
return dataset
nxG: nx.graph.Graph
nxG = load_from_numpy(adj_path)
# add dummy "node feature"
# TODO GG breaks if no node attributes are set.
for nodeIx in nxG.nodes:
onehot = np.zeros(1)
onehot = torch.from_numpy(onehot).to(torch.float)
nxG.node[nodeIx]['node_feature'] = onehot
labels = np.load(labels_path)
# set partition as graph attribute
partition = []
for i in range(labels.max() + 1):
partition.append(set(np.argwhere(labels == i).flatten()))
nxG.graph['partition'] = partition
# TODO GG breaks if no labels are set -- even if labels never really used
# set membership as node attribute
for nodeIx in nxG.nodes:
nxG.node[nodeIx]['node_label'] = torch.tensor(labels[nodeIx]).to(
torch.float)
dsG = deepsnap.graph.Graph(nxG)
return [dsG]
register_loader('modnet_generated_seed', load_modnet_graph)
def sbml_single_hyper(format, name, dataset_dir):
# GG simply uses the first registered loader that does not return None so we
# need to make this check in each loader.
if cfg.dataset.interpretation != 'hyper' or cfg.dataset.format != 'SBML':
return None
# TODO use dataset_dir?
# TODO: how often isthis method called?
# TODO: profile how long this conversion takes for large networks
# right now, this goes sbml -> igraph graph -> pytorch tensor -> dsG/networkx graph
# for alternative method, see loader/SBM.py
nw = Network.from_sbml(name)
pyg_data = nw.to_torch_data_onehot_hyper() # NOTE incidence edge indices
return _make_dataset(nw, pyg_data, nw.get_num_hypernodes())
register_loader('sbml_single_hyper', sbml_single_hyper)
def sbml_single_bipartite(format, name, dataset_dir):
if cfg.dataset.interpretation != 'bipartite' or cfg.dataset.format != 'SBML':
return None
nw = Network.from_sbml(name)
if cfg.dataset.max_node_degree is not None:
nw = nw.limit_node_degrees(cfg.dataset.max_node_degree)
if cfg.dataset.max_edge_degree is not None:
nw = nw.limit_edge_degrees(cfg.dataset.max_edge_degree)
if cfg.dataset.limit_to_largest_component is True:
nw = nw.limit_to_largest_component()
dataset_dir = '{}/{}'.format(dataset_dir, name)
netlists = find_netlists(dataset_dir)
if cfg.dataset.mean:
mean = np.load(cfg.dataset.mean)
stddev = np.load(cfg.dataset.stddev)
dataset = datasets.omitted(netlists,
min_edge_count=5,
resample=cfg.dataset.resample,
mean=mean,
std=stddev)
else:
dataset = datasets.omitted(netlists,
min_edge_count=5,
resample=cfg.dataset.resample)
graphs = h.to_deepsnap(dataset)
dataset = GraphDataset(
graphs,
task=cfg.dataset.task,
edge_train_mode=cfg.dataset.edge_train_mode,
edge_message_ratio=cfg.dataset.edge_message_ratio,
edge_negative_sampling_ratio=cfg.dataset.edge_negative_sampling_ratio,
resample_disjoint=cfg.dataset.resample_disjoint,
minimum_node_per_graph=0)
dataset._num_graph_labels = len(datasets.helpers.component_types)
return dataset
register_loader('omitted_netlists', load_dataset)
from graphgym.register import register_loader
def find_netlists(rootdir='.'):
child_list = [rootdir]
if os.path.isdir(rootdir):
child_list = (os.path.join(rootdir, c) for c in os.listdir(rootdir))
netlist_paths = []
for file_or_dir in child_list:
if os.path.isdir(file_or_dir):
contained_paths = find_netlists(file_or_dir)
netlist_paths.extend(contained_paths)
elif file_or_dir.endswith('.cir') or file_or_dir.endswith('.net'):
netlist_paths.append(file_or_dir)
return netlist_paths
def load_dataset(format, name, dataset_dir):
if format != 'NetlistProtoLinks':
return None
dataset_dir = '{}/{}'.format(dataset_dir, name)
netlists = find_netlists(dataset_dir)
dataset = PrototypeLinkDataset(netlists, normalize=False)
graphs = h.to_deepsnap(dataset)
return graphs
register_loader('netlist_proto_links', load_dataset)