def eccpc_collate(batch):
""" Collates a list of dataset samples into a single batch (adapted in ecc.graph_info_collate_classification())
"""
targets, graphs, clouds_meta, clouds_flag, clouds, clouds_global = list(
zip(*batch))
targets = torch.cat(
[torch.from_numpy(t) for t in targets if t is not None], 0).long()
graphs = [graph for graph in graphs if graph is not None]
GIs = [ecc.GraphConvInfo(graphs, cloud_edge_feats)]
if len(clouds_meta[0]) > 0:
clouds_meta = [
item for sublist in clouds_meta if sublist is not None
for item in sublist
]
clouds_global = torch.cat([
torch.from_numpy(np.asarray(f))
for f in clouds_global if f is not None
], 0)
clouds_flag = torch.cat([
torch.from_numpy(np.asarray(f))
for f in clouds_flag if f is not None
], 0)
clouds = torch.cat(
[torch.from_numpy(np.asarray(f)) for f in clouds if f is not None],
0)
return targets, GIs, (clouds_meta, clouds_flag, clouds, clouds_global)
def graph_info_collate_classification(batch, edge_func):
""" Collates a list of dataset samples into a single batch. We assume that all samples have the same number of resolutions.
Each sample is a tuple of following elements:
features: 2D Tensor of node features
classes: LongTensor of class ids
graphs: list of graphs, each for one resolution
pooldata: list of triplets, each for one resolution: (pooling map, finer graph, coarser graph)
"""
features, classes, graphs, pooldata = list(zip(*batch))
graphs_by_layer = list(zip(*graphs))
pooldata_by_layer = list(zip(*pooldata))
features = torch.cat([torch.from_numpy(f) for f in features])
if features.dim() == 1: features = features.view(-1, 1)
classes = torch.LongTensor(classes)
GIs, PIs = [], []
for graphs in graphs_by_layer:
GIs.append(ecc.GraphConvInfo(graphs, edge_func))
for pooldata in pooldata_by_layer:
PIs.append(ecc.GraphPoolInfo(*zip(*pooldata)))
return features, classes, GIs, PIs
def eccpc_collate(batch):
""" Collates a list of dataset samples into a single batch (adapted in ecc.graph_info_collate_classification())
"""
targets, graphs, clouds_meta, clouds_flag, clouds, clouds_global, fnames, extension_subs, extension_fulls = list(
zip(*batch))
targets = torch.cat([torch.from_numpy(t) for t in targets], 0).long()
GIs = [ecc.GraphConvInfo(graphs, cloud_edge_feats)]
edges_for_ext = GIs[0].edges_for_ext
if len(clouds_meta[0]) > 0:
clouds = torch.cat([torch.from_numpy(f) for f in clouds], 0)
clouds_global = torch.cat([torch.from_numpy(f) for f in clouds_global],
0)
clouds_flag = torch.cat([torch.from_numpy(f) for f in clouds_flag], 0)
clouds_meta = [item for sublist in clouds_meta for item in sublist]
clouds_orig = clouds
mask = clouds_flag == 0
clouds = clouds[mask, :, :]
clouds_global = clouds_global[mask]
num_sp_list = [f.shape[0] for f in extension_subs]
ext_mask = torch.cat([
torch.unsqueeze(torch.from_numpy(t[:, 1]), -1) for t in extension_subs
], 0).long().squeeze(-1)
return targets, GIs, (clouds_meta, clouds_flag, clouds,
clouds_global), clouds_orig, edges_for_ext, fnames, (
ext_mask, extension_subs,
extension_fulls), num_sp_list
def eccpc_collate(batch):
""" Collates a list of dataset samples into a single batch (adapted in ecc.graph_info_collate_classification())
"""
targets, graphs, clouds_meta, clouds_flag, clouds, clouds_global = list(zip(*batch))
targets = torch.cat([torch.from_numpy(t) for t in targets], 0).long()
GIs = [ecc.GraphConvInfo(graphs, cloud_edge_feats)]
if len(clouds_meta[0])>0:
clouds = torch.cat([torch.from_numpy(f) for f in clouds], 0)
clouds_global = torch.cat([torch.from_numpy(f) for f in clouds_global], 0)
clouds_flag = torch.cat([torch.from_numpy(f) for f in clouds_flag], 0)
clouds_meta = [item for sublist in clouds_meta for item in sublist]
return targets, GIs, (clouds_meta, clouds_flag, clouds, clouds_global)
def eccpc_collate_test(batch):
""" Collates a list of dataset samples into a single batch (adapted in ecc.graph_info_collate_classification())
"""
targets, graphs, clouds_meta, clouds_flag, clouds, clouds_global, fnames = list(
zip(*batch))
targets = torch.cat([torch.from_numpy(t) for t in targets], 0).long()
GIs = [ecc.GraphConvInfo(graphs, cloud_edge_feats)]
edges_for_ext = GIs[0].edges_for_ext
if len(clouds_meta[0]) > 0:
clouds = torch.cat([torch.from_numpy(f) for f in clouds], 0)
clouds_global = torch.cat([torch.from_numpy(f) for f in clouds_global],
0)
clouds_flag = torch.cat([torch.from_numpy(f) for f in clouds_flag], 0)
clouds_meta = [item for sublist in clouds_meta for item in sublist]
clouds_orig = clouds
mask = clouds_flag == 0
clouds = clouds[mask, :, :]
clouds_global = clouds_global[mask]
return targets, GIs, (clouds_meta, clouds_flag, clouds,
clouds_global), clouds_orig, edges_for_ext, fnames