1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
|
"""An adaptation of RandLA-Net to the classification task, which was not
addressed in the `"RandLA-Net: Efficient Semantic Segmentation of Large-Scale
Point Clouds" <https://arxiv.org/abs/1911.11236>`_ paper.
"""
import os.path as osp
import torch
import torch.nn.functional as F
from torch import Tensor
from torch.nn import Linear
from tqdm import tqdm
import torch_geometric.transforms as T
from torch_geometric.datasets import ModelNet
from torch_geometric.loader import DataLoader
from torch_geometric.nn import MLP
from torch_geometric.nn.aggr import MaxAggregation
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.nn.pool import knn_graph
from torch_geometric.nn.pool.decimation import decimation_indices
from torch_geometric.typing import WITH_TORCH_CLUSTER
from torch_geometric.utils import softmax
if not WITH_TORCH_CLUSTER:
quit("This example requires 'torch-cluster'")
# Default activation and batch norm parameters used by RandLA-Net:
lrelu02_kwargs = {'negative_slope': 0.2}
bn099_kwargs = {'momentum': 0.01, 'eps': 1e-6}
class SharedMLP(MLP):
"""SharedMLP following RandLA-Net paper."""
def __init__(self, *args, **kwargs):
# BN + Act always active even at last layer.
kwargs['plain_last'] = False
# LeakyRelu with 0.2 slope by default.
kwargs['act'] = kwargs.get('act', 'LeakyReLU')
kwargs['act_kwargs'] = kwargs.get('act_kwargs', lrelu02_kwargs)
# BatchNorm with 1 - 0.99 = 0.01 momentum
# and 1e-6 eps by defaut (tensorflow momentum != pytorch momentum)
kwargs['norm_kwargs'] = kwargs.get('norm_kwargs', bn099_kwargs)
super().__init__(*args, **kwargs)
class LocalFeatureAggregation(MessagePassing):
"""Positional encoding of points in a neighborhood."""
def __init__(self, channels):
super().__init__(aggr='add')
self.mlp_encoder = SharedMLP([10, channels // 2])
self.mlp_attention = SharedMLP([channels, channels], bias=False,
act=None, norm=None)
self.mlp_post_attention = SharedMLP([channels, channels])
def forward(self, edge_index, x, pos):
out = self.propagate(edge_index, x=x, pos=pos) # N, d_out
out = self.mlp_post_attention(out) # N, d_out
return out
def message(self, x_j: Tensor, pos_i: Tensor, pos_j: Tensor,
index: Tensor) -> Tensor:
"""Local Spatial Encoding (locSE) and attentive pooling of features.
Args:
x_j (Tensor): neighboors features (K,d)
pos_i (Tensor): centroid position (repeated) (K,3)
pos_j (Tensor): neighboors positions (K,3)
index (Tensor): index of centroid positions
(e.g. [0,...,0,1,...,1,...,N,...,N])
Returns:
(Tensor): locSE weighted by feature attention scores.
"""
# Encode local neighboorhod structural information
pos_diff = pos_j - pos_i
distance = torch.sqrt((pos_diff * pos_diff).sum(1, keepdim=True))
relative_infos = torch.cat([pos_i, pos_j, pos_diff, distance],
dim=1) # N * K, d
local_spatial_encoding = self.mlp_encoder(relative_infos) # N * K, d
local_features = torch.cat([x_j, local_spatial_encoding],
dim=1) # N * K, 2d
# Attention will weight the different features of x
# along the neighborhood dimension.
att_features = self.mlp_attention(local_features) # N * K, d_out
att_scores = softmax(att_features, index=index) # N * K, d_out
return att_scores * local_features # N * K, d_out
class DilatedResidualBlock(torch.nn.Module):
def __init__(
self,
num_neighbors,
d_in: int,
d_out: int,
):
super().__init__()
self.num_neighbors = num_neighbors
self.d_in = d_in
self.d_out = d_out
# MLP on input
self.mlp1 = SharedMLP([d_in, d_out // 8])
# MLP on input, and the result is summed with the output of mlp2
self.shortcut = SharedMLP([d_in, d_out], act=None)
# MLP on output
self.mlp2 = SharedMLP([d_out // 2, d_out], act=None)
self.lfa1 = LocalFeatureAggregation(d_out // 4)
self.lfa2 = LocalFeatureAggregation(d_out // 2)
self.lrelu = torch.nn.LeakyReLU(**lrelu02_kwargs)
def forward(self, x, pos, batch):
edge_index = knn_graph(pos, self.num_neighbors, batch=batch, loop=True)
shortcut_of_x = self.shortcut(x) # N, d_out
x = self.mlp1(x) # N, d_out//8
x = self.lfa1(edge_index, x, pos) # N, d_out//2
x = self.lfa2(edge_index, x, pos) # N, d_out//2
x = self.mlp2(x) # N, d_out
x = self.lrelu(x + shortcut_of_x) # N, d_out
return x, pos, batch
def decimate(tensors, ptr: Tensor, decimation_factor: int):
"""Decimates each element of the given tuple of tensors."""
idx_decim, ptr_decim = decimation_indices(ptr, decimation_factor)
tensors_decim = tuple(tensor[idx_decim] for tensor in tensors)
return tensors_decim, ptr_decim
class Net(torch.nn.Module):
def __init__(
self,
num_features,
num_classes,
decimation: int = 4,
num_neighboors: int = 16,
return_logits: bool = False,
):
super().__init__()
self.decimation = decimation
# An option to return logits instead of log probabilities:
self.return_logits = return_logits
self.fc0 = Linear(in_features=num_features, out_features=8)
# 2 DilatedResidualBlock converges better than 4 on ModelNet.
self.block1 = DilatedResidualBlock(num_neighboors, 8, 32)
self.block2 = DilatedResidualBlock(num_neighboors, 32, 128)
self.mlp1 = SharedMLP([128, 128])
self.max_agg = MaxAggregation()
self.mlp_classif = SharedMLP([128, 32], dropout=[0.5])
self.fc_classif = Linear(32, num_classes)
def forward(self, x, pos, batch, ptr):
x = x if x is not None else pos
b1 = self.block1(self.fc0(x), pos, batch)
b1_decimated, ptr1 = decimate(b1, ptr, self.decimation)
b2 = self.block2(*b1_decimated)
b2_decimated, _ = decimate(b2, ptr1, self.decimation)
x = self.mlp1(b2_decimated[0])
x = self.max_agg(x, b2_decimated[2])
x = self.mlp_classif(x)
logits = self.fc_classif(x)
return logits if self.return_logits else logits.log_softmax(dim=-1)
def train(epoch):
model.train()
total_loss = 0
for data in tqdm(train_loader):
data = data.to(device)
optimizer.zero_grad()
out = model(data.x, data.pos, data.batch, data.ptr)
loss = F.nll_loss(out, data.y)
loss.backward()
optimizer.step()
total_loss += data.num_graphs * float(loss)
return total_loss / len(train_loader.dataset)
@torch.no_grad()
def test(loader):
model.eval()
correct = 0
for data in loader:
data = data.to(device)
out = model(data.x, data.pos, data.batch, data.ptr)
correct += int((out.argmax(dim=-1) == data.y).sum())
return correct / len(loader.dataset)
if __name__ == '__main__':
path = osp.dirname(osp.realpath(__file__))
path = osp.join(path, '..', 'data/ModelNet10')
pre_transform, transform = T.NormalizeScale(), T.SamplePoints(1024)
train_dataset = ModelNet(path, '10', True, transform, pre_transform)
test_dataset = ModelNet(path, '10', False, transform, pre_transform)
train_loader = DataLoader(train_dataset, 32, shuffle=True, num_workers=6)
test_loader = DataLoader(test_dataset, 32, shuffle=False, num_workers=6)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net(3, train_dataset.num_classes).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=20,
gamma=0.5)
for epoch in range(1, 201):
loss = train(epoch)
test_acc = test(test_loader)
print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}, Test: {test_acc:.4f}')
scheduler.step()
|
