-
Notifications
You must be signed in to change notification settings - Fork 2
/
segment.py
209 lines (150 loc) · 7.75 KB
/
segment.py
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import os
import logging
import build_tree
import random
from encoder import MLP
class MLPCounter(nn.Module):
def __init__(self, *args, **kwargs):
super().__init__()
self.mlp = MLP(*args, **kwargs)
def forward(self, ans):
ans, counter = ans.split(ans.size(-1) - 1, dim=-1)
ans = self.mlp(ans)
ans = torch.cat([ans, counter], dim=-1)
return ans
class Segment(nn.Module):
def __init__(self, ancestor_dim, encoder, num_parts=0, num_classes=0, force_sample=False, part_cls_dropout=None):
super().__init__()
self.encoder = encoder
self.layers = []
self.tree = self.encoder.tree
self.num_parts = num_parts
self.num_classes = num_classes
self.has_sample = force_sample
for encoder_layer in self.encoder.layers:
if encoder_layer.layer_type == 'sampled':
self.has_sample = True
MLPClass = MLPCounter if self.has_sample else MLP
globe_dim = []
for encoder_layer in reversed(self.encoder.layers):
globe_dim.append(max(ancestor_dim, encoder_layer.feature_dim))
for i, encoder_layer in enumerate(reversed(self.encoder.layers)):
if i == 0:
assert globe_dim[i] == encoder_layer.feature_dim
input_dim = globe_dim[i]
push_down = nn.Identity()
else:
input_dim = globe_dim[i - 1] + encoder_layer.feature_dim
push_down = MLPClass([input_dim, globe_dim[i]])
push_down_sample = nn.Identity()
if encoder_layer.layer_type == 'sampled':
# push_down_sample = MLPClass([globe_dim[i], globe_dim[i + encoder.sample_layers - 1]])
push_down_sample = MLPClass([input_dim, globe_dim[i + encoder.sample_layers - 1]])
logging.info(f"layer {i} feature_dim = {encoder_layer.feature_dim} globe_dim = {globe_dim[i]}")
self.layers.append(nn.ModuleList([push_down, push_down_sample]))
self.layers = nn.ModuleList(self.layers)
# self.cloud_classifier = MLP([self.encoder.dim, self.encoder.dim // 2, self.encoder.dim // 4, num_classes], last_bn=False)
if num_parts > 0:
self.part_classfier = MLP([ancestor_dim, ancestor_dim // 2, ancestor_dim // 4, num_parts],
last_bn=False, last_dropout=part_cls_dropout)
else:
self.part_classfier = torch.nn.Identity()
def forward(self, *args, **kwargs):
if self.has_sample:
return self.forward_with_sample(*args, **kwargs)
assert not self.has_sample
features = self.encoder(*args, **kwargs)
self.features = features
batch_size = features.size(0)
inputs = None
for i, ((push_down, push_down_sample), (encoder_layer, features)) in enumerate(
zip(self.layers, reversed(tuple(zip(self.encoder.layers, self.encoder.layer_output))))):
if inputs is None:
inputs = features
else:
inputs = torch.cat([features, inputs], dim=-1)
ans = push_down(inputs)
if encoder_layer.layer_type == 'leaf':
break
n = ans.size(1) * 2
dim = ans.size(2)
assert encoder_layer.child_lr.size(0) == n
outputs = torch.empty(batch_size, n, dim, device='cuda')
outputs.scatter_(1, encoder_layer.child_l[None, :, None].expand(batch_size, -1, dim), ans)
outputs.scatter_(1, encoder_layer.child_r[None, :, None].expand(batch_size, -1, dim), ans)
inputs = outputs
def data_part(ans):
return ans[:, :, :-1]
def counter_part(ans):
return ans[:, :, -1, None]
node_features = torch.cat([ans, torch.ones(*ans.shape[:-1], 1, device='cuda')], dim=-1)
dim = node_features.size(-1)
arrange = self.encoder.arrange
ans = torch.zeros(batch_size, arrange.max().item() + 1, dim, dtype=torch.float, device='cuda')
ans.scatter_add_(1, arrange[:, :, None].cuda().expand(batch_size, -1, dim), node_features)
ans = data_part(ans) / counter_part(ans)
self.align_reg_loss = self.encoder.align_reg_loss
return ans
def forward_with_sample(self, *args, **kwargs):
features = self.encoder(*args, **kwargs)
self.features = features
batch_size = features.size(0)
# if calc_cloud_logits:
# cloud_logits = self.cloud_classifier(features)
scatter_list = [[] for _ in self.layers]
scatter_list[0].append([torch.tensor([0], device='cuda'), torch.ones(batch_size, 1, 1, device='cuda')])
def fetch(scatter_list):
n = sum([val.size(1) for ind, val in scatter_list])
dim = scatter_list[0][1].size(-1)
ans = torch.empty(batch_size, n, dim, device='cuda')
# print(f"fetch n = {n} dim = {dim}")
for ind, val in scatter_list:
# print(f"scatter item {val.shape}")
assert val.size(-1) == dim
ans.scatter_(1, ind[None, :, None].expand(batch_size, -1, dim), val)
return ans
for i, ((push_down, push_down_sample), (encoder_layer, features)) in enumerate(
zip(self.layers, reversed(tuple(zip(self.encoder.layers, self.encoder.layer_output))))):
inputs = fetch(scatter_list[i])
# print(f"decoder fwd {i} {features.shape} {inputs.shape}")
inputs = torch.cat([features, inputs], dim=-1)
ans = push_down(inputs)
# print(f"decoder push_down {ans.shape}")
if encoder_layer.layer_type != 'leaf':
scatter_list[i + 1].append([encoder_layer.child_l, ans])
scatter_list[i + 1].append([encoder_layer.child_r, ans])
if encoder_layer.layer_type == 'sampled':
# ans_sample = push_down_sample(ans)
ans_sample = push_down_sample(inputs)
# print(f"decoder push_down_sample {ans_sample.shape}")
ans_sample[:, :, -1] /= 2 ** self.encoder.sample_layers
scatter_list[i + self.encoder.sample_layers].append([encoder_layer.child_s, ans_sample])
def data_part(ans):
return ans[:, :, :-1]
def counter_part(ans):
return ans[:, :, -1, None]
counter = counter_part(ans)
node_features = torch.cat([data_part(ans) * counter, counter], dim=-1)
dim = node_features.size(-1)
arrange = self.encoder.arrange
ans = torch.zeros(batch_size, arrange.max().item() + 1, dim, dtype=torch.float, device='cuda')
ans.scatter_add_(1, arrange[:, :, None].cuda().expand(batch_size, -1, dim), node_features)
# print(f"scatter {ans.shape}")
# print(f"counters = {ans[:, :, -1]}")
ans = data_part(ans) / counter_part(ans)
# ans = self.part_classfier(ans)
# if calc_cloud_logits:
# ans = (ans, cloud_logits)
return ans
@staticmethod
def convert_raw(raw_logits, cloud_logits, class_parts, cloud_logits_coef=1):
# use static to make sure the model cannot see "class_parts"
import torch.nn.functional as F
cloud_logits = F.log_softmax(cloud_logits * cloud_logits_coef, dim=-1)
logits = torch.empty_like(raw_logits).cuda()
for c, cpart in enumerate(class_parts):
logits[:, :, cpart] = cloud_logits[:, c, None, None] + F.log_softmax(raw_logits[:, :, cpart], dim=-1)
return logits