-
Notifications
You must be signed in to change notification settings - Fork 0
/
Paper_Tree.py
202 lines (163 loc) · 8.32 KB
/
Paper_Tree.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
import torch
import torch.nn as nn
import torch.optim as optim
from Paper_global_vars import global_vars
from convmixer import ConvMixer
from resnet import ResNet8
import numpy as np
import timm
class DecisionNode(nn.Module):
def __init__(self, model, judge=[-1,-1]):
super(DecisionNode, self).__init__()
self.model = model
self.judge = judge
def forward(self, x):
outputs = self.model(x)
outputs = torch.sigmoid(outputs)
return outputs
class SequentialDecisionTreeForRDNet(nn.Module):
def __init__(self, isTest=False):
super(SequentialDecisionTreeForRDNet, self).__init__()
self.isTree = True
self.debug = True
self.isTest = isTest
self.nodes = nn.ModuleList([
DecisionNode(self.create_rdnet(2), judge=[[0,1,8,9],[2,3,4,5,6,7]]),
DecisionNode(self.create_rdnet(2), judge=[[0,8],[1,9]]),
DecisionNode(self.create_rdnet(2), judge=[[0],[8]]),
DecisionNode(self.create_rdnet(2), judge=[[1],[9]]),
DecisionNode(self.create_rdnet(3), judge=[[2,6],[3,5],[4,7]]),
DecisionNode(self.create_rdnet(2), judge=[[2],[6]]),
DecisionNode(self.create_rdnet(2), judge=[[3],[5]]),
DecisionNode(self.create_rdnet(2), judge=[[4],[7]])
])
def create_rdnet(self, num_classes):
model = timm.create_model('rdnet_tiny', pretrained=False, num_classes=num_classes)
if self.isTest==False:
local_pretrained_path = 'rdnet_tiny/pytorch_model.bin'
state_dict = torch.load(local_pretrained_path)
for key in ['head.fc.weight', 'head.fc.bias']:
if key in state_dict:
del state_dict[key]
model.load_state_dict(state_dict, strict=False)
model.head.fc = nn.Linear(model.head.fc.in_features, num_classes)
return model
def forward(self, x):
final_outputs = torch.ones(x.size(0), 10, device=x.device)
for node in self.nodes:
outputs = node(x)
if self.debug:
print(outputs.shape)
self.debug = False
for i, class_indices in enumerate(node.judge):
final_outputs[:, class_indices] *= outputs[:, i].unsqueeze(1)
return final_outputs
class SequentialDecisionTree(nn.Module):
def __init__(self):
super(SequentialDecisionTree, self).__init__()
self.isTree = True
self.nodes = nn.ModuleList([
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[0,1,8,9],[2,3,4,5,6,7]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[0,8],[1,9]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[0],[8]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[1],[9]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=3), judge=[[2,6],[3,5],[4,7]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[2],[6]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[3],[5]]),
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=2), judge=[[4],[7]])
])
def forward(self, x):
final_outputs = torch.ones(x.size(0), 10, device=x.device)
for node in self.nodes:
outputs = node(x)
for i, class_indices in enumerate(node.judge):
final_outputs[:, class_indices] *= outputs[:, i].unsqueeze(1)
return final_outputs
class SequentialDecisionTreeCIFAR100ForRDNet(nn.Module):
def __init__(self,isTest=False):
super(SequentialDecisionTreeCIFAR100ForRDNet, self).__init__()
self.isTree = True
self.isTest=isTest
self.debug = True
# 第一层节点：区分20个大类
self.root_node = DecisionNode(self.create_rdnet(20))
# 创建20个子节点，每个对应一个大类
self.sub_nodes = nn.ModuleList([
DecisionNode(self.create_rdnet(5))
for _ in range(20)
])
# Define the correct CIFAR-100 coarse to fine mapping
self.coarse_labels = [
[4, 30, 55, 72, 95], # aquatic mammals
[1, 32, 67, 73, 91], # fish
[54, 62, 70, 82, 92], # flowers
[9, 10, 16, 28, 61], # food containers
[0, 51, 53, 57, 83], # fruit and vegetables
[22, 39, 40, 86, 87], # household electrical devices
[5, 20, 25, 84, 94], # household furniture
[6, 7, 14, 18, 24], # insects
[3, 42, 43, 88, 97], # large carnivores
[12, 17, 37, 68, 76], # large man-made outdoor things
[23, 33, 49, 60, 71], # large natural outdoor scenes
[15, 19, 21, 31, 38], # large omnivores and herbivores
[34, 63, 64, 66, 75], # medium-sized mammals
[26, 45, 77, 79, 99], # non-insect invertebrates
[2, 11, 35, 46, 98], # people
[27, 29, 44, 78, 93], # reptiles
[36, 50, 65, 74, 80], # small mammals
[47, 52, 56, 59, 96], # trees
[8, 13, 48, 58, 90], # vehicles 1
[41, 69, 81, 85, 89] # vehicles 2
]
def create_rdnet(self, num_classes):
model = timm.create_model('rdnet_tiny', pretrained=False, num_classes=num_classes)
if self.isTest==False:
local_pretrained_path = 'rdnet_tiny/pytorch_model.bin'
state_dict = torch.load(local_pretrained_path)
for key in ['head.fc.weight', 'head.fc.bias']:
if key in state_dict:
del state_dict[key]
model.load_state_dict(state_dict, strict=False)
model.head.fc = nn.Linear(model.head.fc.in_features, num_classes)
return model
def forward(self, x):
# 获取大类的概率分布
coarse_probs = self.root_node(x)
# 对每个大类，获取其小类的概率分布
fine_probs = torch.stack([sub_node(x) for sub_node in self.sub_nodes], dim=1)
# 创建用于存储重新排列后的概率的张量
reordered_probs = torch.zeros(x.size(0), 100, device=x.device)
# 重新排列概率以匹配原始的100个类别
if self.debug:
print("Mapping results:")
for coarse_idx, fine_indices in enumerate(self.coarse_labels):
for fine_idx, class_idx in enumerate(fine_indices):
reordered_probs[:, class_idx] = fine_probs[:, coarse_idx, fine_idx]
if self.debug:
print(f"Class {class_idx}: fine_probs[{coarse_idx}, {fine_idx}] -> reordered_probs[{class_idx}]")
self.debug = False
# 展平并扩大5倍的coarse_probs
expanded_coarse_probs = coarse_probs.unsqueeze(2).expand(-1, -1, 5).reshape(x.size(0), 100)
# 和reordered_probs相乘
final_probs = expanded_coarse_probs * reordered_probs
return final_probs
class SequentialDecisionTreeCIFAR100(nn.Module):
def __init__(self):
super(SequentialDecisionTreeCIFAR100, self).__init__()
self.isTree = True
# 第一层节点：区分10个大类
self.root_node = DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=10))
# 10个节点：每个对应一个大类，区分其下的10个小类
self.sub_nodes = nn.ModuleList([
DecisionNode(ConvMixer(dim=256, depth=8, kernel_size=5, patch_size=1, n_classes=10))
for _ in range(10)
])
def forward(self, x):
# 获取大类的概率分布
coarse_probs = self.root_node(x)
# 对每个大类，获取其小类的概率分布
fine_probs = torch.stack([sub_node(x) for sub_node in self.sub_nodes], dim=1)
# 计算最终的100类概率分布
final_probs = coarse_probs.unsqueeze(2) * fine_probs
final_probs = final_probs.view(x.size(0), -1) # 展平为 (batch_size, 100)
return final_probs