-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsummary.txt
143 lines (110 loc) · 4.02 KB
/
summary.txt
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
- All trained on 4 T4 GPUs
PP-UMamba (Ours)
----------------------------------
Train time: 48m 27.6s
#Params: 10076356 ~10M
Size(MB): 109.15
IOU: 0.8004
Dice: 0.8867
UMambaBot
----------------------------------
Train time: 73m 22.9s
#Params: 9813700 ~9.8M
Size(MB): 108.10
IOU: 0.7908
Dice: 0.8799
UNet
----------------------------------
Train time: 32m 13.1s
#Params: 7765985 ~7.8M
Size(MB): 31.09
IOU: 0.8071
Dice: 0.8929
Segformer
----------------------------------
Train time: 77m 49.3s
#Params: 17782913 ~17.8M
Size(MB): 71.13
IOU: 0.7260
Dice: 0.8454
ResNet18
----------------------------------
Train time: 28m 47.2s
#Params: 13930497 ~13.9M
Size(MB): 55.76
IOU: 0.7623
Dice: 0.8648
ResNet50
----------------------------------
Train time: 47m 3.9s
#Params: 42909633 ~42.9M
Size(MB): 171.85
IOU: 0.7661
Dice: 0.8672
## Original model, c_list=[8,16,24,32,48,64]
Ultra LightM-UNet
----------------------------------
Train time: 304m 41.9s
#Params: 49457 ~0.049M
Size(MB): 0.20
IOU: 0.6725
Dice: 0.7955
## One depth more shallow, changed c_list=[16, 32, 64, 128, 256]
Ultra LightM-UNet Modified v1
---------------------------------- # Highest params, highest scores .. higher dims in c_list give more accurate results but more params
Train time: 310m 14.4s
#Params: 217165 ~0.20M
Size(MB): 0.87
IOU: 0.7359
Dice: 0.8409
## doubled mamba parallels in PVM layer, 4 -> 8, c_list=[8,16,24,32,48,64]
Ultra LightM-UNet Modified v2
---------------------------------- # Smallest size w/ higher score over original
Train time:
#Params: 42737 ~0.042M
Size(MB): 0.17
IOU: 0.7100
Dice: 0.8243
## Combination of v1 and v2, one shallower and 8 parallels in mamba layer, c_list=[16, 32, 64, 128, 256]
Ultra LightM-UNet Modified v3
---------------------------------- # More shallow with more parallels gives lower score
Train time:
#Params: 176453 ~0.176M
Size(MB): 0.71
IOU: 0.7068
Dice: 0.8109
Ultra LightM-UNet Modified v4 *Trained on 1 NVIDIA GeForce RTX 3060
---------------------------------- # Pyramidal Pooling on 6 layer UNet, c_list=[16,32,64,128,256,512],
Train time: 103m 10.5s
#Params: 1135309 ~1.1M
Size(MB): 4.54
IOU: 0.7531
Dice: 0.8548
Ultra LightM-UNet Modified v5
---------------------------------- # 6 layer UNet, c_list=[16,32,64,128,256,512], to compare to v4 & v6
Train time: 489m 43.2s
#Params: 703885 ~0.7M
Size(MB): 2.82
IOU: 0.7638
Dice: 0.8633
Ultra LightM-UNet Modified v6
---------------------------------- # 6 layer UNet, c_list=[16,32,64,128,256,512], PP at bottleneck
Train time: 536 43.5s
#Params: 769677 ~0.77M
Size(MB): 3.08
IOU: 0.7630
Dice: 0.8628
## Description of changes for ULMUNet_v1
Initially, `c_list=[8, 16, 24, 32, 48, 64]` which determines the size of the output at each step of the encoder and reflects
in the decoder by iterating back up the c_list.
In ULMUNet_v1, we modified `c_list=[16, 32, 64, 128, 256]` as well as making the U-Net one depth more shallow.
The intention was it improve performance by increasing the amount of parameters in each step of the encoder and decoder.
(Increased parameter count but reduced UNet depth to keep params# low)
## Description of changes for ULMUNet_v2
Increased the number of parallels in the PVM (Parallel Vision Mamba) block to 8, previously 4.
This has a direct effect on the parameter count, decreasing it. Further proving the idea in this paper: https://arxiv.org/pdf/2403.20035.pdf
## Description of changes for ULMUNet_v3
Modified with a combination of increasing to 8 parallels in the PVM block.
Modified with increasing the size of the output in each step of the encoder-decoder with `c_list=[16, 32, 64, 128, 256]`
as well as reducing the depth of the UNet to 5 layers.
(Combination of v1 and v2)