-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathexperimental results
127 lines (104 loc) · 3.04 KB
/
experimental results
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
## Deep Food
#### 8 classes (dropout: 0.5)
**10 iterations**: 50% test accuracy <br/>
**20 iterations**: 55.38% <br/>
**30 iterations**: loss: 1.3320 - val_loss: 1.3258
Test = 0.5502, Train = 0.6747 <br/>
**50 iterations**: loss: 1.0033 - val_loss: 1.1839
Test = 0.618, Train = 0.7912 <br/>
**100 iterations**: loss: 1.3703 - val_loss: 1.3921
Test = 0.4594
Train = 0.5132 <br/>
**250 iterations**:
loss: 1.1524 - val_loss: 1.3816
Test = 0.5370
Train = 0.7350
#### Add one more layer 128 filters (now lesser parameters) dropout 0.5
Test Accuracy = 0.5343 Train Accuracy = 0.5992 <br/>
#### Adding one dropout after the 5th conv layer
bad performance:
loss: 1.3748 val_loss: 1.6631
Test Accuracy = 0.2583 Train Accuracy = 0.2928
#### Remove the dropout between fc layers
Overfitting: loss: 0.1633 val_loss: 2.6325
Test Accuracy = 0.5696 (need investigation why it is ok)
Train Accuracy = 0.9119
#### No dropout between conv layers, but 0.5 between fc (still 5 conv layers)
loss: 1.1716 val_loss: 1.4838
Accuracy = 0.5934
Train Accuracy = 0.7508
##### Early stopping (20 iterations) - 5 conv layers
Test Accuracy = 0.5467
Train Accuracy = 0.7785
#### 4 conv layers, no dropout b/w conv layers
**overfitting **<br/>
loss: 0.7156 val_loss: 2.1416
Test Accuracy = 0.5035
Train Accuracy = 0.9434
#### 4 conv layers, one dropout (0.5) between conv layers
loss: 1.2606 - val_loss: 2.3975
Test Accuracy = 0.4523
Train Accuracy = 0.6652
#### 4 conv layers, one dropout (0.25) between conv layers
loss: 0.9270 - val_loss: 1.5388
Test Accuracy = 0.4444
Train Accuracy = 0.7466
#### 5th convlayer without pooling (before dropout layer)
loss: 1.0435 - val_loss: 1.3928
Test Accuracy = 0.5687
Train Accuracy = 0.7642
### FC layers
#### 40
loss: 1.0245 - val_loss: 1.1747
Test Accuracy = 0.6031
Train Accuracy = 0.7809
#### 30
loss: 0.9632 - val_loss: 1.2167
Test = 0.5784
Train = 0.7899
#### 50
loss: 0.8682 - val_loss: 1.4710
Test = 0.5837
Train = 0.8791
#### 100
loss: 0.6193 - val_loss: 1.8436
Test = 0.6005
Train = 0.9024
#### 10
loss: 1.9945 - val_loss: 1.9959
Test = 0.1358
Train = 0.1778
## Horizontally Flipped Data
#### 4 conv layers (standard architecture + fc 40)
loss: 1.0658 - val_loss: 1.1294
Test = 0.5661
Train = 0.6833
#### fc 100
loss: 0.7515 - val_loss: 1.2590
Test = 0.6406
Train = 0.8400
#### fc 256
loss: 0.0851 - val_loss: 2.1954
Test Accuracy = 0.6194
Train Accuracy = 0.9998
#### 5th conv layer without pooling
loss: 1.9844 - val_loss: 1.8158
Test = 0.2579
Train = 0.2639
#### 6 Conv layers (as 4 layesrs with dropout, pooling)
loss: 0.3683 - val_loss: 1.4240
Test= 0.6208
Train = 0.9760
## VGG Bottleneck features
#### fc 256
loss: 0.1754 - acc: 0.9643 - val_loss: 2.2353 - val_acc: 0.7504
#### fc 128
loss: 0.3384 - acc: 0.9115 - val_loss: 1.7735 - val_acc: 0.7407
#### fc 512
loss: 0.1003 - acc: 0.9783 - val_loss: 2.1777 - val_acc: 0.7663
## To do
* More training data (flip images horizontally and vertically)
* Pre trained nets
* What's happening while overfitting. Do a class wise study
* Ensemble of VGG and custom network
* Visualizing the filters