In this project, we build ANNs using tensorflow library as an extension to ann-numpy repo where I build ANN architecture from scratch in numpy.
MNIST:
Our tensorflow ANN with 1 hidden layer with only 30 nodes on 10 classes handwritten digit MNIST dataset achieves test error of 3.4% (accuracy 96.6% !!).
training_error_rate
final training error: 0.033
test error: 0.03428571428571429
Confusion Matrix:
[[67 0 0 0 0 0 2 0 0 0]
[ 0 77 1 0 0 0 0 0 0 0]
[ 0 0 62 0 2 0 0 0 0 0]
[ 0 0 1 86 0 1 0 0 0 0]
[ 0 0 0 0 62 0 1 0 0 1]
[ 0 0 0 3 0 60 0 0 0 0]
[ 0 0 0 0 0 1 66 0 0 0]
[ 0 0 0 0 0 0 0 67 0 0]
[ 2 0 0 0 1 3 1 0 63 0]
[ 0 0 0 0 0 1 0 1 2 66]]
Classification Report:
precision recall f1-score support
0.0 0.97 0.97 0.97 69
1.0 1.00 0.99 0.99 78
2.0 0.97 0.97 0.97 64
3.0 0.97 0.98 0.97 88
4.0 0.95 0.97 0.96 64
5.0 0.91 0.95 0.93 63
6.0 0.94 0.99 0.96 67
7.0 0.99 1.00 0.99 67
8.0 0.97 0.90 0.93 70
9.0 0.99 0.94 0.96 70
accuracy 0.97 700
macro avg 0.97 0.97 0.96 700
weighted avg 0.97 0.97 0.97 700
CIFAR 10
For 10 class 3 channel images CIFAR dataset we try deeper archecture with 3 hidden layers of sizes 128, 64 and 32 nodes. The test error is 61.3% which is a bit worse than using my numpy version with test error of 53.1%. I've used CNNs to get down to 18% test error in my CNN folders.
training_error_rate
final training error: 0.647
test error: 0.613333333333
Confusion Matrix:
[[47 3 2 3 0 4 1 3 0 2]
[10 38 2 3 0 1 2 1 0 10]
[ 9 4 16 5 7 5 8 3 0 0]
[ 3 4 6 14 0 22 4 2 0 1]
[12 1 8 2 19 5 7 5 0 0]
[ 0 3 1 6 5 17 7 2 0 1]
[ 1 1 6 6 15 1 29 4 0 2]
[ 7 3 4 3 13 3 1 32 0 2]
[31 5 2 3 1 10 0 2 0 15]
[ 7 12 2 2 2 1 2 4 0 20]]
Classification Report:
precision recall f1-score support
0.0 0.37 0.72 0.49 65
1.0 0.51 0.57 0.54 67
2.0 0.33 0.28 0.30 57
3.0 0.30 0.25 0.27 56
4.0 0.31 0.32 0.31 59
5.0 0.25 0.40 0.31 42
6.0 0.48 0.45 0.46 65
7.0 0.55 0.47 0.51 68
8.0 0.00 0.00 0.00 69
9.0 0.38 0.38 0.38 52
accuracy 0.39 600
macro avg 0.35 0.38 0.36 600
weighted avg 0.35 0.39 0.36 600
