7. Results

Please note that the following results are when compared with images of different classes used in training of the model; for a real- life accuracy statistics, the model must be compared with non-vector carrier images as well as vector-carrier images to compute the accuracy. Each epoch took approximately 5 hrs to run on 16 GB RAM with a 2.2 GHz Quad-Core Intel Core i7 processor. Please note that the loss function did not reduce exponentially as typically expected due to fine-tuning of the learning rate resulting in abnormal variables.

Epoch	Loss	Accuracy	Notes
0	52	0	1 epoch equates to > 5 hrs of learning
1	24	0.1113	learning rate 0.01
2	16	0.1113
3	16	0.1113
4	26.15	0.1113	Irregular data due to high learning rate (eventually regularized)
5	26.95	0.1113
6	27.05	0.1113
7	14.71	0.1113
8	7.8	0.1113
9	8.7	0.1113
10	5.95	0.1113
11	6.7953	0.1113
12	6.7195	0.1113
13	10.4273	0.1113
14	7.5758	0.1113
15	10.4273	0.1113	Reduced the learning rate from 0.01 to 0.0005 for stabilised learning
16	1.3632	0.5049	reduced learning rate to 0.0001 and began shuffling images (error in code resulted in only training one batch) - finally accuracy is beginning to improve
17	1.0958	0.6061	learning rate increased to 0.001
18	0.8411	0.697
19	0.7158	0.7672	Loss and accuracy appear to have stabilised
20	0.2745	0.9073
21	0.2111	0.9258	Again accuracy is plateauing- reduced learning rate to 0.00005
22	0.9659	0.8913
23	0.1059	0.9678	Finally reached industrial level precision (>95%)  after approximately 120 hrs of training