Using the perceptron learning algorithm to differentiate between Thai currency, specifically Thai coins: (เหรียญ riianR).
This repository contains two files:
- perceptron_THB.py- Implentation of the perceptron classifcation algorithm and perceptron learning algorithm from scratch. This file does the testing and training.
- THB_data.py - Creates an artificial data set using the specifications of Thai coins.
specs = {
"1 satang": {
"diameter": 15, # mm
"mass": 0.5 # g
},
"5 satang": {
"diameter": 16, # mm
"mass": 0.6 # g
},
"10 satang": {
"diameter": 17.5, # mm
"mass": 0.8 # g
},
"25 satang": {
"diameter": 16, # mm
"mass": 1.9 # g
},
"50 satang": {
"diameter": 18, # mm
"mass": 2.4 # g
},
"1 baht": {
"diameter": 20, # mm
"mass": 3 # g
},
"2 baht": {
"diameter": 21.75, # mm
"mass": 4 # g
},
"5 baht": {
"diameter": 24, # mm
"mass": 6 # g
},
"10 baht": {
"diameter": 26, # mm
"mass": 8.5 # g
}
}
Artificial data is generated by the THB_data.py file. This file contains two importable methods:
- gen_dataset- Genrates an artificial dataset by randomly selecting two coins from the specs dictionary shown above, introduces random error to the diameter and mass of each coin, generates an N length sample.
- plot_data - Use matplotlib library to create simple data visualization.
The more error I introduce into the dataset, the more interesting the results become. The separating hyperplane is rather obvious without error.
Sample output after running "perceptron-THB.py":
| Using perceptron to classify | ['1 satang', '50 satang'] |
| Weights | [ 11. 3.56406557 -63.21304217] |
| Misclassifications: | 0 |
| Found a separating hyperplace in: | 968 iterations |
| Accuracy: | 100.00% |
- The algorithm is only capable of binary classifcation which is kind of boring in this case because there are much more than 2 THai coins.
- No consideration was made on finding the optimal hyperplane.
- It might be interesting to add a parameter to the perceptron learning algorithm "allowed_misclassifcations", this will allow more leniency.
- Add hyperparameter "learning rate" to PLA