Gender Detection From Bangla Handwritten Images
Handwriting is a unique quality of human being, every person has unique handwriting that is very hard to copy but there can be some similarities Between male and female handwriting. we will classify the difference between Male and Female Handwritting Digit using logistic regression.
Differences between Male and Female Handwriting Digit can be used in various applications like Archeology, forensic department. In this assignment, our main goal will be to achieve a good result using logistic regression and tuning hyperparameters correctly to get a better result.
Dataset
For this experiment, we will use the dataset Ekush which is available in Github. There will be a total of 30830 images, and it was split in a 90:10 ratio. 90% (27747) of data is used in training and 10% (3083) was used in testing.Whole dataset was shuffled before spliting , so that traning and testing dataset can have both label data.
Snapshot from Dataset
Experimental Setup
All of these experiments were performed using Google Colab free GPU, Models were created in PyTorch.
During the whole experiment,
- The height and width of the input was 28*28 =784
- Output dimension was (0,1)=2
- Each batch size was 256
- The number of iteration was 4000
- Softmax activation function was used
- GPU Tesla T4 was available
- Number of Epochs 332
- totaldata: 30830
- minibatch: 256
- iterations: 4,000
-
epochs
$epochs = iterations \div \frac{totaldata}{minibatch} = 4000 \div \frac{30830}{256} = 332.14 == 332 $
We will use different optimizer(SGD and Adam) and learning rate to achieve better performance
Result
Using Logistic Regression
| Experiment Number | Optimizer | Learning Rate | Accurecy of last 500 iterations |
|---|---|---|---|
| 1 | SGD | 0.0001 | 51.41 |
| 2 | SGD | 0.0005 | 50.02 |
| 3 | SGD | 0.0010 | 49.92 |
| 4 | Adam | 0.0001 | 51.05 |
| 5 | Adam | 0.02 | 50.86 |
| 6 | Adam | 0.05 | 48.3 |
Using Deep Neural Network
| Experiment Number | Optimizer | Learning Rate | Num of Hidden Layer | Btach Size | Num of epoch | Accurecy of last 1000 iterations |
|---|---|---|---|---|---|---|
| 1 | SGD | 0.02 | 4 | 256 | 830 | 50.14 |
| 2 | SGD | 0.01 | 4 | 512 | 1660 | 49.17 |
| 3 | SGD | 0.10 | 3 | 256 | 830 | 50.66 |
| 4 | SGD | 0.01 | 2 | 1024 | 3321 | 49.82 |