A web app where user can draw Bengali digit and the AI model can detect handwritten digit and predict the digit.
| First Image | Second Image |
|---|---|
 |
 |
Web App: Bengali Digit Recognizer
Framework: Pytorch
Datasets: NumtaDB: Bengali Handwritten Digits
Custom CNN model is used
| CNN Layer | Linear Layer |
|---|---|
| Layers: 7 | Layers: 2 |
| Normalization: BatchNorm2d | |
| Pooling: MaxPool2d | |
| Activation Function: ReLU | Activation Function: ReLU, Softmax |
| Dropout: 0.25 | Dropout: 0.4 |
| Hyper Parameters |
|---|
| Epoch: 25 |
| Batch Size: 128 |
| Learning Rate: 1e-3 |
| Loss function: NLLLoss |
| Optimizer: Adam |
| Scheduler: StepLR |
| Transformations: Resize(180), RandomRotation(30), |
| Data Split: 20% |
Test Accuracy: 99.2890%
| Class wise Accuracy |
|---|
| Test Accuracy of 0: 99% (1104/1111) |
| Test Accuracy of 1: 99% (1095/1105) |
| Test Accuracy of 2: 99% (1094/1099) |
| Test Accuracy of 3: 98% (1080/1091) |
| Test Accuracy of 4: 99% (1078/1087) |
| Test Accuracy of 5: 99% (1088/1096) |
| Test Accuracy of 6: 99% (1057/1067) |
| Test Accuracy of 7: 99% (1050/1053) |
| Test Accuracy of 8: 99% (1124/1127) |
| Test Accuracy of 9: 99% (1051/1059) |
| Test Accuracy (Overall): 99% (10821/10895) |
- Pandas
- Numpy
- PIL
- Matplotlib
The web app is deployed in Heroku
Web App: Bengali Digit Recognizer
| First Image | Second Image |
|---|---|
|
|
I have no experience in web design, so I use this website template from repo How to deploy a keras model to production Author: Siraj Raval
Shudipto Trafder
Slack Name: @Shudipto Trafder
Email: Shudiptotrafder@gmail.com
Linkedin: Shudipto Trafder