Scale invariant CNN

"Size doesn't matter for me" - That's what our neural network said

Keras - Baseline CNN

Requirements

Install them first on your laptop

scikit-images
tqdm

Baseline CNN

Took Mnist Data, resized it to 16x16, 24x24 and 28x28. Then I padded all of them to make it look like 32x32

Training Loss and Accuracy:

loss: 0.0620 - acc: 98.23

Validation Accuracy:

acc:98.44444444444445

To Run it just do:

python3 BaselineCNN.py

Testing:

X_tester = np.pad(resize(X_test[1174],(6,6)),[13,13],mode='constant')

np.argmax(model.predict(X_tester.reshape(1,32,32,1)))

plt.subplot(121)
plt.imshow(X_tester, cmap='gray')