Solve the Saw

Solved the puzzle using a CNN achieving a 94.98% accuracy on testing data!

Comment

Inspiration

Machine learning is a very hot topic in today's world, and the possibilities of computational methods is very enticing.

What it does

Our project is a CNN to predict the order of image puzzle pieces.

How we built it

We used tensorflow and keras to build a CNN and trained the model on training data.

Challenges we ran into

Although the model performed well on training data, the accuracy on testing data dropped significantly.

Accomplishments that we're proud of

We started from little knowledge of machine learning and finished training a complete model.

What we learned

We learned a lot about tensorflow, layers, and machine learning in general.

Loss Function Optimizer Number training pics Testing Learning Rate epochs Accuracy
KLD Adam 30000 9000 0.001 3 20.17%
Poisson adam 30000 9000 0.01 3 23.82%
Poisson sgd 30000 9000 0.001 3 24.85%
KLD sgd 30000 9000 0.001 3 26.39%
Poisson sgd 30000 9000 0.001 3 27.01%
SCCE sgd 30000 9000 0.001 3 47.10%
SCCE Adam 30000 9000 0.001 1 47.67%
SCCE Adam 30000 9000 0.01 3 47.78%
SCCE Adam 30000 9000 0.001 5 48.97%
SCCE Adam 30000 9000 0.001 3 49.21%
SCCE Adam 90% 5% 0.01 3 73.03%
SCCE Adam 80% 10% 0.001 10 80.00%
SCCE Adam 70.00% 15% 0.001 3 88.71%
SCCE Adam 90% 5% 0.001 3 91.15%
SCCE Adam 90% 5% 0.01 10 91.85%
SCCE Adam 90% 5% 0.001 10 94.25%
SCCE Adam 90% 5% 0.001 30 94.98%

What's next for Puzzle Solver

The accuracy could be much improved.

Built With

Share this project:

Updates