PV021 Neural networks project

Simple feedforward neural network trained by SGD, backpropagation and mean square error.

• Repository

Authors

• Róbert Kolcún @NoName115 487564@mail.muni.cz
• Ondrej Kurák @H00N24 487589@mail.muni.cz

Usage

$ ./RUN

Outputs

• trainPredictions - predicted labels for train vectors (train + eval)
• actualTestPredictions - predicted labels fot test vectors

Makefile

src/Makefile

Debug

• Enables debug prints

$ make debug
$ ./network
==Start==

==Training data loading: Start==
File opened: ../MNIST_DATA/mnist_train_vectors.csv
File opened: ../MNIST_DATA/mnist_train_labels.csv
==Training data loading: End (1.00s)==

==Network initialization: Start==
==Network initialization: End (0.02s)==

==Training: Start==
2 epochs, 0.200 training rate, 32 batch size
Training data shape: (48000, 785)
Training labels shape: (48000, 10)
Eval data shape: (12000, 785)
Eval labels shape: (12000, 10)
Epoch 1: 100%
  Time: 76.25s
  Eval accuracy: 94.17%
Epoch 2: 100%
  Time: 76.17s
  Eval accuracy: 95.88%
==Training: End==

==Predicting training data: Start ==
Test data accuracy 96.48%
==Predicting training data: End (37.96s)==

==Predicting testing data: Start ==
File opened: ../MNIST_DATA/mnist_test_vectors.csv
File opened: ../MNIST_DATA/mnist_test_labels.csv
Test data accuracy 96.20%
==Predicting testing data: End (6.51s)==

==Total time 212.80s==

Test

• Compiles and runs unit tests
• CuTest website
• Licence can be found in src/tests/license.txt file

$ make test
./test
............

OK (12 tests)

Profile

• profiling using gprof

$ make profile
$ ./network
$ gprof ./network