An extensible and modular implementation of Convolution Neural Networks in C++.
- C++ 11
The dataset used for training is MNIST handwritten digits dataste. It contains 60,000 training samples and 10,000 testing examples. You can download it from MNIST link. Please extract all four files and place it in a separate folder named 'MNIST'.
To run, execute following commands
g++ load_image.cpp layers.cpp cnn.cpp -o test.out
./test.out
Accuracy around 92.47% when trained for 20 epochs with 10 batch size and 0.5 learning rate.
60000 images
60000 labels
Epoch : 1/20
Batch : 6000/6000 Training Loss: 0.323261
Time taken : 130.9434204102sec
Test Accuracy : 0.6905666590 Test Loss : 0.2095038593
Epoch : 2/20
Batch : 6000/6000 Training Loss: 0.1709278524
Time taken : 128.9757385254sec
Test Accuracy : 0.7898333073 Test Loss : 0.1530781686
Epoch : 3/20
Batch : 6000/6000 Training Loss: 0.1308346391
Time taken : 129.6773834229sec
Test Accuracy : 0.8605833054 Test Loss : 0.1087474748
Epoch : 4/20
Batch : 6000/6000 Training Loss: 0.1005350351
Time taken : 128.2032928467sec
Test Accuracy : 0.8832333088 Test Loss : 0.0921756253
Epoch : 5/20
Batch : 6000/6000 Training Loss: 0.0876044407
Time taken : 127.9117736816sec
Test Accuracy : 0.8980500102 Test Loss : 0.0819654241
Epoch : 6/20
Batch : 6000/6000 Training Loss: 0.0798420981
Time taken : 127.6298980713sec
Test Accuracy : 0.8978333473 Test Loss : 0.0818339363
Epoch : 7/20
Batch : 6000/6000 Training Loss: 0.0752588511
Time taken : 127.9868087769sec
Test Accuracy : 0.9021333456 Test Loss : 0.0782186687
Epoch : 8/20
Batch : 6000/6000 Training Loss: 0.0717870668
Time taken : 127.7466049194sec
Test Accuracy : 0.9084166884 Test Loss : 0.0727907568
Epoch : 9/20
Batch : 6000/6000 Training Loss: 0.0690624490
Time taken : 127.7570114136sec
Test Accuracy : 0.9107166529 Test Loss : 0.0722316727
Epoch : 10/20
Batch : 6000/6000 Training Loss: 0.0663161054
Time taken : 127.9231414795sec
Test Accuracy : 0.9190833569 Test Loss : 0.0643773302
Epoch : 11/20
Batch : 6000/6000 Training Loss: 0.0639779940
Time taken : 127.9275512695sec
Test Accuracy : 0.9213166833 Test Loss : 0.0632545203
Epoch : 12/20
Batch : 6000/6000 Training Loss: 0.0630218387
Time taken : 127.8605957031sec
Test Accuracy : 0.9217833281 Test Loss : 0.0625450537
Epoch : 13/20
Batch : 6000/6000 Training Loss: 0.0606022254
Time taken : 127.7766952515sec
Test Accuracy : 0.9232333302 Test Loss : 0.0612221025
Epoch : 14/20
Batch : 6000/6000 Training Loss: 0.0596275069
Time taken : 127.7988128662sec
Test Accuracy : 0.9236999750 Test Loss : 0.0606820211
Epoch : 15/20
Batch : 6000/6000 Training Loss: 0.0583804324
Time taken : 127.9415817261sec
Test Accuracy : 0.9226333499 Test Loss : 0.0616037995
Epoch : 16/20
Batch : 6000/6000 Training Loss: 0.0563948229
Time taken : 128.7188110352sec
Test Accuracy : 0.9245833158 Test Loss : 0.0605751537
Epoch : 17/20
Batch : 6000/6000 Training Loss: 0.0559208356
Time taken : 128.3111114502sec
Test Accuracy : 0.9254500270 Test Loss : 0.0596362650
Epoch : 18/20
Batch : 6000/6000 Training Loss: 0.0545428991
Time taken : 128.6391754150sec
Test Accuracy : 0.9259833097 Test Loss : 0.0595779866
Epoch : 19/20
Batch : 6000/6000 Training Loss: 0.0537633859
Time taken : 128.4654388428sec
Test Accuracy : 0.9280333519 Test Loss : 0.0578881949
Epoch : 20/20
Batch : 6000/6000 Training Loss: 0.0531267338
Time taken : 127.8627777100sec
Test Accuracy : 0.9269833565 Test Loss : 0.0591480248
10000 images
10000 labels
Test Accuracy : 0.9247999787 Test Loss : 0.0607771948
Convolution neural network for the multi-classification.
Network flow: Input-Convolution Layer-Pooling layer-FC1-FC2-Output
Contributions are always welcome! For implementing custom layers, you can just inherit Layer class within layers.cpp and implement all virtual functions. Few enhancements planned, implementation of softmax layer, corssentropy loss, adam optimizer, etc,.
Distributed under the GPL-3.0 License. See LICENSE.txt for more information.