An MNIST handwriting images classifier implemented using numpy and Convolutional Neural Network(CNN).
Overall architecture of the network is shown above.
- Input shape:
(B, 1, 28, 28) - Output shape:
(B, 28, 26, 26) - Class:
nn_layers.nn_convolutional_layer - Details: convolution with 28 filters of shape
(3, 3, 1)
- Input shape:
(B, 28, 26, 26) - Output shape:
(B, 28, 26, 26) - Class:
nn_layers.nn_activation_layer - Details: ReLU function is used
- Input shape:
(B, 28, 26, 26) - Output shape:
(B, 28, 13, 13) - Class:
nn_layers.nn_activation_layer - Details: 2x2 maxpooling with stride 2
- Input shape:
(B, 28, 13, 13) - Output shape:
(B, 128) - Class:
nn_layers.nn_fc_layer - Details: Flattens input, i.e. considers input as
Bvectors of size 28 × 28 × 13=4732.
- Input shape:
(B, 128) - Output shape:
(B, 128) - Class:
nn_layers.nn_activation_layer - Details: ReLU function is used
- Input shape:
(B, 128) - Output shape:
(B, 10) - Class:
nn_layers.nn_fc_layer
- Input shape:
(B, 10) - Output shape:
(B, 10) - Class:
nn_layers.nn_softmax_layer
- Class:
nn_layers.nn_cross_entropy_layer - Details: Cross-entropy loss function is used
Among 60k MNIST training images, 50k images were used for training and 10k images were used for validation. The model was trained for 2 epochs, with 48 iterations for each epoch, total 96 iterations, with batch size of 1024.
The training was tried four times, and average training time was 302.5(5min 2.5sec) ± 1.5 seconds. Loss graph over iterations for each trial is shown below.
10k images from MNIST test images were used for test. Average test accuracy was 97.27 ± 0.43%. Some prediction results are shown below.
- The project was submitted for an assignment in COSE474 Deep Learning course, and is no longer maintained.
- Redistribution of source code is permitted, but the following consequences is your responsibility.