ConvNet

A Convolutional neural network for image classification (CIFAR-10).

Pre-processing

The dataset CIFAR10 in torchvision provides images with values between 0 and 1, but they are not normalized to have zero mean and unit variance, so a PyTorch Transforms is used.

Furthermore. the original dataset does not have a validation set so a SubsetRandomSampler is used.

Layers

The convolutional neural network uses rectified linear activation functions (ReLUs) as activation function. It has the following layers:

• 2 x Convolutional layer
• 1 x Max-pooling layer
• 2 x Convolutional layer
• 1 x Max-pooling layer
• 1 Fully connected layer: 512 units

Main

• show_9_images is a function that takes the first 9 images and plot them in order to better understand the dataset, by visualizing some samples.
• run is the function that run the model with the default values (num_epochs = 20, batch_size = 32, learning_rate = 10e-3, momentum = 0.9, kernel_size = 3)
• find_best_channels_momentum_kernel is a function that try different hyperparameters in order to find the ones that provide the higher accuracy.
• best_run runs the CN with the preset better hyperparameters provided by the find_best_channels_momentum_kernel function.

Results

Default configuration without dropout

Test accuracy is: 69.72 %.

The best epoch is the 3, with a validation accuracy equals to 74.8 %.

Default configuration with dropout

Test accuracy is: 71.15 %.

The best epoch is the 12, with a validation accuracy equals to 74.6 %.

Best configuration with dropout

Best channel 1 value: 82. Best channel 2 value: 64. Best momentum value: 0.7. Best kernel size: 3.

Test accuracy is: 81.39 %.

The best epoch is the 19, with a validation accuracy equals to 82.1 %.