CIFAR-100 All-CNN-C

class deepobs.pytorch.testproblems.cifar100_allcnnc.cifar100_allcnnc(batch_size, l2_reg=0.0005)

DeepOBS test problem class for the All Convolutional Neural Network C on Cifar-100.

Details about the architecture can be found in the original paper.

The paper does not comment on initialization; here we use Xavier for conv filters and constant 0.1 for biases.

L2-Regularization is used on the weights (but not the biases) which defaults to 5e-4.

The reference training parameters from the paper are batch size = 256, num_epochs = 350 using the Momentum optimizer with \(\mu = 0.9\) and an initial learning rate of \(\alpha = 0.05\) and decrease by a factor of 10 after 200, 250 and 300 epochs.

Parameters:

• batch_size (int) -- Batch size to use.
• l2_reg (float) -- L2-regularization factor. L2-Regularization (weight decay) is used on the weights but not the biases. Defaults to 5e-4.

get_regularization_groups()

Creates regularization groups for the parameters.

Returns:A dictionary where the key is the regularization factor and the value is a list of parameters. Return type:dict

set_up()

Set up the All CNN C test problem on Cifar-100.