DeepLearning

Comparative report of deep learning architectures on the CIFAR dataset.

By Panagiotis Giagkoulas and Lukas Edman.

Requirements

You need to have Anaconda or Miniconda installed:

Miniconda

Anaconda

You also need a GPU with CUDA support.

Environment

To install the environment run:

conda env create -f environment.yml

After installing, you may need to activate the environment, using either

source activate keras_env

or

activate keras_env

Running the code

To run an experiment, run:

python main.py

There are a number of flags available to modify the experiment:

--model_name               Name of the model. Loads model with same name automatically. Flag must be used for the model to be saved.
--architecture             Architecture to use. [list of architectures below] 
--pretrained               If provided, will load a model pretrained on ImageNet and only the last layer is trained. Valid for VGG/ResNet/DenseNet.
--dataset                  Dataset to use. [cifar10/cifar100]
--save_interval            Save every x epochs.
--batch_size               Batch size for training and testing.
--n_epochs                 Number of epochs to train for.
--optimizer                Optimizer to use. [adam/rmsprop/sgd]
--lr                       Learning rate.
--export                   Export the training statistics to a file.

Supported architectures:

all_conv                   Regular All-CNN-C from Springenberg et al.
all_bn_conv                All-CNN-C with batch normalization
model_c                    Model C from Springenberg et al.
simple_conv                Simple CNN from Keras
simple_bn_conv             Simple CNN with batch norm
lenet5                     LeNet5
lenet5_do                  LeNet5 with dropout
lenet5_bn                  LeNet5 with batch norm
vgg16                      VGG16
resnet50                   ResNet50
densenet                   DenseNet

An example run of Simple CNN with batch normalization on CIFAR-10 using Adam looks like:

python main.py --architecture simple_bn_conv --dataset cifar10 --save_interval 1 --batch_size 64 --n_epochs 100 --optimizer adam --lr 1e-3