MNIST experiments

MNIST32 experiments

NAME=mnist_baseline
python task_launcher.py \
--dataset=mnist \
--arch=architectures/arch_acai_kyle.py \ `#This architecture is from the ACAI repo`\
--save_every=100                         `#Save model every 100 epochs`\
--save_images_every=100 \
--epochs=1000 \
--resume=auto \
--n_channels=1                           `#MNIST is black and white`\
--ngf=0 \                                `#arch_acai_kyle.py ignores ngf, so this can be anything`\
--ndf=0 \                                `#arch_acai_kyle.py ignores ndf, so this can be anything`\
--name=${NAME} \
--batch_size=64 \
--beta=0.0 \                             `#no consistency loss`\
--lamb=10.0 \                            `#reconstruction weight`\
--cls=0.0 \                              `#we're not doing supervised mixes`\
--disable_mix \                          `#disable mixup (this reduces the class into an adversarial AE)`\
--mixer=mixup \                          `#mixing function (ignored due to disable_mix above)`\
--seed=1 \
--classify_encoding='32,10' \            `#bottleneck is 32 units, train a classifier over it which predicts 10 classes`\
--weight_decay=1e-5 \                    `#L2 norm on the weights`\
--beta1=0.5 \                            `#beta1 for ADAM optimiser`\
--beta2=0.99 \                           `#beta2 for ADAM optimiser`\
--lr=1e-4                                `#learning rate for ADAM`