This is the base repository for the Master's thesis in Computer Science at the University of St. Gallen. To start training run the script main.py with all necessary arguments. These include:

• path
  • Directory to where data should be stored
• model
  • Choose which model should be used for training, they should be defined in src/models/model_archs
• model_size
  • Depending on the model will allow to determine its size
• data
  • Choose the dataset for optimization (MNIST, Fashion MNIST, CIFAR10, CIFAR100, SVHN)
• dtype
  • Precision used (float32, float64)
• optimizer
  • Decide on the optimizer (NLCG, Adam, RMSProp, SGD)
• batch_size
  • Choose the size of your batches
• gpu
  • In case of multiple available GPUs, decide how many and which ones you want to use
• run_eagerly
  • Use eager execution in TensorFlow
• max_epochs
  • Limit the number of epochs
• use_wandb
  • Track your experiments with WandB
• max_calls
  • Limit the number of optimization steps

Here is an exemplary run for the script:

python main.py --path "/home/user/code/data/tf_data" --model FlatMLP --model_size small --data MNIST --dtype float64 --optimizer ADAM --batch_size 6000 --gpu 4 --run_eagerly --max_epochs 50000 --use_wandb --max_calls 6697

Find more in the experiment folder.