trainingparameters.yaml

# Description: Training parameters for the training script

# Model selection
model: 'CNNMNIST' # 'FCMNIST' or 'CNNMNST' This is the class name of the model as defined in models.py.
dataset: 'MNIST' # 'MNIST', or EMNIST splits: EMNIST_BALANCED, EMNIST_BYCLASS, EMNIST_BYMERGE, EMNIST_LETTERS, EMNIST_DIGITS, EMNIST_MNIST
  
# Quantization settings
QuantType: '4bitsym' # 'Ternary', 'Binary', 'BinaryBalanced', '2bitsym', '4bit', '4bitsym', '8bit', 'None", 'FP130', 'NF4'
NormType: 'RMS' # 'RMS', 'Lin', 'BatchNorm'
WScale: 'PerTensor' # 'PerTensor', 'PerOutput'

# Clipping parameters - only used for 2 bit and higher quantization
maxw_algo: 'octav' # 'octav', 'prop' Algorithm used to calculate the clipping parameters (maximum weight)
maxw_update_until_epoch: 60 # Update clipping parameters until this epoch, they are frozen afterwards
maxw_quantscale: 0.25  # Used only for clipping_algo='prop'. Determines the relation between stddev of weights and max_weight

# Learning parameters
num_epochs: 60  # 5, 20, 80
batch_size: 64
scheduler: "Cosine" # "StepLR", "Cosine", "CosineWarmRestarts"
learning_rate: 0.001

# CosineWarmRestarts parameters
# T_0: 5     # Period of the first restart for CosineWarmRestarts - 10+20+40 = 70 epochs, need to step in epoch 69 at minimum LR
# T_mult: 4  # Factor increasing T_i after a restart

# StepLR parameters
# lr_decay: 0.1     # lr_decay and step size for StepLR
# step_size: 10

# halve_lr_epoch: 30  # Epoch at which to halve the learning rate - to be used with Cosine schedule 

# Data augmentation
augmentation: True
rotation1: 10  # rotation1 and rotation2 are used for data augmentation
rotation2: 10
elastictransformprobability: 0.0 # probability of applying elastic transform

# channel pruning settings. Requires "MaskLayer" in the model, otherwise these settings have no effect
lambda_l1: 0.0005         # L1 regularization parameter for mask learning
prune_epoch: -1           # Epoch at which to start pruning. -1 means no pruning
prune_groupstoprune: 32   # number of groups to prune
prune_totalgroups: 96     # total number of groups. e.g. if there are 384 channels and 96 groups, then each group has 4 channels

# Model parameters
cnn_width: 64       # Width of CNN layers (CNNMNIST only)
network_width1: 96
network_width2: 64
network_width3: 0

# name
runtag: "octav" # runtag is prefix for runname