There are two configuration folders to operate DA2Lite.
The first is a train folder that contains .yaml files for training configuration options (e.g. model, dataset, optimizer, loss, ...).
And, a configuration file in compress folder allows to user select what compressipon methods will be used sequentially. Besides, when right after compression a model, you can control post-train configurations using its .yaml file.
The followings indicate example configurations and available choices of configurations (represented by comments).
DATASET:
NAME: cifar10 # Choice of ['cifar10', 'cifar100']
IMG_SHAPE: [3, 32, 32] # [Channel, Height, Width] of the dataset
DATA_AUG: True # If True, standard data augmentation will be used.
MODEL:
NAME: resnet50 #
NUM_CLASSES: 100 # The number of classes
PRE_TRAINED:
IS_USE: False # Load your pretrained model???
PATH: './cifar10_resnet50.pt' # Model path for the pretrained model path
TRAIN:
IS_USE: True # Train a model??
OPTIMIZER: sgd # Choice of ['sgd', 'adam']
OPTIMIZER_ARGS: # Arguments of a choosed optimizer
MOMENTUM : 0.9 #for sgd
WEIGHT_DECAY: 5e-4 #for sgd
"""
BETAS = (0.9, 0.999) for adam
EPSILON = 1e-8 for adam
"""
LOSS: categorical_crossentropy # Choice of [categorical_crossentropy, kld, mae, mse]
SCHEDULER: stepLR # [stepLR, exponentialLR, cosineannealLR]
SCHEDULER_ARGS: # Arguments of a choosed scheduler
STEP_SIZE: 60 # for stepLR
GAMMA: 0.1 # for stepLR
"""
GAMMA: 0.95 for exponentialLR
T_MAX: 140(The number of epochs) for cosineannealLR
"""
BATCH_SIZE: 128 # Batch size
EPOCHS: 140 # Epochs
LR: 0.1 # Learning rate
GPU:
IS_USE: True # Use CUDA GPU? PRUNING:
POST_TRAIN:
IS_USE: True # Use post-train after pruning?
OPTIMIZER: sgd # Choice of ['sgd', 'adam']
OPTIMIZER_ARGS: # Arguments of a choosed optimizer
MOMENTUM: 0.9 #for sgd
WEIGHT_DECAY: 5e-4 #for sgd
"""
BETAS = (0.9, 0.999) for adam
EPSILON = 1e-8 for adam
"""
LOSS: categorical_crossentropy # Choice of [categorical_crossentropy, kld, mae, mse]
SCHEDULER: cosineannealLR # [stepLR, exponentialLR, cosineannealLR]
SCHEDULER_ARGS: # Arguments of a choosed scheduler
T_MAX: 30 #for
"""
STEP_SIZE: 60 # for stepLR
GAMMA: 0.1 # for stepLR
GAMMA: 0.95 # for exponentialLR
"""
BATCH_SIZE: 256 # Batch size
EPOCHS: 30 # Epochs
LR: 0.001 # Learning rate
METHOD:
CRITERIA: EagleEye # Choice of ['EagleEye', 'L1Criteria', 'L2Criteria', 'RandomCriteria']
CRITERIA_ARGS:
NUM_CANDIDATES: 20 # for EagleEye
STRATEGY: MinMaxStrategy # Choice of ['MinMaxStrategy', 'RandomStrategy', 'StaticStrategy']
STRATEGY_ARGS:
PRUNING_RATIO: [0.0, 0.5] # for MinMaxStrategy
"""
PRUNING_RATIO: None # for RandomStrategy
PRUNING_RATIO: 0.5 # for StaticStrategy
"""
FD:
POST_TRAIN:
IS_USE: True # Use post-train after filter decomposition??
OPTIMIZER: sgd
OPTIMIZER_ARGS:
MOMENTUM: 0.9
WEIGHT_DECAY: 5e-4
LOSS: categorical_crossentropy
SCHEDULER: cosineannealLR
SCHEDULER_ARGS:
T_MAX: 30
BATCH_SIZE: 256
EPOCHS: 30
LR: 0.001
METHOD:
DECOMPOSITION: Tucker # Choice of ['Tucker']
START_IDX: 6 # Start index of filter decomposition (based on an order of conv layer)
RANK: 6 # Choice of ['4', '5', '6', 'VBMF']