Configuration files are structured in 6 parts as shown in the following example:
# 1. Experiment
experiment:
name (string): Name of the experiment.
output_dir (string): Folder where the outputs of the experiment will be saved.
# For reproducibility
seed (int): Random seed for the experiment.
deterministic (boolean): If true, will apply PyTorch deterministic operations to guarantee reproducibility.# 2. Data
data:
dataset (string): Name of the dataset
dataset_path (string): If dataset is not supported by Torchvision, a path can be provided (e.g., ImageNet)
target_size (int): Images will be resized to the target size (if specified): 32
num_workers (int): Number of threads for data loading # 3. Model
model:
architecture (string): Name of the model architecture
mode:
type (string): Name of the feedback alignment method
options:
init (string): Layer initialization method
gradient_clip (boolean): If true, will apply gradient clipping (-1, 1)
pretrained (boolean): If true, will load a pre-trained model and copy the original weights to the new layers
# checkpoint (string): Path to a custom model
loss_function:
name (string) : Name of the loss function# 4. Training Parameters
training:
hyperparameters:
epochs (int): Number of epochs to run the training
batch_size (int): Data batch size
optimizer:
type: "SGD"
lr: 0.1
weight_decay: 0.0001
momentum: 0.9
lr_scheduler:
type: "multistep_lr"
gamma: 0.1
milestones: [100, 150, 200]
metrics:
top_k: 5
display_iterations: 500
weight_alignment (boolean): Track the angle between layers forward and backward weights
weight_ratio (boolean): Track the norm ratio between layers forward and backward weights# 5. Infrastructure
infrastructure:
gpus (int/list): If -1 will use CPU, else will use the corresponding GPUs devices# 6. Evaluation
evaluation (boolean): If true, it will run an evaluation in the test and store the metrics after the training