# TinyTorch Default Configuration
# This file contains default settings for training neural networks

# Model configuration
model:
  type: "mlp"  # mlp, cnn
  input_size: 784  # for MNIST (28*28)
  hidden_sizes: [128, 64]
  output_size: 10
  activation: "relu"
  dropout: 0.0

# CNN-specific settings (used when model.type == "cnn")
cnn:
  channels: [32, 64]
  kernel_sizes: [3, 3]
  pool_sizes: [2, 2]
  fc_hidden: 128

# Dataset configuration
dataset:
  name: "mnist"  # mnist, cifar10
  batch_size: 64
  shuffle: true
  num_workers: 0
  download: true
  data_dir: "./data"

# Data augmentation (for training)
augmentation:
  enabled: false
  transforms:
    - "random_crop"
    - "random_flip"

# Training configuration
training:
  epochs: 10
  learning_rate: 0.001
  optimizer: "adam"  # sgd, adam
  weight_decay: 0.0
  momentum: 0.9  # for SGD
  
  # Learning rate scheduling
  lr_scheduler:
    enabled: false
    type: "step"  # step, cosine
    step_size: 3
    gamma: 0.1

# Validation and testing
validation:
  enabled: true
  split: 0.1  # fraction of training data for validation
  frequency: 1  # validate every N epochs

# Checkpointing
checkpointing:
  enabled: true
  save_best: true
  save_frequency: 5  # save every N epochs
  max_checkpoints: 3  # keep only N latest checkpoints

# Logging and monitoring
logging:
  level: "INFO"  # DEBUG, INFO, WARNING, ERROR
  log_frequency: 100  # log every N batches
  metrics:
    - "loss"
    - "accuracy"
    - "learning_rate"
  
  # Profiling
  profiling:
    enabled: false
    memory: true
    compute: true

# System configuration
system:
  device: "auto"  # auto, cpu, cuda
  mixed_precision: false
  seed: 42
  deterministic: false