config_cosine_restarts.yaml

# Cosine schedule with hard restarts configuration
# Best for: Training that benefits from periodic LR resets (helps escape local minima)
# To use: python learn_monodecoder.py --config config_cosine_restarts.yaml

# Dataset and data loading
dataset: structs_traininffttest.h5
batch_size: 8
gradient_accumulation_steps: 3  # Effective batch size = 24

# Model architecture
hidden_size: 256
embedding_dim: 128
num_embeddings: 40
se3_transformer: false
hetero_gae: false

# Training parameters
epochs: 150
learning_rate: 0.0002  # Higher initial LR works well with restarts
clip_grad: true
burn_in: 0
EMA: true

# Learning rate scheduling - Cosine with hard restarts
lr_schedule: cosine_restarts
lr_warmup_ratio: 0.05  # 5% warmup period
lr_warmup_steps: 0
lr_min: 1.0e-06
num_cycles: 5  # 5 restart cycles over training

# Commitment cost scheduling
commitment_cost: 0.95
use_commitment_scheduling: true
commitment_schedule: cosine
commitment_warmup_steps: 3000  # Shorter warmup for restarts
commitment_start: 0.2  # Higher start for stability with restarts

# Output options
output_fft: false
output_rt: false
output_foldx: false

# Directories and naming
output_dir: ./models/
model_name: monodecoder_restarts_model
tensorboard_dir: ./runs/
run_name: cosine_restarts_experiment

# System settings
device: null
seed: 42
overwrite: false

# Notes:
# - Cosine with restarts periodically resets LR to initial value
# - Each cycle gets progressively longer
# - Good for finding better local minima during training
# - num_cycles=5 means 5 restarts over the entire training