configs/paper.toml

# DEF-roboticattack — Paper-aligned defense detector training config
# Defense model: PatchDetectorNet (multi-branch CNN)
# Task: binary classification (clean vs adversarial-patched VLA inputs)

[training]
batch_size = "auto"
learning_rate = 0.001
epochs = 50
optimizer = "adamw"
weight_decay = 0.01
scheduler = "cosine"
warmup_steps = 200
precision = "fp16"
gradient_accumulation = 1
max_grad_norm = 1.0
seed = 42

[model]
in_channels = 3
image_size = 224

[data]
num_train_samples = 50000
num_val_samples = 5000
patch_ratio_min = 0.01
patch_ratio_max = 0.20
attack_prob = 0.5
num_workers = 4
pin_memory = true

[checkpoint]
save_every_n_steps = 500
keep_top_k = 2
metric = "val_loss"
mode = "min"
save_dir = "/mnt/artifacts-datai/checkpoints/DEF-roboticattack"

[early_stopping]
enabled = true
patience = 10
min_delta = 0.001

[defense]
edge_threshold = 0.15
clamp_percentile = 99.5
blur_strength = 0.15

[logging]
log_dir = "/mnt/artifacts-datai/logs/DEF-roboticattack"
log_every_n_steps = 50