Parameter Sweeps

Sweeps run multiple training configurations in parallel. Define which hyperparameters to search, and Canard creates one task per combination.

Basic Sweep

from canard import Client

client = Client(api_url="https://api.canard.cloud")

run = client.submit_training_sweep(
    name="lr-sweep",
    task_name="Template-Go2-Standing-Direct-v0",
    code_url="https://github.com/canard-cloud/go2-standing-env",
    sweep_params={
        "learning_rate": [1e-4, 3e-4, 1e-3, 3e-3],
    },
    num_envs=4096,
    max_iterations=5000,
    gpu_count=2,
)

This creates 4 tasks, one per learning rate. They run in parallel on separate GPUs.

Submitting training sweep: lr-sweep
  Configs: 4
  Est. ~$7.23/task · ~$28.92 total · ~7.2 hr wall time

Multi-Parameter Sweep

Sweep across multiple parameters — Canard creates the full grid:

run = client.submit_training_sweep(
    name="lr-entropy-sweep",
    task_name="Template-Go2-Standing-Direct-v0",
    code_url="https://github.com/canard-cloud/go2-standing-env",
    sweep_params={
        "learning_rate": [1e-4, 3e-4, 1e-3],
        "entropy_coef": [0.005, 0.01, 0.05],
    },
    num_envs=4096,
    max_iterations=5000,
    gpu_count=2,
)
# Creates 3 x 3 = 9 tasks

Multi-Seed Evaluation

For publishable results, conferences require multiple seeds per configuration. Add a seed sweep:

run = client.submit_training_sweep(
    name="publication-seeds",
    task_name="Template-Go2-Standing-Direct-v0",
    code_url="https://github.com/canard-cloud/go2-standing-env",
    sweep_params={
        "learning_rate": [1e-4, 3e-4, 1e-3],
        "seed": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
    },
    num_envs=4096,
    max_iterations=5000,
    gpu_count=2,
)
# 3 LRs x 10 seeds = 30 tasks

Sweepable Parameters

Any TrainingConfig field can be swept:

Parameter	Type	Example Values
learning_rate	float	[1e-4, 3e-4, 1e-3]
clip_param	float	[0.1, 0.2, 0.3]
entropy_coef	float	[0.005, 0.01, 0.05]
gamma	float	[0.95, 0.99, 0.999]
lam	float	[0.9, 0.95, 0.99]
desired_kl	float	[0.005, 0.01, 0.02]
num_learning_epochs	int	[3, 5, 8]
num_mini_batches	int	[2, 4, 8]
seed	int	[1, 2, 3, 4, 5]

Analyzing Results

After the sweep completes:

run.wait_for_completion(show_progress=True)

# Get all tasks with their configs and results
tasks = run.get_tasks()

for task in tasks:
    if task.status == "completed" and task.metrics:
        lr = task.params.get("learning_rate", "default")
        reward = task.metrics.training_progress.mean_reward
        print(f"LR={lr}: reward={reward:.2f}")

Cost Estimation

Before submitting, estimate the cost:

from canard.models import TrainingConfig

config = TrainingConfig(
    task_name="Template-Go2-Standing-Direct-v0",
    num_envs=4096,
    max_iterations=5000,
    gpu_count=2,
)

est = config.estimate_cost(total_tasks=30)
print(f"Per task: ${est['task_cost']:.2f}")
print(f"Total:   ${est['total_cost']:.2f}")
print(f"Time:    ~{est['wall_time_hr']:.1f} hr (parallel)")

See Cost Estimation for details on the pricing model.