def test_gym_task_reset(mock_gym, fix_env):
# Assign
mock_gym.make.return_value = fix_env
task = GymTask("example")
# Act
out = task.reset()
# Assert
assert fix_env.reset.called_once()
assert len(out) > 0
def state_transform(state):
"""
Simple cropping of the top and bottom edge and converting to blackwhite scale.
"""
return (state[40:-10].sum(-1) > 0)[None, ...]
def agent_state_tranform(state):
return state
env_name = 'Breakout-v0'
task = GymTask(env_name, state_transform=state_transform)
state_size = np.array(task.reset()).shape
writer = SummaryWriter()
config = {
"update_freq":
10,
"batch_size":
100,
"warm_up":
100,
"lr":
1e-4,
"network_fn":
lambda: QNetwork2D(state_size, task.action_size, hidden_layers=(200, 200)),
"state_transform":
agent_state_tranform,
"gamma":
0.99,
"lr":
1e-3,
"update_freq":
150,
"batch_size":
400,
"buffer_size":
int(5e3),
"device":
"cuda",
}
for _ in range(prev_states):
task.reset()
agent = DQNAgent(task.state_size, task.action_size, **config)
env_runner = EnvRunner(task, agent, data_logger=data_logger)
# env_runner.interact_episode(0, render=True)
scores = env_runner.run(
reward_goal=1000,
max_episodes=20000,
log_every=1,
eps_start=0.9,
gif_every_episodes=200,
force_new=True,
)
# env_runner.interact_episode(render=True)
data_logger.close()