def test_param_grid(self):
grid = ParamGrid([
('p1', [0, 1]),
('p2', ['a', 'b', 'c']),
('p3', [None, {}]),
])
param_combinations = grid.generate_params(randomize=True)
for p in param_combinations:
for key in ('p1', 'p2', 'p3'):
self.assertIn(key, p)
param_combinations = list(grid.generate_params(randomize=False))
self.assertEqual(param_combinations[0], {
'p1': 0,
'p2': 'a',
'p3': None
})
self.assertEqual(param_combinations[1], {'p1': 0, 'p2': 'a', 'p3': {}})
self.assertEqual(param_combinations[-2], {
'p1': 1,
'p2': 'c',
'p3': None
})
self.assertEqual(param_combinations[-1], {
'p1': 1,
'p2': 'c',
'p3': {}
})
def test_experiment(self):
params = ParamGrid([('p1', [3.14, 2.71]), ('p2', ['a', 'b', 'c'])])
cmd = 'python super_rl.py'
ex = Experiment('test', cmd, params.generate_params(randomize=False))
cmds = ex.generate_experiments()
for command, name in cmds:
self.assertTrue(command.startswith(cmd))
self.assertTrue(name.startswith('test'))
def test_descr(self):
params = ParamGrid([('p1', [3.14, 2.71]), ('p2', ['a', 'b', 'c'])])
experiments = [
Experiment('test1', 'python super_rl1.py',
params.generate_params(randomize=False)),
Experiment('test2', 'python super_rl2.py',
params.generate_params(randomize=False)),
]
rd = RunDescription('test_run', experiments)
cmds = rd.generate_experiments()
for command, name, root_dir in cmds:
self.assertIn('--experiment', command)
self.assertIn('--experiments_root', command)
self.assertTrue(name.startswith('test'))
self.assertTrue(root_dir.startswith('test_run'))
def test_simple_cmd(self):
logging.disable(logging.INFO)
echo_params = ParamGrid([
('p1', [3.14, 2.71]),
('p2', ['a', 'b', 'c']),
('p3', list(np.arange(3))),
])
experiments = [
Experiment('test_echo1', 'echo',
echo_params.generate_params(randomize=True)),
Experiment('test_echo2', 'echo',
echo_params.generate_params(randomize=False)),
]
rd = RunDescription('test_run', experiments, max_parallel=10)
run_many(rd)
logging.disable(logging.NOTSET)
from runner.run_descriptions.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('icm_beta', [0.2]),
])
_experiments = [
Experiment(
'doom_test_basic',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_basic --gpu_mem_fraction=0.1 --train_for_env_steps=80000',
_params.generate_params(randomize=False),
),
Experiment(
'doom_test_maze',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze --gpu_mem_fraction=0.1 --train_for_env_steps=80000',
_params.generate_params(randomize=False),
),
]
DOOM_TEST_RUN = RunDescription('doom_test',
experiments=_experiments,
max_parallel=3)
from runner.run_descriptions.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ignore_timer', [True, False]),
])
_experiments = [
Experiment(
'doom_maze',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze --gpu_mem_fraction=0.1 --train_for_env_steps=80000000',
_params.generate_params(randomize=False),
),
Experiment(
'doom_basic',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_basic --gpu_mem_fraction=0.1 --train_for_env_steps=10000000',
_params.generate_params(randomize=False),
),
]
DOOM_TIMER = RunDescription('doom_timer',
experiments=_experiments,
max_parallel=2)
from runner.run_descriptions.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('prediction_bonus_coeff', [0.00, 0.05]),
])
_experiments = [
Experiment(
'doom_maze_very_sparse',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze_very_sparse --gpu_mem_fraction=0.1 --train_for_env_steps=2000000000',
_params.generate_params(randomize=False),
),
# Experiment(
# 'doom_maze_sparse',
# 'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze_sparse --gpu_mem_fraction=0.1 --train_for_env_steps=100000000',
# _params.generate_params(randomize=False),
# ),
# Experiment(
# 'doom_maze',
# 'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze --gpu_mem_fraction=0.1 --train_for_env_steps=50000000',
# _params.generate_params(randomize=False),
# ),
# Experiment(
# 'doom_basic',
# 'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_basic --gpu_mem_fraction=0.1 --train_for_env_steps=10000000',
# _params.generate_params(randomize=False),
# ),
]
DOOM_CURIOUS_VS_VANILLA = RunDescription('doom_curious_vs_vanilla',
experiments=_experiments,
from runner.run_descriptions.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('icm_beta', [0.2, 0.5, 0.8]),
('prediction_bonus_coeff', [0.01, 0.02, 0.05]),
])
_experiment = Experiment(
'doom_sweep',
'python -m algorithms.curious_a2c.train_curious_a2c --env=doom_maze_very_sparse --gpu_mem_fraction=0.1 --train_for_env_steps=80000000',
_params.generate_params(randomize=False),
)
DOOM_SWEEP_RUN = RunDescription('doom_sweep',
experiments=[_experiment],
max_parallel=5)