Files
Gymnasium/tests/envs/test_determinism.py

84 lines
2.8 KiB
Python
Raw Normal View History

"""Test environment determinism by performing a rollout."""
import pytest
from gym.utils.env_checker import data_equivalence
from tests.envs.spec_list import spec_list
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
2021-07-29 02:26:34 +02:00
@pytest.mark.parametrize("spec", spec_list, ids=[spec.id for spec in spec_list])
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
def test_env(spec):
"""Run a rollout with two environments and assert equality.
This test run a rollout of NUM_STEPS steps with two environments
initialized with the same seed and assert that:
- observation after first reset are the same
- same actions are sampled by the two envs
- observations are contained in the observation space
- obs, rew, done and info are equals between the two envs
Args:
spec (EnvSpec): Environment specification
"""
# Note that this precludes running this test in multiple
# threads. However, we probably already can't do multithreading
# due to some environments.
SEED = 0
NUM_STEPS = 50
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
env1, env2 = spec.make(), spec.make()
initial_observation1 = env1.reset(seed=SEED)
initial_observation2 = env2.reset(seed=SEED)
env1.action_space.seed(SEED)
env2.action_space.seed(SEED)
assert data_equivalence(
initial_observation1, initial_observation2
), f"Initial Observations 1 and 2 are not equivalent. initial obs 1={initial_observation1}, initial obs 2={initial_observation2}"
for i in range(NUM_STEPS):
action1 = env1.action_space.sample()
action2 = env2.action_space.sample()
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
2017-03-28 11:25:21 -07:00
try:
assert data_equivalence(
action1, action2
), f"Action 1 and 2 are not equivalent. action 1={action1}, action 2={action2}"
2017-03-28 11:25:21 -07:00
except AssertionError:
print(f"env 1 action space={env1.action_space}")
print(f"env 2 action space={env2.action_space}")
print(f"[{i}] action sample 1={action1}, action sample 2={action2}")
2017-03-28 11:25:21 -07:00
raise
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
# Don't check rollout equality if it's a nondeterministic
# environment.
if spec.nondeterministic:
return
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
obs1, rew1, done1, info1 = env1.step(action1)
obs2, rew2, done2, info2 = env2.step(action2)
assert data_equivalence(
obs1, obs2
), f"Observation 1 and 2 are not equivalent. obs 1={obs1}, obs 2={obs2}"
assert env1.observation_space.contains(obs1)
assert env2.observation_space.contains(obs2)
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
assert rew1 == rew2, f"[{i}] reward1: {rew1}, reward2: {rew2}"
assert done1 == done2, f"[{i}] done1: {done1}, done2: {done2}"
assert data_equivalence(
info1, info2
), f"Info 1 and 2 are not equivalent. info 1={info1}, info 2={info2}"
[WIP] add support for seeding environments (#135) * Make environments seedable * Fix monitor bugs - Set monitor_id before setting the infix. This was a bug that would yield incorrect results with multiple monitors. - Remove extra pid from stats recorder filename. This should be purely cosmetic. * Start uploading seeds in episode_batch * Fix _bigint_from_bytes for python3 * Set seed explicitly in random_agent * Pass through seed argument * Also pass through random state to spaces * Pass random state into the observation/action spaces * Make all _seed methods return the list of used seeds * Switch over to np.random where possible * Start hashing seeds, and also seed doom engine * Fixup seeding determinism in many cases * Seed before loading the ROM * Make seeding more Python3 friendly * Make the MuJoCo skipping a bit more forgiving * Remove debugging PDB calls * Make setInt argument into raw bytes * Validate and upload seeds * Skip box2d * Make seeds smaller, and change representation of seeds in upload * Handle long seeds * Fix RandomAgent example to be deterministic * Handle integer types correctly in Python2 and Python3 * Try caching pip * Try adding swap * Add df and free calls * Bump swap * Bump swap size * Try setting overcommit * Try other sysctls * Try fixing overcommit * Try just setting overcommit_memory=1 * Add explanatory comment * Add what's new section to readme * BUG: Mark ElevatorAction-ram-v0 as non-deterministic for now * Document seed * Move nondetermistic check into spec
2016-05-29 09:07:09 -07:00
if done1: # done2 verified in previous assertion
env1.reset(seed=SEED)
env2.reset(seed=SEED)
env1.close()
env2.close()