Renamed VoxelEnv to Megaverse

Author: alex-petrenko

Makefile

@@ -1,6 +1,5 @@
 PYTHON ?= python
 
-export REPO=voxel-rl
 export BASE_TAG=$(shell ${PYTHON} -c 'import hashlib; sha = hashlib.sha1((open("docker/Dockerfile.base").read() + open("requirements/requirements.txt").read()).encode()); print(sha.hexdigest())')
 BRANCH = $(shell git rev-parse --abbrev-ref HEAD)
 VERSION = $(shell git rev-parse --short HEAD)

README.md

@@ -1,8 +1,8 @@
-# voxel-rl
+# Megaverse
 
 ```
 1) Clone the repo
-git clone https://github.com/alex-petrenko/voxel-rl.git
+git clone https://github.com/alex-petrenko/megaverse.git
 
 2) Init submodules
 git submodule update --init --recursive
@@ -60,7 +60,7 @@ hash -r
 sudo snap install cmake --classic
 
 9) Build the repo
-cd voxel-rl
+cd megaverse
 mkdir build
 cd build
 
@@ -78,9 +78,9 @@ make -j10
 
 10) Run benchmark
 cd Release/bin
-./voxel_env_app
+./megaverse_test_app
 
-(see global boolean flags in voxel_env_app.cpp, they control the scenario and rendering settings
+(see global boolean flags in megaverse_test_app.cpp, they control the scenario and rendering settings
 TODO: make configurable)
 
 11) Run viewer
@@ -103,7 +103,7 @@ pip install -e .
 13) Run tests
 python -m unittest
 
-14) You are ready to use the VoxelWorld Python API!
+14) You are ready to use the Megaverse Python API!
 
 ```
 
@@ -117,7 +117,7 @@ python -m megaverse_rl.train --train_for_seconds=360000000 --train_for_env_steps
 
 Example runner script:
 
-python -m sample_factory.runner.run --run=megaverse_rl.runs.voxel_single_agent --runner=processes --max_parallel=8 --pause_between=10 --experiments_per_gpu=2 --num_gpus=4
+python -m sample_factory.runner.run --run=megaverse_rl.runs.megaverse_single_agent --runner=processes --max_parallel=8 --pause_between=10 --experiments_per_gpu=2 --num_gpus=4
 
 
 ```

voxel_env/.gitignore renamed to megaverse/.gitignore

@@ -5,10 +5,10 @@
 from gym.spaces import Discrete
 
 # noinspection PyUnresolvedReferences
-from voxel_env.extension.voxel_env import VoxelEnvGym, set_voxel_env_log_level
+from megaverse.extension.megaverse import MegaverseGym, set_megaverse_log_level
 
 
-VOXELWORLD8 = [
+MEGAVERSE8 = [
     'TowerBuilding',
     'ObstaclesEasy',
     'ObstaclesHard',
@@ -26,8 +26,8 @@
 
 def make_env_multitask(multitask_name, task_idx, num_envs, num_agents_per_env, num_simulation_threads, use_vulkan=False, params=None):
     assert 'multitask' in multitask_name
-    if multitask_name.endswith('voxelworld8'):
-        tasks = VOXELWORLD8
+    if multitask_name.endswith('megaverse8'):
+        tasks = MEGAVERSE8
     elif multitask_name.endswith('obstacles'):
         tasks = OBSTACLES_MULTITASK
     else:
@@ -36,17 +36,17 @@ def make_env_multitask(multitask_name, task_idx, num_envs, num_agents_per_env, n
     scenario_idx = task_idx % len(tasks)
     scenario = tasks[scenario_idx]
     print('Multi-task, scenario', scenario_idx, scenario)
-    return VoxelEnv(scenario, num_envs, num_agents_per_env, num_simulation_threads, use_vulkan, params)
+    return MegaverseEnv(scenario, num_envs, num_agents_per_env, num_simulation_threads, use_vulkan, params)
 
 
-class VoxelEnv(gym.Env):
+class MegaverseEnv(gym.Env):
     def __init__(self, scenario_name, num_envs, num_agents_per_env, num_simulation_threads, use_vulkan=False, params=None):
         scenario_name = scenario_name.casefold()
         self.scenario_name = scenario_name
 
         self.is_multiagent = True
 
-        set_voxel_env_log_level(2)
+        set_megaverse_log_level(2)
 
         self.img_w = 128
         self.img_h = 72
@@ -69,7 +69,7 @@ def __init__(self, scenario_name, num_envs, num_agents_per_env, num_simulation_t
 
         # float_params['episodeLengthSec'] = 1.0
 
-        self.env = VoxelEnvGym(
+        self.env = MegaverseGym(
             self.scenario_name,
             self.img_w, self.img_h, num_envs, num_agents_per_env, num_simulation_threads, use_vulkan, float_params,
         )

voxel_env/__init__.py renamed to megaverse/__init__.py

@@ -6,7 +6,7 @@
 
 from unittest import TestCase
 
-from voxel_env.voxel_env_gym import VoxelEnv, make_env_multitask
+from megaverse.megaverse_env import MegaverseEnv, make_env_multitask
 
 
 def sample_actions(e):
@@ -15,7 +15,7 @@ def sample_actions(e):
 
 def make_test_env(num_envs, num_agents_per_env, num_simulation_threads, use_vulkan=False, params=None):
     """Making env with a default scenario name."""
-    return VoxelEnv('ObstaclesEasy', num_envs, num_agents_per_env, num_simulation_threads, use_vulkan, params)
+    return MegaverseEnv('ObstaclesEasy', num_envs, num_agents_per_env, num_simulation_threads, use_vulkan, params)
 
 
 class TestEnv(TestCase):
@@ -121,7 +121,7 @@ def test_performance(self):
         # print(fps1, fps2, fps4)
 
     def test_reward_shaping(self):
-        e = VoxelEnv('TowerBuilding', num_envs=3, num_agents_per_env=2, num_simulation_threads=2, use_vulkan=True)
+        e = MegaverseEnv('TowerBuilding', num_envs=3, num_agents_per_env=2, num_simulation_threads=2, use_vulkan=True)
         default_reward_shaping = e.get_default_reward_shaping()
         self.assertEqual(default_reward_shaping, e.get_current_reward_shaping(0))
         self.assertEqual(default_reward_shaping, e.get_current_reward_shaping(1))
@@ -147,12 +147,12 @@ def mem_usage_kb():
 
         # params = {'episodeLengthSec': 0.1}
         params = {}
-        e = VoxelEnv('Rearrange', num_envs=32, num_agents_per_env=1, num_simulation_threads=1, use_vulkan=True, params=params)
+        e = MegaverseEnv('Rearrange', num_envs=32, num_agents_per_env=1, num_simulation_threads=1, use_vulkan=True, params=params)
         e.reset()
 
         orig_mem_usage = mem_usage_kb()
 
-        for i in range(10000):
+        for i in range(1000):
             print('Mem difference: ', mem_usage_kb() - orig_mem_usage, 'kb')
             e.step(sample_actions(e))
 
@@ -165,7 +165,7 @@ def test_multitask(self):
         num_processes = 2
 
         def run_single_task(i):
-            e = make_env_multitask('voxelworld8', i, 1, 1, 1, use_vulkan=True, params={})
+            e = make_env_multitask('megaverse8', i, 1, 1, 1, use_vulkan=True, params={})
             e.reset()
             e.render()  # TODO: if this call is omitted we have rendering bugs. Fixme!
 
@@ -186,11 +186,11 @@ def run_single_task(i):
 
     def test_viewer(self):
         params = {'episodeLengthSec': 1.0}
-        e1 = VoxelEnv('ObstaclesHard', 2, 2, 2, True, params)
+        e1 = MegaverseEnv('ObstaclesHard', 2, 2, 2, True, params)
         e1.reset()
         e1.render()
 
-        for i in range(10000):
+        for i in range(500):
             e1.step(sample_actions(e1))
             e1.render()
             time.sleep(0.01)

voxel_env/voxel_env_gym.py renamed to megaverse/megaverse_env.py

@@ -3,7 +3,7 @@
 from sample_factory.algorithms.appo.enjoy_appo import enjoy
 from sample_factory.algorithms.utils.arguments import parse_args
 
-from megaverse_rl.voxel_env_utils import register_env
+from megaverse_rl.megaverse_utils import register_env
 
 
 def main():

voxel_env/tests/__init__.py renamed to megaverse/tests/__init__.py

@@ -1,7 +1,7 @@
 import gym
 from sample_factory.envs.env_registry import global_env_registry
 
-from voxel_env.voxel_env_gym import VoxelEnv, make_env_multitask
+from megaverse.megaverse_env import MegaverseEnv, make_env_multitask
 
 from sample_factory.envs.env_utils import RewardShapingInterface, TrainingInfoInterface
 from sample_factory.utils.utils import str2bool, log
@@ -64,8 +64,8 @@ def step(self, action):
         return obs, rewards, dones, infos
 
 
-def make_voxel_env(env_name, cfg=None, env_config=None, **kwargs):
-    scenario_name = env_name.split('voxel_env_')[-1].casefold()
+def make_megaverse(env_name, cfg=None, env_config=None, **kwargs):
+    scenario_name = env_name.split('megaverse_')[-1].casefold()
     log.debug('Using scenario %s', scenario_name)
 
     if 'multitask' in scenario_name:
@@ -78,26 +78,26 @@ def make_voxel_env(env_name, cfg=None, env_config=None, **kwargs):
         env = make_env_multitask(
             scenario_name,
             task_idx,
-            num_envs=cfg.voxel_num_envs_per_instance,
-            num_agents_per_env=cfg.voxel_num_agents_per_env,
-            num_simulation_threads=cfg.voxel_num_simulation_threads,
-            use_vulkan=cfg.voxel_use_vulkan,
+            num_envs=cfg.megaverse_num_envs_per_instance,
+            num_agents_per_env=cfg.megaverse_num_agents_per_env,
+            num_simulation_threads=cfg.megaverse_num_simulation_threads,
+            use_vulkan=cfg.megaverse_use_vulkan,
         )
     else:
-        env = VoxelEnv(
+        env = MegaverseEnv(
             scenario_name=scenario_name,
-            num_envs=cfg.voxel_num_envs_per_instance,
-            num_agents_per_env=cfg.voxel_num_agents_per_env,
-            num_simulation_threads=cfg.voxel_num_simulation_threads,
-            use_vulkan=cfg.voxel_use_vulkan,
+            num_envs=cfg.megaverse_num_envs_per_instance,
+            num_agents_per_env=cfg.megaverse_num_agents_per_env,
+            num_simulation_threads=cfg.megaverse_num_simulation_threads,
+            use_vulkan=cfg.megaverse_use_vulkan,
         )
 
-    env = Wrapper(env, cfg.voxel_increase_team_spirit, cfg.voxel_max_team_spirit_steps)
+    env = Wrapper(env, cfg.megaverse_increase_team_spirit, cfg.megaverse_max_team_spirit_steps)
     return env
 
 
-def voxel_env_override_defaults(env, parser):
-    """RL params specific to VoxelEnv envs."""
+def megaverse_override_defaults(env, parser):
+    """RL params specific to Megaverse envs."""
     parser.set_defaults(
         encoder_type='conv',
         encoder_subtype='convnet_simple',
@@ -110,22 +110,22 @@ def voxel_env_override_defaults(env, parser):
     )
 
 
-def add_voxel_env_args(env, parser):
+def add_megaverse_args(env, parser):
     p = parser
-    p.add_argument('--voxel_num_envs_per_instance', default=1, type=int, help='Num simulated envs per instance of VoxelEnv')
-    p.add_argument('--voxel_num_agents_per_env', default=4, type=int, help='Number of agents in a single env withing a VoxelEnv instance. Total number of agents in one VoxelEnv = num_envs_per_instance * num_agents_per_env')
-    p.add_argument('--voxel_num_simulation_threads', default=1, type=int, help='Number of CPU threads to use per instance of VoxelEnv')
-    p.add_argument('--voxel_use_vulkan', default=True, type=str2bool, help='Whether to use Vulkan renderer')
+    p.add_argument('--megaverse_num_envs_per_instance', default=1, type=int, help='Num simulated envs per instance of Megaverse')
+    p.add_argument('--megaverse_num_agents_per_env', default=4, type=int, help='Number of agents in a single env withing a Megaverse instance. Total number of agents in one Megaverse = num_envs_per_instance * num_agents_per_env')
+    p.add_argument('--megaverse_num_simulation_threads', default=1, type=int, help='Number of CPU threads to use per instance of Megaverse')
+    p.add_argument('--megaverse_use_vulkan', default=True, type=str2bool, help='Whether to use Vulkan renderer')
 
     # Team Spirit options
-    p.add_argument('--voxel_increase_team_spirit', default=False, type=str2bool, help='Increase team spirit from 0 to 1 over max_team_spirit_steps during training. At 1, the reward will be completely selfless.')
-    p.add_argument('--voxel_max_team_spirit_steps', default=1e9, type=float, help='Number of training steps when team spirit will hit 1.')
+    p.add_argument('--megaverse_increase_team_spirit', default=False, type=str2bool, help='Increase team spirit from 0 to 1 over max_team_spirit_steps during training. At 1, the reward will be completely selfless.')
+    p.add_argument('--megaverse_max_team_spirit_steps', default=1e9, type=float, help='Number of training steps when team spirit will hit 1.')
 
 
 def register_env():
     global_env_registry().register_env(
-        env_name_prefix='voxel_env_',
-        make_env_func=make_voxel_env,
-        add_extra_params_func=add_voxel_env_args,
-        override_default_params_func=voxel_env_override_defaults,
+        env_name_prefix='megaverse_',
+        make_env_func=make_megaverse,
+        add_extra_params_func=add_megaverse_args,
+        override_default_params_func=megaverse_override_defaults,
     )

voxel_env/tests/test_env.py renamed to megaverse/tests/test_env.py

@@ -1,27 +1,27 @@
 from sample_factory.runner.run_description import Experiment, ParamGrid
 
 _params = ParamGrid([
-    ('env', ['voxel_env_TowerBuilding', 'voxel_env_ObstaclesEasy', 'voxel_env_ObstaclesHard', 'voxel_env_Collect', 'voxel_env_Sokoban', 'voxel_env_HexMemory', 'voxel_env_HexExplore', 'voxel_env_Rearrange']),
+    ('env', ['megaverse_TowerBuilding', 'megaverse_ObstaclesEasy', 'megaverse_ObstaclesHard', 'megaverse_Collect', 'megaverse_Sokoban', 'megaverse_HexMemory', 'megaverse_HexExplore', 'megaverse_Rearrange']),
     ('use_cpc', ['True']),
     ('seed', [11111, 22222, 33333]),
 ])
 
-_cli = 'python -m megaverse_rl.train --train_for_seconds=360000000 --train_for_env_steps=2000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --voxel_num_simulation_threads=1 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=1 --reward_clip=30'
+_cli = 'python -m megaverse_rl.train --train_for_seconds=360000000 --train_for_env_steps=2000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --megaverse_num_simulation_threads=1 --megaverse_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=1 --reward_clip=30'
 
 EXPERIMENT_1AGENT = Experiment(
-    'voxel_env_1ag',
-    _cli + ' --voxel_num_envs_per_instance=36 --voxel_num_agents_per_env=1',
+    'megaverse_1ag',
+    _cli + ' --megaverse_num_envs_per_instance=36 --megaverse_num_agents_per_env=1',
     _params.generate_params(randomize=False),
 )
 
 EXPERIMENT_2AGENTS = Experiment(
-    'voxel_env_2ag',
-    _cli + ' --voxel_num_envs_per_instance=18 --voxel_num_agents_per_env=2',
+    'megaverse_2ag',
+    _cli + ' --megaverse_num_envs_per_instance=18 --megaverse_num_agents_per_env=2',
     _params.generate_params(randomize=False),
 )
 
 EXPERIMENT_4AGENTS = Experiment(
-    'voxel_env_4ag',
-    _cli + ' --voxel_num_envs_per_instance=9 --voxel_num_agents_per_env=4',
+    'megaverse_4ag',
+    _cli + ' --megaverse_num_envs_per_instance=9 --megaverse_num_agents_per_env=4',
     _params.generate_params(randomize=False),
 )

megaverse_rl/enjoy.py

@@ -1,15 +1,15 @@
 from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
 
 _params = ParamGrid([
-    ('env', ['voxel_env_rearrange', 'voxel_env_collect', 'voxel_env_obstaclesEasy', 'voxel_env_hexMemory']),
-    ('voxel_num_simulation_threads', [1]),
+    ('env', ['megaverse_rearrange', 'megaverse_collect', 'megaverse_obstaclesEasy', 'megaverse_hexMemory']),
+    ('megaverse_num_simulation_threads', [1]),
     ('rnn_num_layers', [2]),
 ])
 
 _experiment = Experiment(
-    'voxel_env_pbt',
-    'python -m megaverse_rl.train --train_for_seconds=360000000 --algo=APPO --gamma=0.997 --use_rnn=True --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --voxel_num_envs_per_instance=36 --voxel_num_agents_per_env=1 --voxel_num_simulation_threads=1 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=4',
+    'megaverse_pbt',
+    'python -m megaverse_rl.train --train_for_seconds=360000000 --algo=APPO --gamma=0.997 --use_rnn=True --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --megaverse_num_envs_per_instance=36 --megaverse_num_agents_per_env=1 --megaverse_num_simulation_threads=1 --megaverse_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=4',
     _params.generate_params(randomize=False),
 )
 
-RUN_DESCRIPTION = RunDescription('voxel_env_v114_env_v52', experiments=[_experiment])
+RUN_DESCRIPTION = RunDescription('megaverse_v114_env_v52', experiments=[_experiment])

megaverse_rl/voxel_env_utils.py renamed to megaverse_rl/megaverse_utils.py

@@ -1,4 +1,5 @@
 from sample_factory.runner.run_description import RunDescription
-from sample_factory.runner.runs.voxel_base_experiments import EXPERIMENT_2AGENTS, EXPERIMENT_4AGENTS
 
-RUN_DESCRIPTION = RunDescription('voxel_env_v115_multi_agent_v55', experiments=[EXPERIMENT_2AGENTS, EXPERIMENT_4AGENTS])
+from megaverse_rl.runs.megaverse_base_experiments import EXPERIMENT_4AGENTS, EXPERIMENT_2AGENTS
+
+RUN_DESCRIPTION = RunDescription('megaverse_v115_multi_agent_v55', experiments=[EXPERIMENT_2AGENTS, EXPERIMENT_4AGENTS])

megaverse_rl/runs/voxel_base_experiments.py renamed to megaverse_rl/runs/megaverse_base_experiments.py

@@ -1,17 +1,17 @@
 from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
 
 _params = ParamGrid([
-    ('env', ['voxel_env_multitask_voxelworld8']),
+    ('env', ['megaverse_multitask_megaverse8']),
     ('use_cpc', ['True']),
     ('seed', [11111, 22222, 33333, 44444, 55555]),
 ])
 
-_cli = 'python -m megaverse_rl.train --train_for_seconds=360000000 --train_for_env_steps=2000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --voxel_num_simulation_threads=1 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=1 --reward_clip=30 --pbt_mix_policies_in_one_env=False'
+_cli = 'python -m megaverse_rl.train --train_for_seconds=360000000 --train_for_env_steps=2000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --megaverse_num_simulation_threads=1 --megaverse_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=1 --reward_clip=30 --pbt_mix_policies_in_one_env=False'
 
 EXPERIMENT_1AGENT = Experiment(
-    'voxel_env_multitask_obs',
-    _cli + ' --voxel_num_envs_per_instance=36 --voxel_num_agents_per_env=1',
+    'megaverse_multitask_obs',
+    _cli + ' --megaverse_num_envs_per_instance=36 --megaverse_num_agents_per_env=1',
     _params.generate_params(randomize=False),
 )
 
-RUN_DESCRIPTION = RunDescription('voxel_env_v115_multitask8_v55', experiments=[EXPERIMENT_1AGENT])
+RUN_DESCRIPTION = RunDescription('megaverse_v115_multitask8_v55', experiments=[EXPERIMENT_1AGENT])