File size: 14,496 Bytes
05c9ac2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 |
from unittest import mock
import pytest
from typing import List
import mlagents.trainers.tests.mock_brain as mb
import numpy as np
from mlagents.trainers.agent_processor import (
AgentProcessor,
AgentManager,
AgentManagerQueue,
)
from mlagents.trainers.action_info import ActionInfo
from mlagents.trainers.torch_entities.action_log_probs import LogProbsTuple
from mlagents.trainers.trajectory import Trajectory
from mlagents.trainers.stats import StatsReporter, StatsSummary
from mlagents.trainers.behavior_id_utils import get_global_agent_id
from mlagents_envs.side_channel.stats_side_channel import StatsAggregationMethod
from mlagents.trainers.tests.dummy_config import create_observation_specs_with_shapes
from mlagents_envs.base_env import ActionSpec, ActionTuple
def create_mock_policy():
mock_policy = mock.Mock()
mock_policy.reward_signals = {}
mock_policy.retrieve_previous_memories.return_value = np.zeros(
(1, 1), dtype=np.float32
)
mock_policy.retrieve_previous_action.return_value = np.zeros((1, 1), dtype=np.int32)
return mock_policy
def _create_action_info(num_agents: int, agent_ids: List[str]) -> ActionInfo:
fake_action_outputs = {
"action": ActionTuple(
continuous=np.array([[0.1]] * num_agents, dtype=np.float32)
),
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"log_probs": LogProbsTuple(
continuous=np.array([[0.1]] * num_agents, dtype=np.float32)
),
}
fake_action_info = ActionInfo(
action=ActionTuple(continuous=np.array([[0.1]] * num_agents, dtype=np.float32)),
env_action=ActionTuple(
continuous=np.array([[0.1]] * num_agents, dtype=np.float32)
),
outputs=fake_action_outputs,
agent_ids=agent_ids,
)
return fake_action_info
@pytest.mark.parametrize("num_vis_obs", [0, 1, 2], ids=["vec", "1 viz", "2 viz"])
def test_agentprocessor(num_vis_obs):
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=2,
observation_specs=create_observation_specs_with_shapes(
[(8,)] + num_vis_obs * [(84, 84, 3)]
),
action_spec=ActionSpec.create_continuous(2),
)
fake_action_info = _create_action_info(2, mock_decision_steps.agent_id)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
for _ in range(5):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Assert that two trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 2
# Assert that the trajectory is of length 5
trajectory = tqueue.put.call_args_list[0][0][0]
assert len(trajectory.steps) == 5
# Make sure ungrouped agents don't have team obs
for step in trajectory.steps:
assert len(step.group_status) == 0
# Assert that the AgentProcessor is empty
assert len(processor._experience_buffers[0]) == 0
# Test empty steps
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=0,
observation_specs=create_observation_specs_with_shapes(
[(8,)] + num_vis_obs * [(84, 84, 3)]
),
action_spec=ActionSpec.create_continuous(2),
)
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
# Assert that the AgentProcessor is still empty
assert len(processor._experience_buffers[0]) == 0
def test_group_statuses():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=4,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
grouped=True,
)
fake_action_info = _create_action_info(4, mock_decision_steps.agent_id)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
for _ in range(2):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Make terminal steps for some dead agents
_, mock_terminal_steps_2 = mb.create_mock_steps(
num_agents=2,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
done=True,
grouped=True,
agent_ids=[2, 3],
)
# Make decision steps continue for other agents
mock_decision_steps_2, _ = mb.create_mock_steps(
num_agents=2,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
done=False,
grouped=True,
agent_ids=[0, 1],
)
processor.add_experiences(
mock_decision_steps_2, mock_terminal_steps_2, 0, fake_action_info
)
# Continue to add for remaining live agents
fake_action_info = _create_action_info(4, mock_decision_steps_2.agent_id)
for _ in range(3):
processor.add_experiences(
mock_decision_steps_2, mock_terminal_steps, 0, fake_action_info
)
# Assert that four trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 4
# Get the first trajectory, which should have been agent 2 (one of the killed agents)
trajectory = tqueue.put.call_args_list[0][0][-1]
assert len(trajectory.steps) == 3
# Make sure trajectory has the right Groupmate Experiences.
# All three steps should contain all agents
for step in trajectory.steps:
assert len(step.group_status) == 3
# Last trajectory should be the longest. It should be that of agent 1, one of the surviving agents.
trajectory = tqueue.put.call_args_list[-1][0][-1]
assert len(trajectory.steps) == 5
# Make sure trajectory has the right Groupmate Experiences.
# THe first 3 steps should contain all of the obs (that 3rd step is also the terminal step of 2 of the agents)
for step in trajectory.steps[0:3]:
assert len(step.group_status) == 3
# After 2 agents has died, there should only be 1 group status.
for step in trajectory.steps[3:]:
assert len(step.group_status) == 1
def test_agent_deletion():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"log_probs": LogProbsTuple(continuous=np.array([[0.1]], dtype=np.float32)),
}
mock_decision_step, mock_terminal_step = mb.create_mock_steps(
num_agents=1,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
)
mock_done_decision_step, mock_done_terminal_step = mb.create_mock_steps(
num_agents=1,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
done=True,
)
fake_action_info = ActionInfo(
action=ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
env_action=ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
outputs=fake_action_outputs,
agent_ids=mock_decision_step.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_step, mock_terminal_step, 0, ActionInfo.empty()
)
# Run 3 trajectories, with different workers (to simulate different agents)
add_calls = []
remove_calls = []
for _ep in range(3):
for _ in range(5):
processor.add_experiences(
mock_decision_step, mock_terminal_step, _ep, fake_action_info
)
add_calls.append(
mock.call([get_global_agent_id(_ep, 0)], fake_action_outputs["action"])
)
processor.add_experiences(
mock_done_decision_step, mock_done_terminal_step, _ep, fake_action_info
)
# Make sure we don't add experiences from the prior agents after the done
remove_calls.append(mock.call([get_global_agent_id(_ep, 0)]))
policy.save_previous_action.assert_has_calls(add_calls)
policy.remove_previous_action.assert_has_calls(remove_calls)
# Check that there are no experiences left
assert len(processor._experience_buffers.keys()) == 0
assert len(processor._last_take_action_outputs.keys()) == 0
assert len(processor._episode_steps.keys()) == 0
assert len(processor._episode_rewards.keys()) == 0
assert len(processor._last_step_result.keys()) == 0
# check that steps with immediate dones don't add to dicts
processor.add_experiences(
mock_done_decision_step, mock_done_terminal_step, 0, ActionInfo.empty()
)
assert len(processor._experience_buffers.keys()) == 0
assert len(processor._last_take_action_outputs.keys()) == 0
assert len(processor._episode_steps.keys()) == 0
assert len(processor._episode_rewards.keys()) == 0
assert len(processor._last_step_result.keys()) == 0
def test_end_episode():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"log_probs": LogProbsTuple(continuous=np.array([[0.1]], dtype=np.float32)),
}
mock_decision_step, mock_terminal_step = mb.create_mock_steps(
num_agents=1,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
)
fake_action_info = ActionInfo(
action=ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
env_action=ActionTuple(continuous=np.array([[0.1]], dtype=np.float32)),
outputs=fake_action_outputs,
agent_ids=mock_decision_step.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_step, mock_terminal_step, 0, ActionInfo.empty()
)
# Run 3 trajectories, with different workers (to simulate different agents)
remove_calls = []
for _ep in range(3):
remove_calls.append(mock.call([get_global_agent_id(_ep, 0)]))
for _ in range(5):
processor.add_experiences(
mock_decision_step, mock_terminal_step, _ep, fake_action_info
)
# Make sure we don't add experiences from the prior agents after the done
# Call end episode
processor.end_episode()
# Check that we removed every agent
policy.remove_previous_action.assert_has_calls(remove_calls)
# Check that there are no experiences left
assert len(processor._experience_buffers.keys()) == 0
assert len(processor._last_take_action_outputs.keys()) == 0
assert len(processor._episode_steps.keys()) == 0
assert len(processor._episode_rewards.keys()) == 0
def test_agent_manager():
policy = create_mock_policy()
name_behavior_id = "test_brain_name"
manager = AgentManager(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
assert len(manager._trajectory_queues) == 1
assert isinstance(manager._trajectory_queues[0], AgentManagerQueue)
def test_agent_manager_queue():
queue = AgentManagerQueue(behavior_id="testbehavior")
trajectory = mock.Mock(spec=Trajectory)
assert queue.empty()
queue.put(trajectory)
assert not queue.empty()
queue_traj = queue.get_nowait()
assert isinstance(queue_traj, Trajectory)
assert queue.empty()
def test_agent_manager_stats():
policy = mock.Mock()
stats_reporter = StatsReporter("FakeCategory")
writer = mock.Mock()
stats_reporter.add_writer(writer)
manager = AgentManager(policy, "MyBehavior", stats_reporter)
all_env_stats = [
{
"averaged": [(1.0, StatsAggregationMethod.AVERAGE)],
"most_recent": [(2.0, StatsAggregationMethod.MOST_RECENT)],
"summed": [(3.1, StatsAggregationMethod.SUM)],
},
{
"averaged": [(3.0, StatsAggregationMethod.AVERAGE)],
"most_recent": [(4.0, StatsAggregationMethod.MOST_RECENT)],
"summed": [(1.1, StatsAggregationMethod.SUM)],
},
]
for env_stats in all_env_stats:
manager.record_environment_stats(env_stats, worker_id=0)
expected_stats = {
"averaged": StatsSummary(
full_dist=[1.0, 3.0], aggregation_method=StatsAggregationMethod.AVERAGE
),
"most_recent": StatsSummary(
full_dist=[4.0], aggregation_method=StatsAggregationMethod.MOST_RECENT
),
"summed": StatsSummary(
full_dist=[3.1, 1.1], aggregation_method=StatsAggregationMethod.SUM
),
}
stats_reporter.write_stats(123)
writer.write_stats.assert_any_call("FakeCategory", expected_stats, 123)
# clean up our Mock from the global list
StatsReporter.writers.remove(writer)
|