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 | class GeneticTraining:
"""
Class for the genetic training routine.
"""
def __init__(
self,
trainer: ContinuousTrainer,
simulation_runner_generator: callable,
n_episodes: int = 100,
episode_length: int = 20,
number_of_generations: int = 10,
population_size: int = 10,
number_of_parents: int = 2,
parent_selection_method: str = "sum",
output_directory: str = ".",
routine_name: str = "genetic_algorithm",
parallel_jobs: int = None,
cluster: JobQueueCluster = None,
):
"""
Constructor for the genetic training routine.
Parameters
----------
n_episodes : int
Number of episodes in each lifespan
number_of_generations : int (default: 10)
Number of generations to run
population_size : int (default: 10)
Number of individuals in the population
number_of_parents : int (default: 2)
Number of parents to select for each generation
parent_selection_method : str
How to reduce the life reward. Either sum or mean.
output_directory : str (default: ".")
Output directory of the run.
routine_name : str (default: "genetic_algorithm")
Name of the training routine.
cluster : JobQueueCluster
The cluster to run the jobs on.
If None, the jobs will be run locally.
parallel_jobs : int
The number of parallel jobs to run.
If None, the population size is used.
Notes
-----
Currently the client is fixed to run on the local machine. This will be
changed to a parameter in the future. The problem lies in espresso not being
able to handle multiple threads and us not being able to force Dask to refresh
a worker after each training is finished.
"""
self.trainer = trainer
self.simulation_runner_generator = simulation_runner_generator
self.n_episodes = n_episodes
self.episode_length = episode_length
self.number_of_generations = number_of_generations
self.population_size = population_size
self.number_of_parents = number_of_parents
self.output_directory = Path(f"{output_directory}/{routine_name}")
if parallel_jobs is None:
parallel_jobs = population_size
self.parallel_jobs = parallel_jobs
# Use default local cluster if None is given.
if cluster is None:
cluster = LocalCluster(
processes=True,
threads_per_worker=2,
silence_logs=logging.ERROR,
resources={"espresso": 1},
)
self.cluster = cluster
self.client = Client(cluster)
self.cluster.scale(n=self.parallel_jobs)
webbrowser.open(self.client.dashboard_link)
# Decide on parent splits
self.identifiers = range(population_size)
lazy_splits = np.array_split(np.ones(population_size), number_of_parents)
self.split_lengths = [len(split) for split in lazy_splits]
# set the select function
if parent_selection_method == "sum":
self._select_fn = onp.sum
elif parent_selection_method == "mean":
self._select_fn = onp.mean
elif parent_selection_method == "max":
self._select_fn = onp.max
# Create the output directory
os.mkdir(Path(self.output_directory))
@staticmethod
def _train_network(
name: Path,
load_directory: str = None,
trainer: ContinuousTrainer = None,
runner_generator: callable = None,
select_fn: callable = None,
episode_length: int = None,
n_episodes: int = None,
) -> tuple:
"""
Train the network.
Parameters
----------
name : Path
Name of the network and where to save the data.
load_directory : str (default: None)
Directory to load the model from. If None, a new model will be created.
trainer : ContinuousTrainer
Trainer to use for training.
runner_generator : callable
Function that returns a system_runner.
select_fn : callable
Function for reducing training rewards to a single
number.
episode_length : int
Length of one episode.
n_episodes : int
Number of episodes in the lifespan of the child.
Returns
-------
reduced_rewards : float
The reduced rewards of the agent.
model_id : str
The id of the model.
"""
model_id = name.as_posix().split("_")[-1]
os.makedirs(name)
os.chdir(name)
system_runner = runner_generator() # get the runner
if load_directory is None:
trainer.initialize_models()
else:
trainer.restore_models(load_directory)
rewards = trainer.perform_rl_training(
system_runner,
episode_length=episode_length,
n_episodes=n_episodes,
load_bar=False,
)
trainer.export_models()
return (select_fn(rewards), model_id)
def _deploy_jobs(
self,
child_names: List[Path],
load_paths: List[Path],
) -> List[float]:
"""
Function to send jobs to the cluster.
Parameters
----------
child_names : List[Path]
List of paths to save the models to.
load_paths : List[Path, None]
List of paths to load the models from.
Returns
-------
Returns the outcome of the job deployment.
"""
futures = []
for i in range(self.population_size // self.parallel_jobs):
block = self.client.map(
self._train_network,
child_names[i * self.parallel_jobs : (i + 1) * self.parallel_jobs],
load_paths[i * self.parallel_jobs : (i + 1) * self.parallel_jobs],
[deepcopy(self.trainer)] * self.parallel_jobs,
[self.simulation_runner_generator] * self.parallel_jobs,
[self._select_fn] * self.parallel_jobs,
[self.episode_length] * self.parallel_jobs,
[self.n_episodes] * self.parallel_jobs,
resources={"espresso": 1},
)
_ = wait(block)
futures += self.client.gather(block)
# Restart and wait for workers
_ = self.client.restart(wait_for_workers=False)
_ = self.client.wait_for_workers(self.parallel_jobs)
return futures
def _run_generation(
self, generation: int, seed: bool = False, parent_ids: list = None
) -> List:
"""
Run a generation of the training.
Parameters
----------
generation : int
The number of the generation to run.
seed : bool (default: False)
Whether to seed the generation or not.
parent_ids : list (default: None)
The ids of the parents to use for the generation. If None, it should
be seeded.
Returns
-------
generation_outputs : list
A list of rewards values form which parents will be chosen.
"""
# Create the children directories
children_names = [
(
self.output_directory / f"_generation_{generation}" / f"_child_{i}"
).resolve()
for i in self.identifiers
]
# deploy the jobs
if seed:
generation_outputs = self._deploy_jobs(
children_names, [None] * self.population_size
)
else:
# get load paths for each parent
load_paths = []
for i, index in enumerate(parent_ids):
load_paths += [
self.output_directory
/ f"_generation_{generation - 1}"
/ f"_child_{index}"
/ "Models"
] * self.split_lengths[i]
load_paths = [item.resolve().as_posix() for item in load_paths]
generation_outputs = self._deploy_jobs(children_names, load_paths)
return generation_outputs
def _select_parents(self, generation_outputs: np.ndarray) -> list:
"""
Select the parents for the next generation.
Parameters
----------
generation_outputs : np.ndarray (n_individuals, )
The outputs of the generation.
Returns
-------
ids : list
The ids of the parents.
chosen_reward: float
Reward of the chosen child.
"""
rewards = [item[0] for item in generation_outputs]
ids = [item[1] for item in generation_outputs]
# First get best parent
max_reward_index = np.argmax(np.array(rewards))
chosen_id = ids[max_reward_index]
# Pick mutations
if self.number_of_parents == 1:
return [chosen_id], rewards[max_reward_index]
else:
random_ids = onp.random.choice(
ids, size=self.number_of_parents - 1, replace=False
)
return [chosen_id] + list(random_ids), rewards[max_reward_index]
def train_model(self):
"""
Train the model.
"""
generation = 0
# Seed genetic process
seed_outputs = self._run_generation(generation=generation, seed=True)
parents, reward = self._select_parents(seed_outputs)
# Loop over generations
progress = Progress(
"Generation: {task.fields[generation]}",
BarColumn(),
"Best generation reward: {task.fields[best_reward]:.2f} ",
TimeRemainingColumn(),
)
with progress:
task = progress.add_task(
"Genetic training",
total=self.number_of_generations - 1,
generation=generation,
best_reward=reward,
)
for _ in range(self.number_of_generations - 1):
generation += 1 # Update the generation
generation_outputs = self._run_generation(
generation=generation, parent_ids=parents
)
parents, reward = self._select_parents(generation_outputs)
progress.update(
task,
advance=1,
generation=generation,
best_reward=reward,
)
best_model_path = (
self.output_directory / f"_generation_{generation}" / f"_child_{parents[0]}"
)
print(f"Best Model: {best_model_path.as_posix()}")
print(f"Best Reward: {reward:.2f}")
# Shutdown the cluster
self.cluster.close()
self.client.close()
return best_model_path
|