-
Notifications
You must be signed in to change notification settings - Fork 1
/
experiment_config.py
377 lines (329 loc) · 14.2 KB
/
experiment_config.py
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
import json
import os
import gym
import stable_baselines3
from stable_baselines3 import DQN, A2C
from stable_baselines3.common.base_class import BaseAlgorithm
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.callbacks import EvalCallback
from stable_baselines3.common.vec_env.dummy_vec_env import DummyVecEnv
from stable_baselines3.common.vec_env.vec_transpose import VecTransposeImage
from callback import InfoCallback
import gym_minigrid
from gym_minigrid.envs import RiskyPathEnv
from gym_minigrid.envs.risky import DEFAULT_REWARDS
from gym_minigrid.wrappers import ImgObsWrapper, RGBImgObsWrapper, TensorObsWrapper
from special_wrappers import IntrinsicMotivationWrapper, RandomizeGoalWrapper
class GridworldExperiment:
SEEDS = (763, 4744, 5672, 4267, 3377, 4356, 2689, 2819, 5224, 529)
def __init__(
self,
exp_id: str,
**kwargs
):
"""Initialize the experiment with environment-specific
information. Algorithm configuration is seperate from the object
initialization, as this leads to reusable environment configurations
when rerunning the same experiment with different algorithm parameters.
Args:
exp_id (str): the unique experiment identifier
max_steps (int): the number of time steps before episode end
slip_probability (float): agent slipping probability -> [0, 1)
wall_rebound (bool): activate rebound against wall
spiky_active (bool): whether spiky tiles should be active
reward_spec (dict): the dictionary containing the reward model
show_agent_dir (bool): only relevant for rgb_array observation
lava_positions (list[tuple]): specify the lava positions
agent_start_pos (tuple): the agent's starting position
goal_positions (list[tuple]): specify the goal positions
spiky_positions (list[tuple]): the spiky tile positions
"""
self.experiment_id = exp_id
# fix experiment not working when no intrinsic motivation wrapping
self.im_config = False
# set goal-tile placement randomization
self.goal_rnd = None
if 'goal_rnd' in kwargs:
self.goal_rnd = kwargs.pop('goal_rnd')
# environment spec
self.env_kwargs = kwargs
def add_dqn_config(
self,
config_name : str,
**kwargs
):
"""Configure the parameters that will be passed to the DQN algorithm.
The kwargs should correspond to the DQN arguments. The environment is
automatically added to the configuration.
"""
assert 'policy' not in kwargs, \
"Policy must be specified in run_experiment method"
self.dqn_kwargs = kwargs
self.dqn_config_name = config_name
def add_a2c_config(
self,
config_name : str,
**kwargs
):
"""Configure the parameters that will be passed to the a2c algorithm
"""
assert 'policy' not in kwargs, \
"Policy must be specified in run_experiment method"
self.a2c_kwargs = kwargs
self.a2c_config_name = config_name
def add_im_config(
self,
config_name : str,
**kwargs
):
"""Configure parameters passed for Intrinsic Motivation wrapper
"""
assert "total_steps" not in kwargs, \
"'total_steps' is automatically set during experiment execution"
self.im_config = True
self.im_config_name = config_name
self.im_kwargs = kwargs
def _check_existing_runs(self, log_path: str) -> list[int]:
"""Will return a list for seeds for which there are already logs in
the given path. List is empty if there are no logs with the given seed.
"""
res = []
if os.path.exists(log_path):
for seed in self.SEEDS:
for entry in os.listdir(log_path):
# use _ to make sure seed is not itself in another seed
if f"_{seed}_" in entry:
res.append(seed)
return res
def run_experiment(
self,
num_runs: int,
algo: str,
total_timesteps: int,
observation_type : str = "tensor",
policy_type : str = "MlpPolicy",
tile_size_px: int = None,
directional_agent: bool = False,
save_log: bool = True,
log_directory: str = "saved_logs/",
save_model: bool = True,
model_directory: str = "saved_models/",
force_cuda: bool = False,
callback: str = 'progress',
ignore_logs: bool = False
):
"""Run the experiment with preset algorithm configurations for a
certain number of runs. The path to the saved log and model are
automatically generated. Each individual run is saved in the folder
corresponding folder structure and is saved with the seed in the
filename for purposes of reproducibility.
Args:
num_runs (int): Number of runs, should be maximally 10.
algo (str): Either 'dqn' or 'a2c'
total_timesteps (int): Number of timesteps to learn for each run
observation_type (str): Either 'tensor' or 'rgb_array'
policy_type (str): Either 'MlpPolicy' or 'CnnPolicy'
tile_size_px (int, optional): Set tile render size for cnn policy
save_log (bool): Whether to save a tensorboard log
log_directory (str): The log folder (e.g. "saved_logs/")
save_model (bool): Whether to save the trained model
model_directory (str): The model folder ("saved_models/")
force_cuda (bool): Force execution with cuda (else error)
callback (str): #TODO
ignore_logs (bool): Will still execute already existing runs
"""
# basic consistency checks
assert num_runs <= len(self.SEEDS), \
"Cannot execute more runs than specified random seeds"
assert algo.lower() in ["a2c", "dqn"], \
"Must use one of the algorithms 'a2c' or 'dqn'"
assert observation_type.lower() in ["tensor", "rgb_array"], \
"Observations must conform to 'tensor' or 'rgb_array'"
assert policy_type in ["MlpPolicy", "CnnPolicy"], \
"policy_type must conform to types defined by stable baselines 3"
if policy_type == "CnnPolicy":
assert observation_type.lower() == "rgb_array", \
"Cnn only works on Images"
assert isinstance(tile_size_px, int), \
"Render size for tiles must be set when using rgb input to cnn"
try:
if algo.lower() == "dqn":
algo_config_name = self.dqn_config_name
elif algo.lower() == "a2c":
algo_config_name = self.a2c_config_name
except AttributeError as e:
# Attribute error raised when no algorithm configuration was passed
print(
"\nWARNING: Try setting algorithm parameters with the " \
+ "add_<a2c, dqn>_config() methods before running " \
+ "an experiment.\n"
)
raise
# update env kwargs for agent directionality
self.env_kwargs.update({"show_agent_dir" : directional_agent})
available_seeds = self.SEEDS
# define path for saving
if save_log:
obs_type_path = "tensor_obs/" if observation_type == "tensor" \
else f"pixel_obs_{tile_size_px}/"
full_path_suffix = \
self.experiment_id + "/" + obs_type_path + algo.lower() + \
"/" + algo_config_name + "/"
if self.im_config:
full_path_suffix += self.im_config_name + "/"
if not ignore_logs:
# check for are already existing logs for certain seeds
pth = log_directory + full_path_suffix
seeds_w_log = self._check_existing_runs(pth)
if len(seeds_w_log) > 0:
seed_enumeration = ", ".join([str(s) for s in seeds_w_log])
print(f"Skipping seeds {seed_enumeration}")
if len(seeds_w_log) == len(self.SEEDS):
print(
"WARNING: There are already runs for each seed" \
+ f" in directory {pth}."
)
# reduce list of seeds to be used
available_seeds = [x for x in self.SEEDS if x not in seeds_w_log]
# available_seeds = list(set(self.SEEDS).difference(seeds_w_log))
else:
full_path_suffix = ""
available_seeds = available_seeds[:num_runs]
for s in available_seeds:
# make environment
env = gym.make(
"MiniGrid-RiskyPath-v0",
**self.env_kwargs
)
eval_env = gym.make(
"MiniGrid-RiskyPath-v0",
**self.env_kwargs
)
# wrap environment according to observation type
if observation_type.lower() == "tensor":
env = TensorObsWrapper(env)
eval_env = TensorObsWrapper(eval_env)
elif observation_type.lower() == "rgb_array":
env = RGBImgObsWrapper(env, tile_size=tile_size_px)
eval_env = RGBImgObsWrapper(eval_env, tile_size=tile_size_px)
env = ImgObsWrapper(env)
eval_env = ImgObsWrapper(eval_env)
# apply same wrapping as base_algorithm for rgb inputs
# eval_env must be VecTransposeImage
eval_env = DummyVecEnv([lambda: eval_env])
eval_env = VecTransposeImage(eval_env)
# env = Monitor(env, info_keywords=("is_success",))
# NOTE EvalEnv only evaluates performance on the basic environment
if self.im_config:
env = IntrinsicMotivationWrapper(
env,
total_timesteps,
**self.im_kwargs
)
if self.goal_rnd is not None:
env = RandomizeGoalWrapper(
env,
randomization=self.goal_rnd
)
eval_env = RandomizeGoalWrapper(
eval_env,
eval_mode=True
)
# initialize model
try:
if algo.lower() == "dqn":
model = DQN(
policy_type,
env,
**self.dqn_kwargs,
tensorboard_log=log_directory \
if save_log else None,
seed=s
)
elif algo.lower() == "a2c":
model = A2C(
policy_type,
env,
**self.a2c_kwargs,
tensorboard_log=log_directory \
if save_log else None,
seed=s
)
print("Model seed:", model.seed)
except AttributeError as e:
print(
"\nWARNING: Try setting algorithm parameters with the " \
+ "add_<a2c, dqn>_config() methods before running " \
+ "an experiment.\n"
)
raise
# check model
print(f"Pytorch device: {model.device}")
if force_cuda and model.device.type != "cuda":
assert False, "Force Cuda execution but cuda not available"
# save paths
log_path = full_path_suffix + f"seed_{s}" \
+ ("_directional" if directional_agent else "")
save_path_model = model_directory \
+ full_path_suffix \
+ f"seed_{s}" \
+ ("_directional" if directional_agent else "")
# Initialize callback if provided
if callback == 'progress':
cb = InfoCallback(total_timesteps)
elif callback == 'save_best':
cb = [
EvalCallback(
eval_env=eval_env,
log_path=log_directory + log_path + "_evals",
best_model_save_path=save_path_model + "_best_model",
n_eval_episodes=6 # dependent on GoalRandomization
),
InfoCallback(total_timesteps)
]
else:
raise NotImplementedError()
print(f"Running experiment '{self.experiment_id}'")
print(f"Saving tensorboard logs to: {log_directory+log_path}")
# learn model, save if necessary
model.learn(
total_timesteps,
tb_log_name=log_path,
callback=cb
)
if save_model:
print(f"Saving trained model to: {save_path_model}")
model.save(save_path_model)
# explicitly remove model and environment
del model
del env
@classmethod
def load_env_json_config(
cls,
json_path: str,
keys: list[str] | str
):
"""Return an experiment object from a json configuration file.
Keys may be passed in a list, which leads to returning a list
of experiment objects.
Args:
json_path (str): the path to the json file
keys (list[str] | str): the experiment id(s) to be passed
"""
# Convenience checking for input
if isinstance(keys, str):
internal_keys = [keys]
else:
internal_keys = keys
exp_objects = []
# Initialize the experiment objects
with open(json_path, "r") as file:
deserialized = json.load(file)
for key in internal_keys:
new_exp = cls(key, **deserialized[key])
exp_objects.append(new_exp)
if len(exp_objects) > 1:
return exp_objects
elif len(exp_objects) == 1:
return exp_objects[0]
return None