-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
346 lines (307 loc) · 15.4 KB
/
train.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
from stable_baselines.common.vec_env import SubprocVecEnv
from stable_baselines import PPO2
from millify import millify
import numpy as np
import sys
from importlib import import_module
import argparse
import os
import shutil
import yaml
from util.tb_logging import Logger
class AslaugTrainer:
def __init__(self):
# Define internal counters
self.counter = {"n_steps": 0, "model_idx": 0, "cl_idx": 0,
"info_idx": 0, "ADR_idx": 1, "ADR_lvl": 0,
"last_adaption": None, "last_adr_idx": 0,
"logger": None}
# Parse arguments
self.parse_args()
# Prepare directory structure
self.prepare_directory()
# Prepare curriculum learning
self.cl_list = self.prepare_curriculum_learning(self.args['cl'])
# Load parameters
with open("params.yaml") as f:
params_all = yaml.load(f)
self.learning_params = params_all["learning_params"]
self.env_params = params_all["environment_params"]
# Prepare learning function
lp = self.learning_params
if type(lp["learning_rate"]) in [list, tuple]:
lr_params = lp["learning_rate"]
lp["learning_rate"] = self.create_custom_lr(*lr_params)
if type(lp["cliprange"]) in [list, tuple]:
lr_params = lp["cliprange"]
lp["cliprange"] = self.create_custom_lr(*lr_params)
# Inizialize gyms as vector environments
aslaug_mod = import_module("envs." + self.model_name)
def create_gym(): return aslaug_mod.AslaugEnv(params=self.env_params)
env = SubprocVecEnv([create_gym for i in range(self.args['n_cpu'])])
g_env = create_gym()
# Obtain observation slicing for neural network adaption
obs_slicing = g_env.obs_slicing if hasattr(g_env,
"obs_slicing") else None
lidar_calib = np.array(g_env.get_lidar_calibration())
np.save("data/saved_models/{}/\
lidar_calib.npy".format(self.folder_name), lidar_calib)
# Prepare model, either new or proceeding training (pt)
if self.args['pt'] is None:
model = PPO2(self.policy, env, verbose=0,
tensorboard_log="data/tb_logs/\
{}".format(self.folder_name),
policy_kwargs={"obs_slicing": obs_slicing},
**self.learning_params)
else:
pfn, pep = self.args['pt'].split(":")
model_path = "data/saved_models/{}/aslaug_{}_{}.pkl\
".format(pfn, self.args['version'], pep)
tb_log_path = "data/tb_logs/{}".format(self.folder_name)
model = PPO2.load(model_path, env=env, verbose=0,
tensorboard_log=tb_log_path,
policy_kwargs={"obs_slicing": obs_slicing},
**self.learning_params)
self.model = model
def train(self):
# Print number of trainable weights
n_els = np.sum([x.shape.num_elements()*x.trainable
for x in self.model.get_parameter_list()])
print("Number of trainable weights: {}".format(n_els))
# Start learning
self.model.learn(total_timesteps=int(self.args['steps']),
callback=self.callback)
# Save model
self.model.save(self.dir_path + self.model_name + ".pkl")
def parse_args(self):
# Parse arguments
parser = argparse.ArgumentParser()
parser.add_argument("-s", "--steps",
help="Define # steps for learning.",
default=10e6)
parser.add_argument("-n", "--n_cpu",
help="Define # processes to use.",
default=16)
parser.add_argument("-v", "--version",
help="Set env version.",
default="v0")
parser.add_argument("-f", "--folder",
help="Name the folder to save models.",
default="None")
parser.add_argument("-p", "--policy",
help="Define policy to use (import path).",
default="policies.aslaug_policy_v0.AslaugPolicy")
parser.add_argument("-cp", "--check_point",
help="# steps in between model checkpoints.",
default=500e3)
parser.add_argument("-cl", "--curriculum_learning", action='append',
help="Enable curriculum learning. Example to \
adjust parameter reward.r1 from 1 to 5 in 3M \
steps, starting at 1M: -cl reward.r1:1:5:1e6:3e6.")
parser.add_argument("-pt", "--proceed_training",
help="Specify model from which training shall be \
proceeded. Format: folder_name:episode")
args = parser.parse_args()
self.args = {"n_cpu": int(float(args.n_cpu)), "version": args.version,
"steps": int(float(args.steps)),
"policy_arg": args.policy, "n_cp": args.check_point,
"folder_name": args.folder,
"cl": args.curriculum_learning,
"pt": args.proceed_training}
# Define model name
self.model_name = "aslaug_{}".format(self.args['version'])
# Load module for policy
policy_mod_name = ".".join(self.args['policy_arg'].split(".")[:-1])
self.policy_name = self.args['policy_arg'].split(".")[-1]
self.policy_mod = import_module(policy_mod_name)
self.policy = getattr(self.policy_mod, self.policy_name)
# Define folder name
if self.args['folder_name'] == "None":
self.folder_name = self.model_name
else:
self.folder_name = self.args['folder_name']
# Define directory to save files to
self.dir_path = "data/saved_models/{}/".format(self.folder_name)
# self.prepare_directory()
def prepare_directory(self):
if not os.path.exists(self.dir_path):
os.mkdir(self.dir_path)
elif len(os.listdir(self.dir_path)) == 0:
print("Directory exists, but is empty. Proceeding.")
else:
print("Attention, {} already exists.".format(self.folder_name))
resp = input("Move it [m], delete it [r] or cancel [c]: ")
if resp == 'c':
exit()
elif resp == 'r':
shutil.rmtree(self.dir_path)
os.mkdir(self.dir_path)
elif resp == 'm':
resp = input("Enter new folder name for \
{}: ".format(self.folder_name))
shutil.move(self.dir_path,
"data/saved_models/{}/".format(resp))
os.mkdir(self.dir_path)
else:
print("Can't understand your expression.")
exit()
# Copy model to directory
shutil.copy("envs/{}.py".format(self.model_name),
self.dir_path + self.model_name + ".py")
# Save command to file for reference
text_file = open(os.path.join(self.dir_path, 'cmd.txt'), "w")
text_file.write(" ".join(sys.argv))
text_file.close()
# Save learning params to file
params_file = "data/saved_models/{}/\
params.yaml".format(self.folder_name)
shutil.copy("params.yaml", params_file)
# Copy policy to models folder
shutil.copy(self.policy_mod.__file__,
"data/saved_models/{}/{}.py".format(self.folder_name,
self.policy_name))
def prepare_curriculum_learning(self, cl):
# Prepare curriculum learning
if cl is not None:
cl_list = []
for clstr in cl:
(cl_param, cl_start, cl_end,
cl_begin, cl_steps) = clstr.split(":")
cl_list.append({"param": cl_param, "start": float(cl_start),
"end": float(cl_end), "begin": float(cl_begin),
"steps": float(cl_steps)})
else:
cl_list = None
return cl_list
# Prepare custom learning rate function
def create_custom_lr(self, lr_max, lr_min, a, b):
m = (lr_max - lr_min) / (a - b)
c = lr_max - m * a
return lambda x: np.min([lr_max, np.max([lr_min, m * x + c])])
# Please excuse the horrible formatting
def callback(self, _locals, _globals):
n_cp_simple = 0
self.counter['n_steps'] += self.model.n_batch
if self.counter['logger'] is None:
ppo_id = 1
ppo_path = 'data/tb_logs/{}/PPO2_{}'.format(self.folder_name,
ppo_id+1)
while os.path.exists(ppo_path):
ppo_id += 1
ppo_path = 'data/tb_logs/{}/PPO2_{}'.format(self.folder_name,
ppo_id+1)
ppo_path = 'data/tb_logs/{}/PPO2_{}/addons'.format(
self.folder_name, ppo_id)
self.counter['logger'] = Logger(ppo_path)
if (self.counter['n_steps'] / float(self.args['n_cp'])
>= self.counter['model_idx']):
n_cp_simple = millify(
float(self.counter['model_idx']) * float(self.args['n_cp']),
precision=6)
suffix = "_{}.pkl".format(n_cp_simple)
cp_name = self.model_name + suffix
self.model.save(self.dir_path + cp_name)
self.counter['model_idx'] += 1
data = {"version": self.args['version'],
"model_path": self.dir_path + cp_name}
with open('latest.json', 'w') as outfile:
yaml.dump(data, outfile)
print("Stored model at episode {}.".format(n_cp_simple))
if (self.cl_list is not None and self.counter['n_steps'] / 25000.0
>= self.counter['cl_idx']):
self.counter['cl_idx'] += 1
self.perform_CL()
if self.counter['n_steps'] / 5000.0 >= self.counter['info_idx']:
self.counter['info_idx'] += 1
print("Current frame_rate: {} fps.".format(_locals["fps"]))
msr_avg = np.average(
self.model.env.env_method("get_success_rate"))
self.counter['logger'].log_scalar('metrics/success_rate', msr_avg,
self.counter['n_steps'])
if self.counter['n_steps'] / 25000.0 >= self.counter['ADR_idx'] \
and len(self.env_params['adr']['adaptions']) > 0:
self.counter['ADR_idx'] += 1
self.perform_ADR()
# Please excuse the horrible formatting
def perform_ADR(self):
avg = np.average(self.model.env.env_method("get_success_rate"))
print("Average success rate: {}".format(avg))
for level in range(len(self.env_params['adr']['adaptions'])):
for adaption in self.env_params['adr']['adaptions'][level]:
val = np.average(self.model.env.env_method(
"get_param", adaption['param']))
self.counter['logger'].log_scalar(
'ADR/{}/{}'.format(level, adaption['param']), val,
self.counter['n_steps'])
if avg >= self.env_params['adr']['success_threshold']:
to_adapt = []
if self.counter['ADR_lvl'] \
< len(self.env_params['adr']['adaptions']):
adaptions = self.env_params['adr']['adaptions']
for adapts in adaptions[self.counter['ADR_lvl']]:
val = np.average(self.model.env.env_method(
"get_param", adapts['param']))
if val != adapts['end']:
to_adapt.append(adapts)
if len(to_adapt) > 0:
rnd_idx = np.random.randint(len(to_adapt))
adapts = to_adapt[rnd_idx]
self.counter['last_adaption'] = adapts
val = np.average(self.model.env.env_method(
"get_param", adapts['param']))
dval = +(adapts['end']-adapts['start'])/adapts['steps']
val = max(min(adapts['end'], adapts['start']), min(
max(adapts['end'], adapts['start']), val + dval))
print("Setting {} to {}(+)".format(adapts['param'], val))
self.model.env.env_method("set_param", adapts['param'], val)
else:
self.counter['ADR_lvl'] = min(
self.counter['ADR_lvl'] + 1,
len(self.env_params['adr']['adaptions']))
if avg <= self.env_params['adr']['fail_threshold']:
if self.counter['last_adaption'] is not None:
adaptions = self.env_params['adr']['adaptions']
if self.counter['last_adaption'] \
not in adaptions[self.counter['ADR_lvl']]:
self.counter['ADR_lvl'] = max(
0, self.counter['ADR_lvl'] - 1)
else:
val = np.average(self.model.env.env_method(
"get_param", self.counter['last_adaption']['param']))
dval = (-(self.counter['last_adaption']['end']
- self.counter['last_adaption']['start'])
/ self.counter['last_adaption']['steps'])
val_h1 = max(self.counter['last_adaption']['end'],
self.counter['last_adaption']['start'])
val = max(min(self.counter['last_adaption']['end'],
self.counter['last_adaption']['start']),
min(val_h1, val + dval))
print("Setting {} to {}(-)\
".format(self.counter['last_adaption']['param'],
val))
self.model.env.env_method(
"set_param", self.counter['last_adaption']['param'],
val)
self.counter['last_adaption'] = None
# Please excuse the horrible formatting
def perform_CL(self):
for cl_entry in self.cl_list:
cl_val = (cl_entry["start"]
+ (cl_entry["end"]
- cl_entry["start"])
* (self.counter['n_steps']-cl_entry["begin"])
/ cl_entry["steps"])
cl_h1 = cl_entry["begin"] + cl_entry["steps"]
if cl_val >= min(cl_entry["start"], cl_entry["end"]) \
and cl_val <= max(cl_entry["start"], cl_entry["end"]) \
and self.counter['n_steps'] >= cl_entry["begin"] \
and self.counter['n_steps'] <= cl_h1:
self.model.env.env_method("set_param",
cl_entry["param"], cl_val)
print("Modifying param {} to {}".format(cl_entry["param"],
cl_val))
self.counter['logger'].log_scalar(
'CL/{}'.format(cl_entry['param']), cl_val,
self.counter['n_steps'])
trainer = AslaugTrainer()
trainer.train()