-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
288 lines (239 loc) · 8.77 KB
/
eval.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
import os.path as osp
import numpy as np
import torch
import argparse
import json
from paper_environments import get_env
from utils import rollout
import plots
VIDEOS_DIR = 'videos'
# Numpy print options
np.set_printoptions(precision=3, suppress=True)
# Script options
options_choices = [
('e', 'evaluate', "Evaluate the policy"),
('p', 'plot', "Plot the expected return from the learning process"),
('pi', 'plot_info', "Plot relevant information from the learning process"),
('er', 'eval_repeat', "Evaluate repeteadly the environment!"),
]
# Script parameters
parser = argparse.ArgumentParser(
description='Evaluate a policy from a log directory.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
exclusive_opts = parser.add_mutually_exclusive_group()
for (short_opt, long_opt, description) in options_choices:
exclusive_opts.add_argument('-'+short_opt, '--'+long_opt, help=description,
action='store_const', dest='script_option',
const=short_opt)
parser.add_argument('log_dir', type=str, help='Full path of the log directory')
parser.add_argument('--seed', '-s', type=int, default=610, help='Seed value')
parser.add_argument('--pol_task', '-pt', type=int, default=None,
help="Policy task number. None is the Main Task")
parser.add_argument('--env_task', '-et', type=int, default=None,
help="Environment task number. None is the Main Task")
parser.add_argument('--horizon', '-n', type=int, default=None,
help="Rollout horizon. None used the max_horizon parameter"
"value from the log directory.")
parser.add_argument('--iteration', '-i', type=int, default=-1,
help="Model iteration. -1 for the last available episode")
parser.add_argument('--gpu', type=int, default=-1,
help='GPU ID. For using the cpu selects -1')
parser.add_argument('--stochastic', action='store_true')
# By default the script evaluates the last policy
parser.set_defaults(script_option='e')
n_rollouts = 0
environment = None
def plot_progress(progress_file, algo_name='hiusac', only_expect_return=True):
"""Plot relevant data from the training process logged in a file.
Args:
progress_file (str): Full path of the progress file.
algo_name (str): Algorithm name.
Returns:
None
"""
if algo_name in ['hiusac', 'hiusac-p']:
num_intentions = 2
else:
num_intentions = None
plots.plot_intentions_eval_returns(
progress_file,
num_intentions=num_intentions,
)
if not only_expect_return:
plots.plot_intentions_info(
progress_file,
num_intentions=num_intentions,
)
def eval_policy(env, policy, max_horizon=50, task=None, stochastic=False):
"""Evaluate a policy in a specific environment.
Args:
env (Env): Environment
policy (torch.nn.Module): Policy
max_horizon (int): Maximum horizon
task (int or None): Policy subtask
stochastic (bool): Select actions by sampling from the policy
Returns:
None
"""
rollout_info = rollout(
env,
policy,
max_horizon=max_horizon,
fixed_horizon=False,
device='cpu',
render=True,
return_info_dict=True,
record_video_name=None,
intention=task,
deterministic=not stochastic,
)
if task is None:
rollout_return = sum(rollout_info['reward'])
else:
rollout_return = sum([info[task]
for info in rollout_info['reward_vector']])
print("The rollout return is: %f" % rollout_return)
def record_policy(env, policy, max_horizon=50, task=None, stochastic=False,
q_fcn=None, video_name='rollout_video', video_dir=None):
if video_dir is None:
video_dir = VIDEOS_DIR
video_name = osp.join(
video_dir,
video_name
)
if not video_name.endswith('.mp4'):
video_name += '.mp4'
rollout(
env, policy,
max_horizon=max_horizon,
fixed_horizon=False,
device='cpu',
render=True,
intention=task, deterministic=not stochastic,
return_info_dict=False,
q_fcn=q_fcn,
record_video_name=video_name,
)
def plot_value_fcn(qf, policy, obs, action_lows, action_highs, actions_dims=(0, 1)):
plots.plot_q_values(
qf,
action_lower=action_lows,
action_higher=action_highs,
obs=obs,
policy=policy,
action_dims=actions_dims,
delta=0.05,
device='cpu'
)
if __name__ == '__main__':
# Parse and print out parameters
args = parser.parse_args()
# the full path of the log directory
log_dir = args.log_dir
# Get experiment variant data from log directory
with open(osp.join(log_dir, 'variant.json')) as json_data:
log_data = json.load(json_data)
env_name = log_data['env_name']
log_env_params = log_data['env_params']
algo_params = log_data['algo_params']
seed = algo_params['seed']
horizon = algo_params['max_horizon']
progress_file = None
env = None
policy = None
env_params = None
run_once = True
models_dir = None
qf = None
# Get models directory
itr_dir = 'itr_%03d' % args.iteration if args.iteration > -1 else 'last_itr'
models_dir = osp.join(log_dir, 'models', itr_dir)
# Get the progress file
if args.script_option in ['p', 'pi']:
if progress_file is None:
progress_file = osp.join(log_dir, 'progress.csv')
# Get environment and policy
if args.script_option in ['e', 'er', 'v']:
env, env_params = get_env(
env_name, args.pol_task, args.seed, render=True,
new_env_params=log_env_params
)
pol_file = osp.join(models_dir, 'policy.pt')
policy = torch.load(pol_file, map_location=lambda storage, loc: storage)
if args.horizon is not None:
horizon = args.horizon
# Get Q-value function
if args.script_option in ['v']:
qf_file = osp.join(models_dir, 'qf1.pt')
qf = torch.load(qf_file, map_location=lambda storage, loc: storage)
# Run the script with the selected option
if args.script_option == 'e':
eval_policy(env, policy,
max_horizon=horizon,
task=args.pol_task,
stochastic=args.stochastic,
)
elif args.script_option == 'er':
try:
while True:
eval_policy(env, policy,
max_horizon=horizon,
task=args.pol_task,
stochastic=args.stochastic,
)
except KeyboardInterrupt:
pass
elif args.script_option == 'p':
plot_progress(progress_file, log_data['algo_name'],
only_expect_return=True)
elif args.script_option == 'pi':
plot_progress(progress_file, log_data['algo_name'],
only_expect_return=False)
elif args.script_option in ['r', 're']:
max_iter = 300
max_rollouts = 10
# range_list = list(range(0, max_iter, 25)) + [None]
range_list = [args.iteration]
env_subtask = None if args.env_task == -1 else args.env_task
env.set_active_subtask(env_subtask)
for rr in range(max_rollouts):
if args.horizon is not None:
horizon = args.horizon
if env_subtask is None:
env_subtask = -1
if args.task is None:
subtask = -1
else:
subtask = args.task
video_name = (
itr_dir +
('_s%03d' % seed) +
('_task%01d' % subtask) +
('_envtask%01d' % env_subtask) +
('_rollout%02d' % rr)
)
video_name = osp.join(
env_name,
video_name
)
record_policy(env, policy,
max_horizon=horizon,
task=subtask,
stochastic=args.stochastic,
q_fcn=qf,
video_name=video_name,
)
n_rollouts += 1
if args.option.lower() == 're':
break
elif args.script_option == 'v':
obs = np.zeros(env.obs_dim)
# TODO: Make this for all envs
obs[0] = -6
obs[1] = -6
plot_value_fcn(qf, policy, obs, env.action_space.low, env.action_space.high)
if args.script_option in ['p', 'pi']:
input('Press a key to close the script')
else:
print("Closing the script. Bye!")