-
Notifications
You must be signed in to change notification settings - Fork 0
/
arguments.py
134 lines (114 loc) · 5.14 KB
/
arguments.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
import argparse
import json
import logging
import sys
from pathlib import Path
import gym_rarl.envs
all_envs = gym_rarl.envs.getList()
def parse_args(cmd_args=None):
logging.info('Parsing args')
if cmd_args is None:
cmd_args = sys.argv[1:]
# Parse name to know which config to do
parser = argparse.ArgumentParser()
parser.add_argument('--name', type=str, default=None, required=True)
parser.add_argument('--trainingconfig', type=str, default=str(Path.cwd() / 'trainingconfig.json'))
name_arguments, remaining_args = parser.parse_known_args(cmd_args)
# Load config file
config_arguments = get_config_arguments(name_arguments)
# Parse other commandline options
# Hyperparameters
parser.add_argument('--N_steps', type=int) # Number of steps in a rolloout, N_traj in Algorithm 1
parser.add_argument('--N_iter', type=int)
parser.add_argument('--N_mu', type=int)
parser.add_argument('--N_nu', type=int)
parser.add_argument('--N_eval_episodes', type=int)
parser.add_argument('--N_eval_timesteps', type=int)
parser.add_argument('--adv_force', type=float, default=None)
parser.add_argument('--mass_percentage', type=float, default=1.0)
parser.add_argument('--friction_percentage', type=float, default=1.0)
parser.add_argument('--seed', type=int)
# The name of the adversarial environment class
parser.add_argument("--env", type=str, required=True,
help=', '.join(all_envs))
parser.add_argument("--force-adv-name", type=str)
parser.add_argument('--save-every', type=int, default=None)
parser.add_argument('--monitor-dir', type=str, default=None)
parser.add_argument('--root', type=str, default=str(Path.cwd()))
# Flags
parser.add_argument('--evaluate', action='store_true')
parser.add_argument('--verbose', action='store_true')
parser.add_argument('--log', action='store_true')
parser.add_argument('--control', action='store_true')
parser.add_argument('--render', action='store_true')
parser.add_argument("--force-adversarial", action='store_true')
parser.add_argument("--force-no-adversarial", action='store_true')
parser.add_argument("--simple-reward", action='store_true')
arguments = parser.parse_args(cmd_args, namespace=config_arguments)
# Set extra variables
populate_derivatives(arguments)
logging.info(f'All arguments: {arguments}')
validate_arguments(arguments)
return arguments
def validate_arguments(arguments):
"""
List of sanity assertions for commandline and trainingconfig arguments
"""
if arguments.adv_force:
assert 0.0 <= arguments.adv_force
if arguments.N_steps:
assert arguments.N_steps % 2 == 0
assert not (arguments.force_adversarial and arguments.force_no_adversarial)
if arguments.save_every:
assert arguments.save_every % (arguments.N_mu * arguments.N_steps) == 0
def populate_derivatives(arguments):
"""
Add derivative arguments from the already parsed ones.
"""
import random
import numpy as np
import torch
random.seed(arguments.seed)
np.random.seed(arguments.seed)
torch.manual_seed(arguments.seed)
arguments.root = Path(arguments.root)
if arguments.monitor_dir is not None:
arguments.monitor_dir = str(arguments.root / arguments.monitor_dir)
arguments.pickle = arguments.root / f'models/{arguments.name}'
arguments.logs = arguments.root / f'logs/{arguments.name}'
# Are we running RARL or control
if arguments.control:
arguments.prot_name = f'control-{arguments.env}'
arguments.adversarial = arguments.force_adversarial
else:
arguments.prot_name = f'prot-{arguments.env}'
arguments.adversarial = not arguments.force_no_adversarial
arguments.adv_name = f'adv-{arguments.env}'
if arguments.force_adv_name is None:
arguments.adv_pickle = f'{arguments.pickle}-{arguments.adv_name}'
else:
arguments.adv_pickle = arguments.root / f'models/{arguments.force_adv_name}-{arguments.adv_name}'
arguments.envname = f'{arguments.prot_name}-env'
def get_config_arguments(existing_arguments):
"""
Parse config arguments given --name.
The config name must be a valid filepath.
It could be simple or compound, where --name="name_version" uses config "name"
and stores logs and models with prefix "name_version".
Returns a Namespace object with parameters from trainingconfig.
"""
all_configs = json.load(open(existing_arguments.trainingconfig))
assert not any('_' in config['name'] for config in all_configs)
assert not any(any(k == 'name' for k in c['params'].keys()) for c in all_configs)
configfile_arguments = argparse.Namespace()
config_name = existing_arguments.name
if '_' in existing_arguments.name:
fields = existing_arguments.name.split('_')
assert len(fields) == 2
config_name = fields[0]
my_config = next(c for c in all_configs if c['name'] == config_name)
params = my_config['params']
for k, v in params.items():
logging.info(f'config file set arguments[{k}] = {v}')
configfile_arguments.__setattr__(k, v)
return configfile_arguments