-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest.py
202 lines (179 loc) · 7.93 KB
/
test.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
import argparse
import torch
import numpy as np
import random
import wandb
from main_code.utils.torch_objects import device
from main_code.nets.pomo import PomoNetwork
from main_code.agents.policy_agent import PolicyAgent
from main_code.agents.mcts_agent.mcts_agent import MCTSAgent, MCTSBatchAgent
from main_code.agents.adaptive_policy_agent import AdaptivePolicyAgent
from main_code.agents.mcts_agent.mcts import MCTS
from main_code.utils.logging.logging import get_test_logger
from main_code.utils.config.config import Config
from main_code.testing.tsp_tester import TSPTester
def main():
pass
def parse_adaptlr_args():
parser = argparse.ArgumentParser()
parser.add_argument("--num_epochs", type=int, default=4)
parser.add_argument("--batch_size", type=int, default=16)
parser.add_argument("--state_space_size", type=int, default=32)
parser.add_argument("--lr_rate", type=float, default=1e-4)
parser.add_argument("--weight_decay", type=float, default=1e-6)
parser.add_argument("--lr_decay_epoch", type=int, default=1)
parser.add_argument("--lr_decay_gamma", type=float, default=1.0)
parser.add_argument("--noise_factor", type=float, default=0.0)
adaptlr_opts = parser.parse_known_args()[0]
return adaptlr_opts
# different argument parsers for different settings
def parse_mcts_args():
parser = argparse.ArgumentParser()
parser.add_argument("--c_puct", type=float, default=2.0)
parser.add_argument("--epsilon", type=float, default=0.6)
parser.add_argument("--weight_fac", type=float, default=50)
parser.add_argument("--node_value_scale", type=str, default="[0,1]")
parser.add_argument("--expansion_limit", type=int, default=None)
parser.add_argument("--node_value_term", type=str, default="smooth")
parser.add_argument("--prob_term", type=str, default="puct")
parser.add_argument("--num_playouts", type=int, default=10)
parser.add_argument("--num_parallel", type=int, default=8)
parser.add_argument("--virtual_loss", type=int, default=20)
mcts_opts = parser.parse_known_args()[0]
return mcts_opts
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--model_path",
type=str,
default="./results/saved_models/saved_tsp20_model",
# default="./results/saved_models/saved_tsp50_model",
)
# test hyperparmeters -> influence performance
parser.add_argument("--num_trajectories", type=int, default=1)
parser.add_argument("--use_pomo_aug", type=int, default=0)
parser.add_argument("--sampling_steps", type=int, default=1)
# whether to use agent with mcts planning
parser.add_argument("--use_mcts", type=int, default=0)
# use agent with adaptive learning
parser.add_argument("--use_adaptlr", type=int, default=0)
# batchsize only relevant for speed, depends on gpu memory
parser.add_argument("--test_batch_size", type=int, default=1024)
# random test set specifications
parser.add_argument(
"--random_test", dest="random_test", default=False, action="store_true"
)
parser.add_argument("--num_samples", type=int, default=10000)
parser.add_argument("--num_nodes", type=int, default=100)
parser.add_argument(
"--tsp_type", type=str, default="uniform"
) # later add clustered
parser.add_argument("--num_workers", type=int, default=4)
# saved test set
parser.add_argument("--test_set", type=str, default="fu_et_al_n_20_10000")
parser.add_argument("--test_type", type=str, default="test")
# save options
parser.add_argument("--save_dir", type=str, default="./results")
# wandb stuff
parser.add_argument("--job_type", type=str, default=None)
parser.add_argument("--experiment_name", type=str, default=None)
parser.add_argument("--wandb_mode", type=str, default=None)
opts = parser.parse_known_args()[0]
return opts
if __name__ == "__main__":
opts = parse_args()
# set seeds for reproducibility
np.random.seed(37)
random.seed(37)
torch.manual_seed(37)
# get config
config = Config(config_json=f"{opts.model_path}/config.json", restrictive=False)
# create new test subconfig
config.test = Config(config_class=opts, restrictive=False)
# config.test = config.test
# update config based on provided bash arguments
# check if test shall be random then skip next step
if config.test.random_test:
config.test.test_set_path = None
else:
# check whether test set exists if not throw error
config.test.test_set_path = f"./data/{opts.test_type}_sets/{opts.test_set}"
# extract some extra info based on the test set name
config.test.num_samples = int(opts.test_set.split("_")[-1])
config.test.num_nodes = int(opts.test_set.split("_")[-2])
# handle arbitrary test batch sizes for various sized inputs
if config.test.use_pomo_aug:
config.test.test_batch_size = max(8, config.test.test_batch_size)
else:
config.test.test_batch_size = max(
config.test.sampling_steps, config.test.test_batch_size
)
# adjust settings for mcts
if config.test.use_mcts:
config.test.test_batch_size = (
8 if config.test.use_pomo_aug else config.test.sampling_steps
)
# parse mcts arguments
mcts_opts = parse_mcts_args()
# prepare mcts_config and use as direct input for agent
config.test.mcts = Config()
# optional if defaults are set in class definiton instead of during parsing
config.test.mcts.set_defaults(MCTS)
# overwrite with parse values
config.test.mcts.from_class(mcts_opts)
if config.test.use_adaptlr:
adaptlr_opts = parse_adaptlr_args()
config.test.adaptlr = Config(config_class=adaptlr_opts, restrictive=False)
config.test.adaptlr.state_space_size = np.clip(
config.test.adaptlr.batch_size, 1, config.test.test_batch_size
)
config.test.adaptlr.batch_size = np.clip(
config.test.adaptlr.batch_size, 1, config.test.adaptlr.state_space_size
)
# Init logger
logger, result_folder_path = get_test_logger(config.test)
wandb.init(
config=config.to_dict(),
mode=config.test.wandb_mode,
group=config.test.experiment_name,
job_type=config.test.job_type,
)
# wandb cant handle sub configs by itself
config = Config(config_dict=wandb.config._items, restrictive=False)
# save config to log folder
config.to_yaml(f"{result_folder_path}/config.yml", nested=True)
# Load Model - could be done inside agent
nnetwork = PomoNetwork(config).to(device)
model_save_path = f"{opts.model_path}/ACTOR_state_dic.pt"
nnetwork.load_state_dict(torch.load(model_save_path, map_location=device))
# select the agent
if config.test.use_mcts:
agent = MCTSAgent(nnetwork, config.test.mcts.to_dict(False))
# agent = MCTSBatchAgent(nnetwork, c_puct=config.test.c_puct, n_playout=config.test.num_playouts, num_parallel=config.test.num_parallel)
elif config.test.use_adaptlr:
agent = AdaptivePolicyAgent(nnetwork, **config.test.adaptlr.to_dict(False))
else:
agent = PolicyAgent(nnetwork)
# log model info
fill_str = (
"=============================================================================="
)
logger.info(fill_str)
logger.info(fill_str)
logger.info(f" <<< MODEL: {model_save_path} >>>")
# init tester
tester = TSPTester(
logger,
num_trajectories=config.test.num_trajectories,
num_nodes=config.test.num_nodes,
num_samples=config.test.num_samples,
sampling_steps=config.test.sampling_steps,
use_pomo_aug=config.test.use_pomo_aug,
test_set_path=config.test.test_set_path,
test_batch_size=config.test.test_batch_size,
num_workers=config.test.num_workers,
)
# run test
test_result = tester.test(agent)
# save results
tester.save_results(file_path=f"{result_folder_path}/result.json")