-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathPhase3_lstm-gnn-ppo_simp.py
277 lines (249 loc) · 14.9 KB
/
Phase3_lstm-gnn-ppo_simp.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
import os
import argparse
import random
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
from modules.dqn.dqn_utils import seed_everything
from modules.sim.graph_factory import GetWorldSet
import modules.gnn.nfm_gen
from modules.gnn.construct_trainsets import ConstructTrainSet
from modules.ppo.ppo_custom import WriteTrainParamsToFile, WriteModelParamsToFile, GetConfigs
from modules.ppo.helpfuncs import CreateEnv, CreateEnvFS, evaluate_lstm_ppo
from modules.rl.rl_utils import GetFullCoverageSample
from modules.rl.rl_policy import ColllisionRiskAvoidancePolicy, LSTM_GNN_PPO_Single_Policy_simp, LSTM_GNN_PPO_Dual_Policy_simp
from modules.rl.rl_utils import EvaluatePolicy
from modules.sim.simdata_utils import SimulateAutomaticMode_PPO, SimulateInteractiveMode_PPO
from modules.ppo.ppo_wrappers import PPO_ActWrapper, PPO_ObsBasicDictWrapper
from modules.ppo.models_basic_lstm import PPO_GNN_Single_LSTM, PPO_GNN_Dual_LSTM
def get_last_checkpoint_filename(tp):
"""
Determine latest checkpoint iteration.
"""
if os.path.isdir(tp['base_checkpoint_path']):
max_checkpoint_iteration_list = [dirname.split('_')[-1].split('.')[0] for dirname in os.listdir(tp['base_checkpoint_path'])]
max_checkpoint_iteration_list = [int(n) for n in max_checkpoint_iteration_list if n.isdigit()]
if max_checkpoint_iteration_list == []:
return None, 0, -1e6
max_checkpoint_iteration = max(max_checkpoint_iteration_list)
fname = tp['base_checkpoint_path'] + 'model_' + str(max_checkpoint_iteration) + '.tar'
it = max_checkpoint_iteration
OF = open(tp['base_checkpoint_path'] + 'model_best_save_history.txt', 'r')
lines = OF.readlines()
reslist = [float(e.split('res:')[1].split('\n')[0])*.99 for e in lines]
best_result = max(reslist)
else:
return None, 0, -1e6
return fname, it, best_result
def main(args):
config, hp, tp = GetConfigs(args, suffix='simp')
##### TRAIN FUNCTION #####
if config['train']:
senv, env_all_train_list = ConstructTrainSet(config, apply_wrappers=True, type_obs_wrap=config['type_obs_wrap'], remove_paths=False, tset=config['train_on']) #TODO check
hp.node_dim = env_all_train_list[0].F
if config['demoruns']:
while True:
a = SimulateInteractiveMode_PPO(senv, filesave_with_time_suffix=False)
if a == 'Q': break
WriteTrainParamsToFile(config,hp,tp)
for seed in config['seedrange']:
seed_everything(seed)
senv.seed(seed)
logdir_=config['logdir']+'/SEED'+str(seed)
tp['writer'] = SummaryWriter(log_dir=f"{logdir_}/logs")
tp["base_checkpoint_path"]=f"{logdir_}/checkpoints/"
tp["seed_path"]=logdir_
if config['lstm_type'] in ['None', 'EMB', 'FE']:
model = PPO_GNN_Single_LSTM(config, hp, tp)
elif config['lstm_type'] in ['Dual','DualCC']:
model = PPO_GNN_Dual_LSTM(config, hp, tp)
if seed == config['seed0']: WriteModelParamsToFile(config, model)
last_checkpoint, it0, best_result = get_last_checkpoint_filename(tp)
if last_checkpoint is not None:
cp = torch.load(last_checkpoint)
model.load_state_dict(cp['weights'])
model.optimizer.load_state_dict(cp['optimizer'])
print(f"Loaded model from {last_checkpoint}")
print('Iteration:', it0, 'best_result:', best_result)
#try:
score = model.learn(senv, it0, best_result)
#except:
# pass#continue
##### EVALUATION FUNCTION FOR TRAINED POLICY for full test dataset #####
if config['eval']:
evalResults={}
evalName='Trainset'
senv, env_all_train_list = ConstructTrainSet(config, apply_wrappers=True, type_obs_wrap='BasicDict', remove_paths=False, tset=config['train_on']) #TODO check
evalResults[evalName]={'num_graphs.........':[],'num_graph_instances':[],'avg_return.........':[],'success_rate.......':[],}
for seed in config['seedrange']:
evalResults = GenerateResults(seed, env_all_train_list, evalName, evalResults, config, hp, tp, maxnodes=senv.max_possible_num_nodes)
SaveResults(evalResults, config, tp)
##### TEST FUNCTION FOR TRAINED POLICY #####
if config['test']:
evalResults={}
world_dict = SelectTestWorlds()
obs_evalmasks = 11*['prob_per_u_test']
obs_evalrates = [1.,.9,.8,.7,.6,.5,.4,.3,.2,.1,0.0] # [1.][0.8]
for obs_mask, obs_rate in zip(obs_evalmasks, obs_evalrates):
for world_name in world_dict.keys():
evalName=world_name+'_obs'+obs_mask+'_evaldet'+str(tp['eval_deterministic'])[0]
if obs_mask != 'None': evalName += str(obs_rate)
if world_name == "full_solvable_3x3subs":
evalenv = CreateEnvFS(config, obs_mask, obs_rate, max_nodes=world_dict[world_name][0], max_edges=world_dict[world_name][1])
else:
env = CreateEnv(world_name, max_nodes=world_dict[world_name][0], max_edges = world_dict[world_name][1], nfm_func_name = config['nfm_func'], var_targets=None, remove_world_pool=None, apply_wrappers=True, type_obs_wrap='BasicDict', obs_mask=obs_mask, obs_rate=obs_rate)
evalenv=[env]
evalResults[evalName]={'num_graphs.........':[],'num_graph_instances':[],'avg_return.........':[],'success_rate.......':[],}
for seed in config['seedrange']:
evalResults = GenerateResults(seed, evalenv, evalName, evalResults, config, hp, tp, world_dict[world_name][0])
SaveResults(evalResults, config, tp)
##### EVALUATION FUNCTION FOR HEURSTIC POLICY for selected dataset #####
if config['test_heur']:
l=config['logdir'].find(config['train_on'])
r=config['logdir'].find(config['nfm_func'])
to=len(config['train_on'])
config['logdir']=config['logdir'][:(l+to+1)]+config['logdir'][r:]
config['rootdir']=config['rootdir'][:l+to]
evalResults={}
world_dict = SelectTestWorlds()
obs_evalmasks = ['None']#2*['prob_per_u_test']
obs_evalrates = [1.0]#[0.1,0.0]#[1.,.9,.8,.7,.6,.5,.4,.3,.2,.1,0.]#[config['eval_rate']] #[1.,.3,.2,.1,0.0]
for obs_mask, obs_rate in zip(obs_evalmasks, obs_evalrates):
for world_name in world_dict.keys():
evalName=world_name+'_obs'+obs_mask
if obs_mask != 'None': evalName += str(obs_rate)
if world_name == "full_solvable_3x3subs":
evalenv = CreateEnvFS(config, obs_mask, obs_rate, max_nodes=world_dict[world_name][0], max_edges=world_dict[world_name][1])
else:
env = CreateEnv(world_name, max_nodes=world_dict[world_name][0], max_edges = world_dict[world_name][1], nfm_func_name = config['nfm_func'], var_targets=None, remove_world_pool=None, apply_wrappers=True, type_obs_wrap='BasicDict', obs_mask=obs_mask, obs_rate=obs_rate)
assert config['type_obs_wrap']=='BasicDict'
evalenv=[env]
evalResults[evalName]={'num_graphs.........':[],'num_graph_instances':[],'avg_return.........':[],'success_rate.......':[],}
evalResults = GenerateResultsHeur(evalenv, evalName, evalResults, config, hp, tp, world_dict[world_name][0])
SaveResults(evalResults, config, tp)
def SelectTestWorlds():
world_dict={ # [max_nodes,max_edges]
#'Manhattan3x3_PredictionExample':[9,9],
#'Manhattan5x5_DuplicateSetB':[25,300],
#'Manhattan3x3_WalkAround':[9,21],
#'MetroU3_e1t31_FixedEscapeInit':[33, 119],
'MemoryTaskU1':[8,16],
#'BifurGraphTask1':[27,52],
#'full_solvable_3x3subs':[9,21],
#'Manhattan5x5_FixedEscapeInit':[25,105],
#'Manhattan5x5_FixedEscapeInit2':[25,105],
#'Manhattan5x5_VariableEscapeInit':[25,105],
#'MetroU3_e17tborder_FixedEscapeInit':[33,119],
#'MetroU3_e17tborder_VariableEscapeInit':[33,119],
#'NWB_test_FixedEscapeInit':[975,1425],
#'NWB_test_FixedEscapeInit_U=15':[975,1425],
#'NWB_test_FixedEscapeInit_U=20':[975,1425],
#'NWB_test_FixedEscapeInit2':[975,1425],
#'NWB_test_VariableEscapeInit':[975,1425],
#'NWB_UTR_FixedEscapeInit':[1182,3204],
#'NWB_UTR_FixedEscapeInit_U=15':[1182,3204],
#'NWB_UTR_FixedEscapeInit_U=20':[1182,3204],
#'NWB_UTR_FixedEscapeInit2':[1182,3204],
#'NWB_UTR_VariableEscapeInit':[1182,3204],
#'NWB_ROT_FixedEscapeInit':[2602,7266],
#'NWB_ROT_FixedEscapeInit_U=15':[2602,7266],
#'NWB_ROT_FixedEscapeInit_U=20':[2602,7266],
#'NWB_ROT_FixedEscapeInit2':[2602,7266],
#'NWB_ROT_VariableEscapeInit':[2602,7266],
#'SparseManhattan5x5':[25,105],
}
return world_dict
def GenerateResults(seed, evalenv, evalName, evalResults, config, hp, tp, maxnodes):
logdir_=config['logdir']+'/SEED'+str(seed)
tp["base_checkpoint_path"]=f"{logdir_}/checkpoints/"
try:
assert os.path.exists(tp['base_checkpoint_path'])
except:
return evalResults
fname=tp['base_checkpoint_path']+'best_model.tar'
checkpoint = torch.load(fname)
if config['lstm_type'] in ['None', 'EMB','FE']:
ppo_model = PPO_GNN_Single_LSTM(config, hp, tp)
elif config['lstm_type'] in ['Dual','DualCC']:
ppo_model = PPO_GNN_Dual_LSTM(config, hp, tp)
ppo_model.load_state_dict(checkpoint['weights'])
print('Loaded model from', fname)
if config['lstm_type'] in ['EMB','FE','None']:
ppo_policy = LSTM_GNN_PPO_Single_Policy_simp(ppo_model, deterministic=tp['eval_deterministic'])
elif config['lstm_type'] in ['Dual','DualCC']:
ppo_policy = LSTM_GNN_PPO_Dual_Policy_simp(ppo_model, deterministic=tp['eval_deterministic'])
else: assert False
if config['demoruns']:
while True:
demoenv=random.choice(evalenv)
a = SimulateAutomaticMode_PPO(demoenv, ppo_policy, t_suffix=False, entries=None)
if a == 'Q': break
result = evaluate_lstm_ppo(logdir=logdir_, config=config, env=evalenv, ppo_policy=ppo_policy, eval_subdir=evalName, max_num_nodes=maxnodes)
num_unique_graphs, num_graph_instances, avg_return, success_rate = result
evalResults[evalName]['num_graphs.........'].append(num_unique_graphs)
evalResults[evalName]['num_graph_instances'].append(num_graph_instances)
evalResults[evalName]['avg_return.........'].append(avg_return)
evalResults[evalName]['success_rate.......'].append(success_rate)
return evalResults
def GenerateResultsHeur(evalenv, evalName, evalResults, config, hp, tp, maxnodes):
logdir_=config['logdir']
ppo_policy = ColllisionRiskAvoidancePolicy(evalenv[0])
if config['demoruns']:
while True:
demoenv=random.choice(evalenv)
a = SimulateAutomaticMode_PPO(demoenv, ppo_policy, t_suffix=False, entries=None)
if a == 'Q': break
result = evaluate_lstm_ppo(logdir=logdir_, config=config, env=evalenv, ppo_policy=ppo_policy, eval_subdir=evalName, max_num_nodes=maxnodes)
num_unique_graphs, num_graph_instances, avg_return, success_rate = result
evalResults[evalName]['num_graphs.........'].append(num_unique_graphs)
evalResults[evalName]['num_graph_instances'].append(num_graph_instances)
evalResults[evalName]['avg_return.........'].append(avg_return)
evalResults[evalName]['success_rate.......'].append(success_rate)
return evalResults
def SaveResults(evalResults, config, tp):
for ename, results in evalResults.items():
OF = open(config['logdir']+'/Eval_det'+str(tp['eval_deterministic'])[0]+'_Results_over_seeds_'+ename+'.txt', 'w')
def printing(text):
print(text)
OF.write(text + "\n")
np.set_printoptions(formatter={'float':"{0:0.3f}".format})
printing('Results over seeds for evaluation on '+ename+'\n')
for category,values in results.items():
printing(category)
printing(' avg over seeds: '+str(np.mean(values)))
printing(' std over seeds: '+str(np.std(values)))
printing(' per seed: '+str(np.array(values))+'\n')
if __name__ == '__main__':
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--train_on', default='None', type=str)
parser.add_argument('--num_step', default=600, type=int)
parser.add_argument('--batch_size', default=2, type=int)
parser.add_argument('--lr', default=5e-4, type=float)
parser.add_argument('--recurrent_seq_len', default=2, type=int)
parser.add_argument('--parallel_rollouts', default=4, type=int)
parser.add_argument('--rollout_steps', default=100, type=int)
parser.add_argument('--patience', default=500, type=int)
parser.add_argument('--checkpoint_frequency', default=5000, type=int)
parser.add_argument('--obs_mask', default='None', type=str, help='U observation masking type', choices=['None','freq','prob','prob_per_u','mix'])
parser.add_argument('--obs_rate', default=1.0, type=float)
parser.add_argument('--eval_rate', default=1.0, type=float)
parser.add_argument('--emb_dim', default=24, type=int)
parser.add_argument('--lstm_type', default='None', type=str, choices=['None','EMB','FE','Dual','DualCC'])
parser.add_argument('--lstm_hdim', default=24, type=int)
parser.add_argument('--lstm_layers', default=1, type=int)
#parser.add_argument('--lstm_dropout', default=0.0, type=float)
parser.add_argument('--emb_iterT', default=5, type=int)
parser.add_argument('--nfm_func', default='NFM_ev_ec_t_dt_at_um_us', type=str)
parser.add_argument('--qnet', default='gat2', type=str)
parser.add_argument('--critic', default='q', type=str, choices=['q','v']) # q=v value route, v=single value route
parser.add_argument('--train', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=False)
parser.add_argument('--eval', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=False)
parser.add_argument('--test', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=False)
parser.add_argument('--test_heur', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=False)
parser.add_argument('--num_seeds', default=1, type=int)
parser.add_argument('--seed0', default=0, type=int)
parser.add_argument('--demoruns', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=False)
parser.add_argument('--eval_deter', type=lambda s: s.lower() in ['true', 't', 'yes', '1'],default=True)
parser.add_argument('--type_obs_wrap', default='BasicDict', type=str)
args=parser.parse_args()
main(args)