-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathutil.py
169 lines (137 loc) · 7.91 KB
/
util.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
import torch as th
from stable_baselines3 import PPO
from stable_baselines3.common import make_vec_env
from stable_baselines3.common.vec_env import DummyVecEnv
from stable_baselines3.common.evaluation import evaluate_policy
def check_optimal(env, model, n_eval_episodes, trunc=None):
model.policy.setup_optimal_mask(env)
wass_dist_main, wass_dist_partner, cnt = 0, 0, 0
for partner_idx in range(model.policy.num_partners):
for _ in range(n_eval_episodes):
obs = env.reset()
done, state = False, None
while not done:
thobs = th.Tensor([obs])
# compare wasserstein distance between main_logits and optimal mask
main_logits, partner_logits = model.policy.get_action_logits_from_obs(obs=thobs, partner_idx=partner_idx)
if trunc is not None: main_logits, partner_logits = main_logits[:, :trunc], partner_logits[:, :trunc]
main_logits, partner_logits = main_logits - main_logits.logsumexp(dim=-1, keepdim=True), partner_logits - partner_logits.logsumexp(dim=-1, keepdim=True)
main_prob, partner_prob = th.exp(main_logits), th.exp(partner_logits)
optimal_mask = model.policy.get_mask(thobs).float()
if not optimal_mask.bool().all():
if trunc is not None: optimal_mask = optimal_mask[:, :trunc]
optimal_prob = optimal_mask / optimal_mask.sum(dim=-1, keepdim=True)
#print(obs, main_prob, partner_prob, optimal_prob)
wass_dist_main += th.abs(main_prob - optimal_prob).sum().data
wass_dist_partner += th.abs(partner_prob - optimal_prob).sum().data
cnt += 1.0
action, _ = model.predict(observation=obs, partner_idx=partner_idx, deterministic=False)
obs, reward, done, _info = env.step(action)
print(wass_dist_main / cnt, wass_dist_partner / cnt)
def learn(model, model_name, model_path, timesteps, save, period, save_thresh=None):
# TRAINING
steps, save_period, loop_period = timesteps, 10*period, period
while True:
for partner_idx in range(model.policy.num_partners):
mean_reward, std_reward = evaluate_policy(model, model.get_env(), partner_idx=partner_idx, n_eval_episodes=200, deterministic=False)
print("#data steps %u\t partner %u\t mean_rew %.2f\t std_rew %.2f\t" % (timesteps - steps, partner_idx, mean_reward, std_reward))
if save_thresh is not None and mean_reward >= save_thresh:
model.save(model_path), print("saving...")
break
if steps == 0: break
cur_steps = min(steps, loop_period)
steps -= cur_steps
model.learn(total_timesteps=cur_steps, tb_log_name=model_name, reset_num_timesteps=False)
time_to_save = steps == 0 or steps // save_period != (steps+cur_steps) // save_period
if save and time_to_save: model.save(model_path), print("saving...")
def load_model(model_path, policy_class, policy_kwargs, env, hp, partners, testing, try_load=True):
load_successful = False
if try_load:
try:
model = PPO.load(model_path)#, policy_kwargs=policy_kwargs)
load_successful = True
print("Model loaded successfully")
except Exception as e:
print("Could not load model", e)
if not load_successful:
print("Create new model")
n_steps, batch_size, n_epochs, = hp['n_steps'], hp['batch_size'], hp['n_epochs']
model = PPO(policy_class, env, policy_kwargs=policy_kwargs, n_steps=n_steps, batch_size=batch_size, n_epochs=n_epochs, verbose=0, ent_coef=0.00, marginal_reg_coef=hp['mreg'])
for name, param in model.policy.named_parameters():
if param.requires_grad:
print(name, param.data.size())
vec_env = DummyVecEnv([lambda: env])
model.set_env(vec_env)
model.policy.set_partners(partners)
if testing:
model.policy.num_partners = 1 # only test 1 partner
model.marginal_reg_coef = 0
model.n_epochs = hp['n_epochs_testing']
model.n_steps = hp['n_steps_testing']
model._init_rollout_buffer()
return model
def adapt_task(load_model_fn, learn_model_fn, train_partners, test_partners, invert_train_partners, invert_test_partners, timesteps1, timesteps2, period):
if len(test_partners) > len(train_partners): test_partners = test_partners[:len(train_partners)]
# ADAPT TASK
# new partners
model = load_model_fn(test_partners, testing=False)
model.policy.do_init_weights(init_main=False, init_partner=True)
model.policy.set_freeze_main(freeze=True)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps1, save=False, period=period)
# invert task on new partners
model.policy.set_partners(invert_test_partners)
model.env.envs[0].set_invert(invert=True)
model.policy.do_init_weights(init_main=False, init_partner=False)
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=True)
learn_model_fn(model, timesteps2, save=False, period=period)
# use new main module on orig partners
print("#section AdaptTask")
model_orig = load_model_fn(invert_train_partners, testing=False)
model_orig.policy.overwrite_main(model.policy)
model_orig.env.envs[0].set_invert(invert=True) # invert task on orig partners
learn_model_fn(model_orig, 1, save=False, period=1)
def adapt_partner_baseline(load_model_fn, learn_model_fn, partners, timesteps, period, do_optimal=False):
model = load_model_fn(partners, testing=True)
print("#section AdaptPartnerBaseline")
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)
if do_optimal:
model = load_model_fn(partners, testing=True)
model.policy.setup_optimal_mask(model.env.envs[0])
print("#section AdaptPartnerBaselineOptimal")
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)
def adapt_partner_modular(load_model_fn, learn_model_fn, partners, timesteps, period, do_optimal=False):
model = load_model_fn(partners, testing=True)
print("#section AdaptPartner")
model.policy.do_init_weights(init_main=False, init_partner=True)
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)
if do_optimal:
model = load_model_fn(partners, testing=True)
model.policy.setup_optimal_mask(model.env.envs[0])
print("#section AdaptPartnerOptimal")
model.policy.do_init_weights(init_main=False, init_partner=True)
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)
def adapt_partner_scratch(load_model_fn, learn_model_fn, partners, timesteps, period, do_optimal=False):
model = load_model_fn(partners, testing=True, try_load=False)
print("#section AdaptPartnerFromScratch")
model.policy.do_init_weights(init_main=True, init_partner=True)
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)
if do_optimal:
model = load_model_fn(partners, testing=True, try_load=False)
model.policy.setup_optimal_mask(env=model.env.envs[0])
print("#section AdaptPartnerFromScratchOptimal")
model.policy.do_init_weights(init_main=True, init_partner=True)
model.policy.set_freeze_main(freeze=False)
model.policy.set_freeze_partner(freeze=False)
learn_model_fn(model, timesteps, save=False, period=period)