forked from IDSIA/policyevaluator
-
Notifications
You must be signed in to change notification settings - Fork 0
/
pssvf.py
373 lines (290 loc) · 13.1 KB
/
pssvf.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
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
import torch
import numpy as np
import gym
import core
import torch.nn as nn
import torch.optim as optim
from torch.distributions.normal import Normal
import time
import pickle
import argparse
parser = argparse.ArgumentParser()
parser.add_argument(
"--env_name",
default="Swimmer-v3",
choices=[
"Swimmer-v3",
"Hopper-v3",
"Ant-v3",
"Walker2d-v3",
"InvertedDoublePendulum-v2",
"HalfCheetah-v3"
],
type=str,
required=False,
)
parser.add_argument("--verbose", default=0, type=int, required=False)
parser.add_argument("--show_plots", default=0, type=int, required=False)
parser.add_argument("--use_gpu", default=1, type=int, required=False)
parser.add_argument("--seed", default=1234, type=int, required=False)
args = parser.parse_args()
verbose = args.verbose
show_plots = args.show_plots
# Default hyperparameters
config = dict(
env_name='Hopper-v3',#'Swimmer-v3', # MountainCarContinuous-v0
neurons_policy=(256,256),
neurons_vf=(256,256),
policy_iters=5,
vf_iters=5,
batch_size=16,
learning_rate_policy=2e-6,
learning_rate_vf=5e-3,
noise_policy=0.05, # std of distribution generating the noise for the perturbed policy
size_buffer=10000,
max_episodes=1000000000,
max_timesteps=3000000,
seed=1,
survival_bonus=False,
deterministic_actor=True,
ts_evaluation=10000,
n_probing_states=200,
print_stats=True,
render_prob_states=False,
save_model=100000000,
act_clipping=False,
activation_policy='tanh',
activation_vf='relu',
start_steps=0,
observation_normalization=True,
algo='pssvf',
act_noise=0.0,
use_virtual_class=True,
update_every_ts=False,
update_every=1000,
render=False,
weighted_sampling=True,
scale=1.1,
)
# Use GPU or CPU
if args.use_gpu:
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
else:
device = torch.device("cpu")
torch.manual_seed(config["seed"])
np.random.seed(config["seed"])
# Create env
env = gym.make(config["env_name"])
env_test = gym.make(config["env_name"])
torch.manual_seed(config['seed'])
np.random.seed(config['seed'])
if config['env_name'] == 'InvertedDoublePendulum-v2':
config.update({'max_timesteps': 300000}, allow_val_change=True)
config.update({'ts_evaluation': 1000}, allow_val_change=True)
if config['env_name'] in ['MountainCarContinuous-v0', 'InvertedPendulum-v2', 'Reacher-v2']:
config.update({'ts_evaluation': 1000,
'max_timesteps': 100000,
}, allow_val_change=True)
activation_policy = nn.ReLU if config['activation_policy'] == 'relu' else nn.Tanh
activation_vf = nn.ReLU if config['activation_vf'] == 'relu' else nn.Tanh
# Create replay buffer, policy, vf
buffer = core.Buffer(config['size_buffer'], scale=config['scale'])
statistics = core.Statistics(env.observation_space.shape)
ac = core.MLPActorCritic(config['algo'], env.observation_space, env.action_space, config['n_probing_states'],
hidden_sizes_actor=tuple(config['neurons_policy']), activation_policy=activation_policy,
activation_vf=activation_vf,
hidden_sizes_critic=tuple(config['neurons_vf']), device=device,
critic=True, deterministic_actor=config['deterministic_actor'],
act_clipping=config['act_clipping'], act_noise=config['act_noise'])
virtual_mlp = core.VirtualMLPPolicy(layer_sizes=[env.observation_space.shape[0]]+
list(tuple(config['neurons_policy']))+
[env.action_space.shape[0]],
act_lim=env.action_space.high[0],
nonlinearity=config['activation_policy'])
print("Number of policy params:", len(nn.utils.parameters_to_vector(list(ac.pi.parameters()))))
print("Number of value function params:", len(nn.utils.parameters_to_vector(list(ac.v.parameters()))))
print("Obs dim", env.observation_space.shape[0])
print("Act dim", env.action_space.shape[0])
# Setup optimizer
q_params = ac.v.parameters()
optimize_policy = optim.Adam(ac.pi.parameters(), lr=config['learning_rate_policy'])
optimize_vf = optim.Adam(q_params, lr=config['learning_rate_vf'])
def compute_policy_loss():
params = nn.utils.parameters_to_vector(list(ac.pi.parameters())).to(device, non_blocking=True).unsqueeze(0)
return -ac.v.forward(params, use_virtual_module=True, virtual_module=virtual_mlp)
def compute_vf_loss(progs, rew):
q = ac.v(progs, use_virtual_module=True, virtual_module=virtual_mlp)
loss_q = ((q - rew) ** 2).mean()
return loss_q
def perturbe_policy(policy):
dist = Normal(torch.zeros(len(torch.nn.utils.parameters_to_vector(policy.parameters()))), scale=1)
delta = dist.sample().to(device=device, non_blocking=True).detach()
# Perturbe policy parameters
params = torch.nn.utils.parameters_to_vector(policy.parameters()).detach()
perturbed_params = params + config['noise_policy'] * delta
# Copy perturbed parameters into a new policy
perturbed_policy = core.MLPActorCritic(config['algo'], env.observation_space, env.action_space,
config['n_probing_states'],
hidden_sizes_actor=tuple(config['neurons_policy']), activation_policy=activation_policy,
activation_vf=activation_vf,
hidden_sizes_critic=tuple(config['neurons_vf']), device=device,
critic=False, deterministic_actor=config['deterministic_actor'],
act_clipping=config['act_clipping'], act_noise=config['act_noise'])
torch.nn.utils.vector_to_parameters(perturbed_params, perturbed_policy.parameters())
return perturbed_policy
def update():
start_time = time.perf_counter()
for _ in range(config['vf_iters']):
# Sample batch
hist = buffer.sample_replay(config['batch_size'], weighted_sampling=config['weighted_sampling'])
prog, rew, _ = zip(*hist)
if config['use_virtual_class']:
prog = torch.stack(prog).to(device)
rew = torch.from_numpy(np.asarray(rew)).float().to(device=device, non_blocking=True).detach()
optimize_vf.zero_grad()
loss_vf = compute_vf_loss(prog, rew)
loss_vf.backward()
optimize_vf.step()
statistics.up_v_time += time.perf_counter() - start_time
start_time = time.perf_counter()
# # Freeze PSSVF
for p in q_params:
p.requires_grad = False
# Update policy
for _ in range(config['policy_iters']):
optimize_policy.zero_grad()
loss_policy = compute_policy_loss()
loss_policy.backward()
optimize_policy.step()
# # Unfreeze PSSVF
for p in q_params:
p.requires_grad = True
statistics.up_policy_time += time.perf_counter() - start_time
log_dict = {'loss_pvf': loss_vf.item(),
'loss_policy': loss_policy.item()}
if verbose:
print(log_dict)
return
def evaluate(policy):
rew_evals = []
with torch.no_grad():
for _ in range(10):
# Simulate a trajectory and compute the total reward
done = False
obs = env_test.reset()
rew_eval = 0
while not done:
obs = torch.as_tensor(obs, dtype=torch.float32)
if config['observation_normalization'] and statistics.episode > 0:
obs = statistics.normalize(obs)
with torch.no_grad():
action = policy.act(obs.to(device, non_blocking=True).detach())
if config['render']:
env_test.render()
#print(action)
obs_new, r, done, _ = env_test.step(action)
# Remove survival bonus
rew_eval += r
obs = obs_new
rew_evals.append(rew_eval)
statistics.rew_eval = np.mean(rew_evals)
statistics.push_rew(np.mean(rew_evals))
# Log results
log_dict = {'rew_eval': statistics.rew_eval,
'average_reward': np.mean(statistics.rewards),
'average_last_rewards': np.mean(statistics.last_rewards),
}
print(log_dict)
if config['print_stats']:
print("Ts", statistics.total_ts, "Ep", statistics.episode, "rew_eval", statistics.rew_eval)
print("time_sim", statistics.sim_time, "time_up_pi", statistics.up_policy_time, "time_up_v", statistics.up_v_time,
"total_time", statistics.total_time)
return
def simulate(policy):
# Simulate a trajectory and compute the total reward
done = False
obs = env.reset()
rew = 0
rew_bonus = 0
while not done:
obs = torch.as_tensor(obs, dtype=torch.float32)
if config['observation_normalization']:
statistics.push_obs(obs)
if statistics.episode > 0:
obs = statistics.normalize(obs)
with torch.no_grad():
action = policy.act(obs.to(device, non_blocking=True).detach())
obs_new, r, done, _ = env.step(action)
# Remove survival bonus
if not config['survival_bonus']:
if config['env_name'] == 'Hopper-v3' or config['env_name'] == 'Ant-v3' or config['env_name'] == 'Walker2d-v3':
rew += r - 1
elif config['env_name'] == 'Humanoid-v3':
rew += r - 5
else:
rew += r
else:
rew += r
rew_bonus += r
statistics.total_ts += 1
if statistics.total_ts % config['save_model'] == 0:
if config['neurons_policy'] == []:
torch.save(ac.v, 'models/model_lin' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts))
torch.save(ac.pi, 'policies/pi_lin' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts))
with open('statistics/stat_lin' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts), 'wb') as fp:
pickle.dump(statistics, fp)
else:
torch.save(ac.v, 'models/model' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts))
torch.save(ac.pi, 'policies/pi' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts))
with open('statistics/stat' + str(int(100*config['n_probing_states'])) + str(int(100*config['seed'])) + config['env_name'] + "_" + str(statistics.total_ts), 'wb') as fp:
pickle.dump(statistics, fp)
# Evaluate current policy
if statistics.total_ts % config['ts_evaluation'] == 0:
evaluate(ac)
# Update
if statistics.total_ts > config['start_steps'] and config['update_every_ts'] and statistics.episode > 0:
if statistics.total_ts % config['update_every'] == 0:
update()
if statistics.total_ts == 1000000:
log_dict = {'rew_eval_1M': statistics.rew_eval,
'average_reward_1M': np.mean(statistics.rewards),
'average_last_rewards_1M': np.mean(statistics.last_rewards)}
print(log_dict)
obs = obs_new
return rew, rew_bonus
def train():
start_time = time.perf_counter()
# Collect data with perturbed policy
perturbed_policy = perturbe_policy(ac.pi)
# Simulate a trajectory and compute the total reward
rew, rew_bonus = simulate(perturbed_policy)
# Store data in replay buffer
perturbed_params = nn.utils.parameters_to_vector(list(perturbed_policy.parameters())).to(device, non_blocking=True).detach()
buffer.history.append((perturbed_params, rew, rew_bonus))
statistics.episode += 1
statistics.sim_time += time.perf_counter() - start_time
# Update
if statistics.total_ts > config['start_steps'] and not config['update_every_ts']:
update()
# Log results
log_dict = {'rew': rew,
'rew_bonus': rew_bonus,
'steps': statistics.total_ts,
'episode': statistics.episode,
'grads_norm_policy': core.grad_norm(ac.pi.parameters()),
'norm_policy': core.norm(ac.pi.parameters()),
'norm_pvf': core.norm(ac.v.parameters()),
'grads_norm_pvf': core.grad_norm(ac.v.parameters()),
'norm_prob_states': core.norm(ac.v.probing_states.parameters()),
'grads_norm_prob_states': core.grad_norm(ac.v.probing_states.parameters()),
}
if verbose:
print(log_dict)
return
# Initial evaluation
evaluate(ac)
# Loop over episodes
while statistics.total_ts < config['max_timesteps'] and statistics.episode < config['max_episodes']:
start_time = time.perf_counter()
train()
statistics.total_time += time.perf_counter() - start_time