-
Notifications
You must be signed in to change notification settings - Fork 0
/
pfn_env.py
252 lines (219 loc) · 10.2 KB
/
pfn_env.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
import math
import time
import gymnasium as gym
import os
import torch
import torch.nn as nn
import numpy as np
from stable_baselines3 import PPO
import grid_world
from train import build_model
import encoders
import simple_env
from decoder import *
# if GPU is to be used
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def get_done_func(env_name):
if env_name == "CartPole-v1" or env_name == "CartPole-v0":
def cartpole_reset(state, steps):
# Limits for resetting
theta_threshold_radians = 12 * 2 * math.pi / 360
x_threshold = 2.4
x = state[0]
theta = state[2]
done = bool(
x < -x_threshold
or x > x_threshold
or theta < -theta_threshold_radians
or theta > theta_threshold_radians
or steps > 199
)
return done
return cartpole_reset
elif env_name == "Pendulum-v1":
def pendulum_reset(state, steps):
return steps > 199
return pendulum_reset
elif env_name == "SimpleEnv":
def simple_env_reset(state, steps):
return steps > 20
return simple_env_reset
elif env_name == "Reacher-v4":
def reacher_reset(state, steps):
return steps > 50
return reacher_reset
elif env_name == "MountainCar-v0":
def mountain_car_reset(state, steps):
return state[0] >= 0.5 or steps > 199
return mountain_car_reset
elif env_name == "GridWorld":
def grid_world_reset(state, steps):
d1 = state[0] <= 0 or state[1] <= 0
d2 = state[0] >= 6 or state[1] >= 7
return d1 or d2 or steps > 25
return grid_world_reset
else:
raise NotImplementedError
class ArtificialEnv(gym.Env):
def __init__(self, env_name, render_mode=None, mixure_context=False, p_expert_context=False):
self.round = False
# Create environment to reset and create train samples
if env_name == "SimpleEnv":
self.real_env = simple_env.SimpleEnv()
elif env_name == "GridWorld":
self.real_env = grid_world.GridWorld()
self.round = True
else:
self.real_env = gym.make(env_name)
self.action_space = self.real_env.action_space
self.observation_space = self.real_env.observation_space
self.done_func = get_done_func(env_name)
self.episode_steps = 0
num_features = 14
batch_size = 1
seq_len = 1001
self.train_x = torch.full((seq_len, batch_size, num_features), 0.)
self.train_y = torch.full((seq_len, batch_size, num_features), float(0.))
observation, info = self.real_env.reset()
if mixure_context or p_expert_context:
if env_name == "CartPole-v0":
print("val_transitions/expert_policies/PPO_" + "CartPole-v1" + ".zip")
policy = PPO.load("val_transitions/expert_policies/PPO_" + "CartPole-v1")
else:
policy = PPO.load("val_transitions/expert_policies/PPO_" + env_name + ".zip")
for b in range(batch_size):
ep = 0
ep_step = 0
steps_after_done = 0
for i in range(1001):
if mixure_context:
if i < seq_len // 3:
action, _ = policy.predict(observation)
elif i < 2 * (seq_len // 3):
eps = np.random.rand()
if eps < 0.5:
action = self.real_env.action_space.sample()
else:
action, _ = policy.predict(observation)
else:
action = self.real_env.action_space.sample()
elif p_expert_context:
eps = np.random.rand()
if eps < 0.5:
action = self.real_env.action_space.sample()
else:
action, _ = policy.predict(observation)
else:
action = self.real_env.action_space.sample()
if isinstance(action, int) or isinstance(action, np.int64):
action_array = np.array([action]) # TODO detect action type
elif action.shape == ():
action_array = np.expand_dims(action, 0)
else:
action_array = action
act = torch.full((3,), 0.)
act[:action_array.shape[0]] = torch.tensor(action_array)
obs = torch.full((num_features - 3,), 0.)
obs[:observation.shape[0]] = torch.tensor(observation)
obs_action_pair = torch.hstack((obs, act))
# obs = torch.full((num_features - 1,), 0.)
# obs[:observation.shape[0]] = torch.tensor(observation)
# obs_action_pair = torch.hstack((obs, torch.tensor(action)))
# batch_features = observation.shape[0] + 1 # action.shape[0]
self.train_x[i, b] = obs_action_pair # * num_features / batch_features
observation, reward, terminated, truncated, info = self.real_env.step(action)
ep_step += 1
obs = torch.full((num_features - 1,), 0.)
obs[:observation.shape[0]] = torch.tensor(observation)
# obs[-1] = float(terminated or truncated) # TODO if possible for all env
next_state_reward_pair = torch.hstack((obs, torch.tensor(reward)))
self.train_y[i, b] = next_state_reward_pair
if terminated or truncated or ep_step >= 50:
steps_after_done += 1
sad = 5 if env_name == "CartPole-v0" else 0
if steps_after_done >= sad:
steps_after_done = 0
observation, info = self.real_env.reset()
ep += 1
ep_step = 0
print(f"Episode {ep}")
self.train_y = self.train_y.to(device)
self.train_x = self.train_x.to(device)
# Normalize the columns to 0 mean and 1 Variance
self.x_mean = torch.mean(self.train_x[:1000, :], dim=0)
self.x_std = torch.std(self.train_x[:1000, :], dim=0)
self.train_x = torch.nan_to_num((self.train_x - self.x_mean) / self.x_std, nan=0)
self.y_mean = torch.mean(self.train_y[:1000, :], dim=0)
self.y_std = torch.std(self.train_y[:1000, :], dim=0)
self.train_y = torch.nan_to_num((self.train_y - self.y_mean) / self.y_std, nan=0)
# building Transformer model and loading weights
criterion = nn.MSELoss(reduction='none')
# TODO test batch?
encoder_decoder_hps = {"decoder_activation": "sigmoid", "decoder_depth": 2, "decoder_res_connection": True,
"decoder_type": "cat", "decoder_use_bias": False, "decoder_width": 64,
"encoder_activation": "gelu", "encoder_depth": 3,
"encoder_res_connection": True, "encoder_type": "cat", "encoder_use_bias": True,
"encoder_width": 512}
if encoder_decoder_hps["encoder_type"] == "mlp":
gen_x = mlp_encoder_generator_generator(encoder_decoder_hps)
gen_y = gen_x
elif encoder_decoder_hps["encoder_type"] == "cat":
gen_x = cat_encoder_generator_generator(encoder_decoder_hps, target=False)
gen_y = cat_encoder_generator_generator(encoder_decoder_hps, target=True)
if encoder_decoder_hps["encoder_type"] == "mlp":
dec_model = mlp_decoder_generator_generator(encoder_decoder_hps)
elif encoder_decoder_hps["decoder_type"] == "cat":
dec_model = cat_decoder_generator_generator(encoder_decoder_hps)
decoder_dict = {"standard": (dec_model, 14)}
hps = {'test': True}
self.pfn = build_model(
criterion=criterion,
encoder_generator=gen_x,
test_batch=None,
n_out=14,
emsize=512, nhead=8, nhid=1024, nlayers=6,
seq_len=1001,
y_encoder_generator=gen_y,
decoder_dict=decoder_dict,
extra_prior_kwargs_dict={'num_features': num_features, 'hyperparameters': hps},
).to(device)
print(
f"Using a Transformer with {sum(p.numel() for p in self.pfn.parameters()) / 1000 / 1000:.{2}f} M parameters"
)
self.pfn.load_state_dict(torch.load("saved_models/exp_seed_1.pt"))
self.pfn.eval()
self.state = None
def step(self, a):
a = a # .item()
# TODO check if all normalizations are correct
if self.state is None:
raise gym.error.ResetNeeded
if isinstance(a, int) or isinstance(a, np.int64):
action_array = np.array([a]) # TODO detect action type
else:
action_array = a
act = torch.full((3,), 0.)
act[:action_array.shape[0]] = torch.tensor(action_array)
obs = torch.full((14 - 3,), 0.)
obs[:self.state.shape[0]] = torch.tensor(self.state)
obs = torch.hstack((obs, act)).to(device)
norm_state_action = torch.nan_to_num((obs - self.x_mean) / self.x_std)
self.train_x[1000, :, :] = norm_state_action.to(device)
with torch.no_grad():
logits = self.pfn(self.train_x[:1000], self.train_y[:1000], self.train_x[:])
ns = logits[1000, :, :].detach().clone() * self.y_std + self.y_mean
ns = torch.mean(ns, dim=0) # TODO discrete steps
self.state = ns[:self.state.shape[0]].cpu().numpy()
if self.round:
self.state = self.state.round()
re = torch.nan_to_num(ns[-1], nan=-1) # TODO if this accurate enough
done = self.done_func(self.state, self.episode_steps)
self.episode_steps += 1
return self.state, re, done, False, {}
def reset(self, **kwargs):
# TODO reset based on real environment
self.episode_steps = 0
self.state, information = self.real_env.reset()
self.train_y[1000:] = 0
self.train_x[1000:] = 0
return self.state, information