-
Notifications
You must be signed in to change notification settings - Fork 0
/
bc.py
119 lines (86 loc) · 3.28 KB
/
bc.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
import numpy as np
import cv2
from doorsenvs import Doors,DoorZ
from agents import PPOAgent
import torch
import torch.optim as optim
def collect_expert_trajs():
traj_len = 32
env = Doors(max_steps=traj_len)
x = 0
obs = env.reset(agent_at=x)
n_episodes = 15
all_actions = [[]]
all_obses = [[]]
all_features = []
for i in range(traj_len*n_episodes):
all_obses[-1].append(obs.copy())
all_features.append(np.hstack((env.goal,env.agent)))
a = env.expert_action()
env.render(scale=20)
obs,reward,done,info = env.step(a)
cv2.waitKey(10)
all_actions[-1].append(a)
if done:
x+=1
obs = env.reset(agent_at=min(x,n_episodes-1))
all_actions.append([])
all_obses.append([])
print(np.array(all_actions[:-1]).shape)
print(np.array(all_obses[:-1]).shape)
np.save(f'expert_actions_15_{traj_len}.npy',np.array(all_actions[:-1]))
np.save(f'expert_obses_15_{traj_len}_225.npy',np.array(all_obses[:-1]))
np.save(f'expert_features_15_{traj_len}.npy',np.array(all_features[:]))
env.close()
def collect_expert_trajs_modes():
traj_len = 32
env = DoorZ(max_steps=traj_len)
x = 0
obs = env.reset(agent_at=x)
n_episodes = 15*3
all_actions = [[]]
all_obses = [[]]
all_features = []
for i in range(traj_len*n_episodes):
all_obses[-1].append(obs.copy())
all_features.append(np.hstack((env.goal,env.agent)))
a = env.expert_action()
env.render(scale=20)
obs,reward,done,info = env.step(a)
cv2.waitKey(10)
all_actions[-1].append(a)
if done:
x+=1
obs = env.reset(agent_at=x%env.gridsize[1])#min(x,n_episodes-1))
all_actions.append([])
all_obses.append([])
print(np.array(all_actions[:-1]).shape)
print(np.array(all_obses[:-1]).shape)
np.save(f'expert_actions_15_{traj_len}_Z.npy',np.array(all_actions[:-1]))
np.save(f'expert_obses_15_{traj_len}_225_Z.npy',np.array(all_obses[:-1]))
np.save(f'expert_features_15_{traj_len}_Z.npy',np.array(all_features[:]))
env.close()
def main():
#collect_expert_trajs()
traj_len = 32
training_epochs = 3000
lr = 9e-4
demos_size = 240
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
env = Doors(max_steps=traj_len)
agent = PPOAgent(envs=[env]).to(device=device).float()
optimizer = optim.Adam(agent.actor.parameters(), lr=lr, eps=1e-5)
expert_actions = torch.as_tensor(np.load(f'expert_actions_15_{traj_len}.npy').flatten(),device=device).long()[:demos_size]
onehot_expert_acts = torch.functional.F.one_hot(expert_actions,num_classes=5).float()
expert_obs = torch.as_tensor(np.load(f'expert_obses_15_{traj_len}_225.npy').reshape(-1,np.prod(env.gridsize)),device=device).float()[:demos_size,:]
agent.actor.train()
for epoch in range(training_epochs):
out = agent.actor(expert_obs)
loss = torch.functional.F.binary_cross_entropy_with_logits(out,onehot_expert_acts)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(loss.item())
torch.save(agent,f'bc_agent_iter_{training_epochs}_mlp.pth')
if __name__=='__main__':
collect_expert_trajs_modes()