-
Notifications
You must be signed in to change notification settings - Fork 39
/
Copy pathrun_toy_bc.py
212 lines (165 loc) · 6.44 KB
/
run_toy_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
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
import os
import torch
import numpy as np
from torch.distributions import Normal
import argparse
import matplotlib.pyplot as plt
from toy_experiments.toy_helpers import Data_Sampler
parser = argparse.ArgumentParser()
parser.add_argument("--seed", default=2022, type=int)
args = parser.parse_args()
seed = args.seed
def generate_data(num, device = 'cpu'):
each_num = int(num / 4)
pos = 0.8
std = 0.05
left_up_conor = Normal(torch.tensor([-pos, pos]), torch.tensor([std, std]))
left_bottom_conor = Normal(torch.tensor([-pos, -pos]), torch.tensor([std, std]))
right_up_conor = Normal(torch.tensor([pos, pos]), torch.tensor([std, std]))
right_bottom_conor = Normal(torch.tensor([pos, -pos]), torch.tensor([std, std]))
left_up_samples = left_up_conor.sample((each_num,)).clip(-1.0, 1.0)
left_bottom_samples = left_bottom_conor.sample((each_num,)).clip(-1.0, 1.0)
right_up_samples = right_up_conor.sample((each_num,)).clip(-1.0, 1.0)
right_bottom_samples = right_bottom_conor.sample((each_num,)).clip(-1.0, 1.0)
data = torch.cat([left_up_samples, left_bottom_samples, right_up_samples, right_bottom_samples], dim=0)
action = data
state = torch.zeros_like(action)
reward = torch.zeros((num, 1))
return Data_Sampler(state, action, reward, device)
torch.manual_seed(seed)
np.random.seed(seed)
device = 'cuda:0'
num_data = int(10000)
data_sampler = generate_data(num_data, device)
state_dim = 2
action_dim = 2
max_action = 1.0
discount = 0.99
tau = 0.005
model_type = 'MLP'
T = 50
beta_schedule = 'vp'
hidden_dim = 128
lr = 3e-4
num_epochs = 1000
batch_size = 100
iterations = int(num_data / batch_size)
img_dir = 'toy_imgs/bc'
os.makedirs(img_dir, exist_ok=True)
fig, axs = plt.subplots(1, 5, figsize=(5.5 * 5, 5))
axis_lim = 1.1
# Plot the ground truth
num_eval = 1000
_, action_samples, _ = data_sampler.sample(num_eval)
action_samples = action_samples.cpu().numpy()
axs[0].scatter(action_samples[:, 0], action_samples[:, 1], alpha=0.3)
axs[0].set_xlim(-axis_lim, axis_lim)
axs[0].set_ylim(-axis_lim, axis_lim)
axs[0].set_xlabel('x', fontsize=20)
axs[0].set_ylabel('y', fontsize=20)
axs[0].set_title('Ground Truth', fontsize=25)
# Plot MLE BC
from toy_experiments.bc_mle import BC_MLE as MLE_Agent
mle_agent = MLE_Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
lr=lr,
hidden_dim=hidden_dim)
for i in range(num_epochs):
mle_agent.train(data_sampler,
iterations=iterations,
batch_size=batch_size)
if i % 100 == 0:
print(f'Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = mle_agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[1].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3)
axs[1].set_xlim(-2.5, 2.5)
axs[1].set_ylim(-2.5, 2.5)
axs[1].set_xlabel('x', fontsize=20)
axs[1].set_ylabel('y', fontsize=20)
axs[1].set_title('BC-MLE', fontsize=25)
# Plot CVAE BC
from toy_experiments.bc_cvae import BC_CVAE as CVAE_Agent
cvae_agent = CVAE_Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
lr=lr,
hidden_dim=hidden_dim)
for i in range(num_epochs):
cvae_agent.train(data_sampler,
iterations=iterations,
batch_size=batch_size)
if i % 100 == 0:
print(f'Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = cvae_agent.vae.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[2].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3)
axs[2].set_xlim(-axis_lim, axis_lim)
axs[2].set_ylim(-axis_lim, axis_lim)
axs[2].set_xlabel('x', fontsize=20)
axs[2].set_ylabel('y', fontsize=20)
axs[2].set_title('BC-CVAE', fontsize=25)
# Plot CVAE BC
from toy_experiments.bc_mmd import BC_MMD as MMD_Agent
mmd_agent = MMD_Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
lr=lr,
hidden_dim=hidden_dim)
for i in range(num_epochs):
mmd_agent.train(data_sampler,
iterations=iterations,
batch_size=batch_size)
if i % 100 == 0:
print(f'Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = mmd_agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[3].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3)
axs[3].set_xlim(-axis_lim, axis_lim)
axs[3].set_ylim(-axis_lim, axis_lim)
axs[3].set_xlabel('x', fontsize=20)
axs[3].set_ylabel('y', fontsize=20)
axs[3].set_title('BC-MMD', fontsize=25)
# Plot Diffusion BC
from toy_experiments.bc_diffusion import BC as Diffusion_Agent
diffusion_agent = Diffusion_Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
beta_schedule=beta_schedule,
n_timesteps=T,
model_type=model_type,
hidden_dim=hidden_dim,
lr=lr)
for i in range(num_epochs):
diffusion_agent.train(data_sampler,
iterations=iterations,
batch_size=batch_size)
if i % 100 == 0:
print(f'Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = diffusion_agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[4].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3)
axs[4].set_xlim(-axis_lim, axis_lim)
axs[4].set_ylim(-axis_lim, axis_lim)
axs[4].set_xlabel('x', fontsize=20)
axs[4].set_ylabel('y', fontsize=20)
axs[4].set_title('BC-Diffusion', fontsize=25)
fig.tight_layout()
fig.savefig(os.path.join(img_dir, f'bc_diffusion_{T}_sd{seed}.pdf'))