-
Notifications
You must be signed in to change notification settings - Fork 39
/
run_toy_ql.py
248 lines (196 loc) · 8.13 KB
/
run_toy_ql.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
import os
import torch
import numpy as np
from torch.distributions import Normal
import seaborn as sns
import matplotlib.pyplot as plt
import argparse
from toy_experiments.toy_helpers import Data_Sampler
parser = argparse.ArgumentParser()
parser.add_argument("--ill", action='store_true')
parser.add_argument("--seed", default=2022, type=int)
parser.add_argument("--exp", default='exp_1', type=str)
parser.add_argument("--x", default=0., type=float)
parser.add_argument("--y", default=0., type=float)
parser.add_argument("--eta", default=2.5, type=float)
parser.add_argument('--device', default=0, type=int)
parser.add_argument("--dir", default='whole_grad', type=str)
parser.add_argument("--r_fun", default='no', type=str)
parser.add_argument("--lr", default=3e-4, type=float)
parser.add_argument('--hidden_dim', default=128, type=int)
parser.add_argument("--mode", default='whole_grad', type=str)
args = parser.parse_args()
r_fun_std = 0.25
device = f"cuda:{args.device}" if torch.cuda.is_available() else "cpu"
eta = args.eta
seed = args.seed
lr = args.lr
hidden_dim = args.hidden_dim
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)
r_left_up = 3.0 + 0.5 * torch.randn((each_num, 1))
r_left_bottom = 0.5 * torch.randn((each_num, 1))
r_right_up = 1.5 + 0.5 * torch.randn((each_num, 1))
r_right_bottom = 5.0 + 0.5 * torch.randn((each_num, 1))
reward = torch.cat([r_left_up, r_left_bottom, r_right_up, r_right_bottom], dim=0)
return Data_Sampler(state, action, reward, device)
torch.manual_seed(seed)
np.random.seed(seed)
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 = 64
# eta = 10.0
# lr = 3e-4
num_epochs = 1000
batch_size = 100
iterations = int(num_data / batch_size)
img_dir = f'toy_imgs/{args.dir}'
os.makedirs(img_dir, exist_ok=True)
num_eval = 100
fig, axs = plt.subplots(1, 5, figsize=(5.5 * 5, 5))
axis_lim = 1.1
pos = 0.8
std = 0.05
left_up_conor = Normal(torch.tensor([-pos, pos]), torch.tensor([std, std])).sample((200,)).clip(-1.0, 1.0).numpy()
left_bottom_conor = Normal(torch.tensor([-pos, -pos]), torch.tensor([std, std])).sample((200,)).clip(-1.0, 1.0).numpy()
right_up_conor = Normal(torch.tensor([pos, pos]), torch.tensor([std, std])).sample((200,)).clip(-1.0, 1.0).numpy()
right_bottom_conor = Normal(torch.tensor([pos, -pos]), torch.tensor([std, std])).sample((200,)).clip(-1.0, 1.0).numpy()
axs[0].scatter(left_up_conor[:, 0], left_up_conor[:, 1], label=r"$r \sim N (3.0, 0.5)$")
axs[0].scatter(left_bottom_conor[:, 0], left_bottom_conor[:, 1], label=r"$r \sim N (0.0, 0.5)$")
axs[0].scatter(right_up_conor[:, 0], right_up_conor[:, 1], label=r"$r \sim N (1.5, 0.5)$")
axs[0].scatter(right_bottom_conor[:, 0], right_bottom_conor[:, 1], label=r"$r \sim N (5.0, 0.5)$")
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('Add Reward', fontsize=25)
axs[0].legend(loc='best', fontsize=15, title_fontsize=15)
#fig.colorbar(c, ax=axs[0])
# Plot QL-MLE
from toy_experiments.ql_mle import QL_MLE
agent = QL_MLE(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
eta=eta,
hidden_dim=hidden_dim,
lr=lr,
r_fun=None)
for i in range(1, num_epochs + 1):
agent.train(data_sampler, iterations=iterations, batch_size=batch_size)
if i % 100 == 0:
print(f'QL-MLE Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[1].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3, color='#d62728')
axs[1].set_xlim(-axis_lim, axis_lim)
axs[1].set_ylim(-axis_lim, axis_lim)
axs[1].set_xlabel('x', fontsize=20)
axs[1].set_ylabel('y', fontsize=20)
axs[1].set_title('TD3+BC', fontsize=25)
# Plot QL-CVAE
from toy_experiments.ql_cvae import QL_CVAE
agent = QL_CVAE(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
hidden_dim=hidden_dim,
lr=lr,
r_fun=None)
for i in range(1, num_epochs + 1):
agent.train(data_sampler, iterations=iterations, batch_size=batch_size)
if i % 100 == 0:
print(f'QL-CVAE Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = agent.vae.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[2].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3, color='#d62728')
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('BCQ', fontsize=25)
# Plot QL-MMD
from toy_experiments.ql_mmd import QL_MMD
agent = QL_MMD(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
hidden_dim=hidden_dim,
lr=lr,
r_fun=None)
for i in range(1, num_epochs + 1):
agent.train(data_sampler, iterations=iterations, batch_size=batch_size)
if i % 100 == 0:
print(f'QL-MMD Epoch: {i}')
new_state = torch.zeros((num_eval, 2), device=device)
new_action = agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[3].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3, color='#d62728')
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('BEAR-MMD', fontsize=25)
# Plot QL-Diffusion
from toy_experiments.ql_diffusion import QL_Diffusion
agent = QL_Diffusion(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=discount,
tau=tau,
eta=eta,
beta_schedule=beta_schedule,
n_timesteps=T,
model_type=model_type,
hidden_dim=hidden_dim,
lr=lr,
r_fun=None,
mode=args.mode)
for i in range(1, num_epochs+1):
b_loss, q_loss = agent.train(data_sampler, iterations=iterations, batch_size=batch_size)
if i % 100 == 0:
print(f'QL-Diffusion Epoch: {i} B_loss {b_loss} Q_loss {q_loss}')
# fig, ax = plt.subplots()
new_state = torch.zeros((num_eval, 2), device=device)
new_action = agent.actor.sample(new_state)
new_action = new_action.detach().cpu().numpy()
axs[4].scatter(new_action[:, 0], new_action[:, 1], alpha=0.3, color='#d62728')
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('Diffusion-QL', fontsize=25)
file_name = f'ql_all_T{T}_eta{eta}_r_fun{args.r_fun}_lr{lr}_hd{hidden_dim}_mode_{args.mode}'
file_name += f'_sd{args.seed}.pdf'
fig.tight_layout()
fig.savefig(os.path.join(img_dir, file_name))