-
Notifications
You must be signed in to change notification settings - Fork 39
/
Copy pathmain.py
264 lines (226 loc) · 13.7 KB
/
main.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
# Copyright 2022 Twitter, Inc and Zhendong Wang.
# SPDX-License-Identifier: Apache-2.0
import argparse
import gym
import numpy as np
import os
import torch
import json
import d4rl
from utils import utils
from utils.data_sampler import Data_Sampler
from utils.logger import logger, setup_logger
from torch.utils.tensorboard import SummaryWriter
hyperparameters = {
'halfcheetah-medium-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 9.0, 'top_k': 1},
'hopper-medium-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 9.0, 'top_k': 2},
'walker2d-medium-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 1.0, 'top_k': 1},
'halfcheetah-medium-replay-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 2.0, 'top_k': 0},
'hopper-medium-replay-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 4.0, 'top_k': 2},
'walker2d-medium-replay-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 4.0, 'top_k': 1},
'halfcheetah-medium-expert-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 7.0, 'top_k': 0},
'hopper-medium-expert-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 5.0, 'top_k': 2},
'walker2d-medium-expert-v2': {'lr': 3e-4, 'eta': 1.0, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 2000, 'gn': 5.0, 'top_k': 1},
'antmaze-umaze-v0': {'lr': 3e-4, 'eta': 0.5, 'max_q_backup': False, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 2.0, 'top_k': 2},
'antmaze-umaze-diverse-v0': {'lr': 3e-4, 'eta': 2.0, 'max_q_backup': True, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 3.0, 'top_k': 2},
'antmaze-medium-play-v0': {'lr': 1e-3, 'eta': 2.0, 'max_q_backup': True, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 2.0, 'top_k': 1},
'antmaze-medium-diverse-v0': {'lr': 3e-4, 'eta': 3.0, 'max_q_backup': True, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 1.0, 'top_k': 1},
'antmaze-large-play-v0': {'lr': 3e-4, 'eta': 4.5, 'max_q_backup': True, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 10.0, 'top_k': 2},
'antmaze-large-diverse-v0': {'lr': 3e-4, 'eta': 3.5, 'max_q_backup': True, 'reward_tune': 'cql_antmaze', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 7.0, 'top_k': 1},
'pen-human-v1': {'lr': 3e-5, 'eta': 0.15, 'max_q_backup': False, 'reward_tune': 'normalize', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 7.0, 'top_k': 2},
'pen-cloned-v1': {'lr': 3e-5, 'eta': 0.1, 'max_q_backup': False, 'reward_tune': 'normalize', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 8.0, 'top_k': 2},
'kitchen-complete-v0': {'lr': 3e-4, 'eta': 0.005, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 250 , 'gn': 9.0, 'top_k': 2},
'kitchen-partial-v0': {'lr': 3e-4, 'eta': 0.005, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 10.0, 'top_k': 2},
'kitchen-mixed-v0': {'lr': 3e-4, 'eta': 0.005, 'max_q_backup': False, 'reward_tune': 'no', 'eval_freq': 50, 'num_epochs': 1000, 'gn': 10.0, 'top_k': 0},
}
def train_agent(env, state_dim, action_dim, max_action, device, output_dir, args):
# Load buffer
dataset = d4rl.qlearning_dataset(env)
data_sampler = Data_Sampler(dataset, device, args.reward_tune)
utils.print_banner('Loaded buffer')
if args.algo == 'ql':
from agents.ql_diffusion import Diffusion_QL as Agent
agent = Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=args.discount,
tau=args.tau,
max_q_backup=args.max_q_backup,
beta_schedule=args.beta_schedule,
n_timesteps=args.T,
eta=args.eta,
lr=args.lr,
lr_decay=args.lr_decay,
lr_maxt=args.num_epochs,
grad_norm=args.gn)
elif args.algo == 'bc':
from agents.bc_diffusion import Diffusion_BC as Agent
agent = Agent(state_dim=state_dim,
action_dim=action_dim,
max_action=max_action,
device=device,
discount=args.discount,
tau=args.tau,
beta_schedule=args.beta_schedule,
n_timesteps=args.T,
lr=args.lr)
early_stop = False
stop_check = utils.EarlyStopping(tolerance=1, min_delta=0.)
writer = None # SummaryWriter(output_dir)
evaluations = []
training_iters = 0
max_timesteps = args.num_epochs * args.num_steps_per_epoch
metric = 100.
utils.print_banner(f"Training Start", separator="*", num_star=90)
while (training_iters < max_timesteps) and (not early_stop):
iterations = int(args.eval_freq * args.num_steps_per_epoch)
loss_metric = agent.train(data_sampler,
iterations=iterations,
batch_size=args.batch_size,
log_writer=writer)
training_iters += iterations
curr_epoch = int(training_iters // int(args.num_steps_per_epoch))
# Logging
utils.print_banner(f"Train step: {training_iters}", separator="*", num_star=90)
logger.record_tabular('Trained Epochs', curr_epoch)
logger.record_tabular('BC Loss', np.mean(loss_metric['bc_loss']))
logger.record_tabular('QL Loss', np.mean(loss_metric['ql_loss']))
logger.record_tabular('Actor Loss', np.mean(loss_metric['actor_loss']))
logger.record_tabular('Critic Loss', np.mean(loss_metric['critic_loss']))
logger.dump_tabular()
# Evaluation
eval_res, eval_res_std, eval_norm_res, eval_norm_res_std = eval_policy(agent, args.env_name, args.seed,
eval_episodes=args.eval_episodes)
evaluations.append([eval_res, eval_res_std, eval_norm_res, eval_norm_res_std,
np.mean(loss_metric['bc_loss']), np.mean(loss_metric['ql_loss']),
np.mean(loss_metric['actor_loss']), np.mean(loss_metric['critic_loss']),
curr_epoch])
np.save(os.path.join(output_dir, "eval"), evaluations)
logger.record_tabular('Average Episodic Reward', eval_res)
logger.record_tabular('Average Episodic N-Reward', eval_norm_res)
logger.dump_tabular()
bc_loss = np.mean(loss_metric['bc_loss'])
if args.early_stop:
early_stop = stop_check(metric, bc_loss)
metric = bc_loss
if args.save_best_model:
agent.save_model(output_dir, curr_epoch)
# Model Selection: online or offline
scores = np.array(evaluations)
if args.ms == 'online':
best_id = np.argmax(scores[:, 2])
best_res = {'model selection': args.ms, 'epoch': scores[best_id, -1],
'best normalized score avg': scores[best_id, 2],
'best normalized score std': scores[best_id, 3],
'best raw score avg': scores[best_id, 0],
'best raw score std': scores[best_id, 1]}
with open(os.path.join(output_dir, f"best_score_{args.ms}.txt"), 'w') as f:
f.write(json.dumps(best_res))
elif args.ms == 'offline':
bc_loss = scores[:, 4]
top_k = min(len(bc_loss) - 1, args.top_k)
where_k = np.argsort(bc_loss) == top_k
best_res = {'model selection': args.ms, 'epoch': scores[where_k][0][-1],
'best normalized score avg': scores[where_k][0][2],
'best normalized score std': scores[where_k][0][3],
'best raw score avg': scores[where_k][0][0],
'best raw score std': scores[where_k][0][1]}
with open(os.path.join(output_dir, f"best_score_{args.ms}.txt"), 'w') as f:
f.write(json.dumps(best_res))
# writer.close()
# Runs policy for X episodes and returns average reward
# A fixed seed is used for the eval environment
def eval_policy(policy, env_name, seed, eval_episodes=10):
eval_env = gym.make(env_name)
eval_env.seed(seed + 100)
scores = []
for _ in range(eval_episodes):
traj_return = 0.
state, done = eval_env.reset(), False
while not done:
action = policy.sample_action(np.array(state))
state, reward, done, _ = eval_env.step(action)
traj_return += reward
scores.append(traj_return)
avg_reward = np.mean(scores)
std_reward = np.std(scores)
normalized_scores = [eval_env.get_normalized_score(s) for s in scores]
avg_norm_score = eval_env.get_normalized_score(avg_reward)
std_norm_score = np.std(normalized_scores)
utils.print_banner(f"Evaluation over {eval_episodes} episodes: {avg_reward:.2f} {avg_norm_score:.2f}")
return avg_reward, std_reward, avg_norm_score, std_norm_score
if __name__ == "__main__":
parser = argparse.ArgumentParser()
### Experimental Setups ###
parser.add_argument("--exp", default='exp_1', type=str) # Experiment ID
parser.add_argument('--device', default=0, type=int) # device, {"cpu", "cuda", "cuda:0", "cuda:1"}, etc
parser.add_argument("--env_name", default="walker2d-medium-expert-v2", type=str) # OpenAI gym environment name
parser.add_argument("--dir", default="results", type=str) # Logging directory
parser.add_argument("--seed", default=0, type=int) # Sets Gym, PyTorch and Numpy seeds
parser.add_argument("--num_steps_per_epoch", default=1000, type=int)
### Optimization Setups ###
parser.add_argument("--batch_size", default=256, type=int)
parser.add_argument("--lr_decay", action='store_true')
parser.add_argument('--early_stop', action='store_true')
parser.add_argument('--save_best_model', action='store_true')
### RL Parameters ###
parser.add_argument("--discount", default=0.99, type=float)
parser.add_argument("--tau", default=0.005, type=float)
### Diffusion Setting ###
parser.add_argument("--T", default=5, type=int)
parser.add_argument("--beta_schedule", default='vp', type=str)
### Algo Choice ###
parser.add_argument("--algo", default="ql", type=str) # ['bc', 'ql']
parser.add_argument("--ms", default='offline', type=str, help="['online', 'offline']")
# parser.add_argument("--top_k", default=1, type=int)
# parser.add_argument("--lr", default=3e-4, type=float)
# parser.add_argument("--eta", default=1.0, type=float)
# parser.add_argument("--max_q_backup", action='store_true')
# parser.add_argument("--reward_tune", default='no', type=str)
# parser.add_argument("--gn", default=-1.0, type=float)
args = parser.parse_args()
args.device = f"cuda:{args.device}" if torch.cuda.is_available() else "cpu"
args.output_dir = f'{args.dir}'
args.num_epochs = hyperparameters[args.env_name]['num_epochs']
args.eval_freq = hyperparameters[args.env_name]['eval_freq']
args.eval_episodes = 10 if 'v2' in args.env_name else 100
args.lr = hyperparameters[args.env_name]['lr']
args.eta = hyperparameters[args.env_name]['eta']
args.max_q_backup = hyperparameters[args.env_name]['max_q_backup']
args.reward_tune = hyperparameters[args.env_name]['reward_tune']
args.gn = hyperparameters[args.env_name]['gn']
args.top_k = hyperparameters[args.env_name]['top_k']
# Setup Logging
file_name = f"{args.env_name}|{args.exp}|diffusion-{args.algo}|T-{args.T}"
if args.lr_decay: file_name += '|lr_decay'
file_name += f'|ms-{args.ms}'
if args.ms == 'offline': file_name += f'|k-{args.top_k}'
file_name += f'|{args.seed}'
results_dir = os.path.join(args.output_dir, file_name)
if not os.path.exists(results_dir):
os.makedirs(results_dir)
utils.print_banner(f"Saving location: {results_dir}")
# if os.path.exists(os.path.join(results_dir, 'variant.json')):
# raise AssertionError("Experiment under this setting has been done!")
variant = vars(args)
variant.update(version=f"Diffusion-Policies-RL")
env = gym.make(args.env_name)
env.seed(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
variant.update(state_dim=state_dim)
variant.update(action_dim=action_dim)
variant.update(max_action=max_action)
setup_logger(os.path.basename(results_dir), variant=variant, log_dir=results_dir)
utils.print_banner(f"Env: {args.env_name}, state_dim: {state_dim}, action_dim: {action_dim}")
train_agent(env,
state_dim,
action_dim,
max_action,
args.device,
results_dir,
args)